DoorDash has moved task delegation from individual engineers to a centralized platform that runs 130,000 agent tasks a month with human review happening after execution, not before.

Airbnb has collapsed the PRD-to-delivery handoff into prototype-first collaboration where 60% of code is already AI-authored by the time a pull request appears.

The uncomfortable implication is that both companies have made speed structurally dependent on governance infrastructure that most engineering organizations have not built, and copying the output metrics without copying the sandboxed execution environments, permissioned tooling, and institutional knowledge layers means absorbing the autonomy without any of the control.

AI Operating Model Case — DoorDash

Organization Type: Scale-up

Industry: Logistics & Supply Chain

Initiative starting date: 2026
Evidence Quality Score: 3/5

Scope of AI in Development Lifecycle: AI in engineering delivery, AI in testing, AI in release, AI in post launch analysis

Description: DoorDash built Flux, an internal cloud platform that runs autonomous AI agents inside isolated Firecracker microVM environments to execute engineering tasks across coding, review, CI investigation, and recurring maintenance. Engineers delegate work through GitHub, Slack, CLI triggers, and scheduled jobs rather than relying on individual, laptop-based assistants. Flux standardizes execution with a centralized tool-access gateway, permission controls, and more than 300 reusable playbooks, replacing fragmented team-specific integrations with shared infrastructure. The shift moves engineers toward task delegation and review responsibilities and restructures how routine engineering work is organized and governed.

Nature of Integration: DoorDash built Flux as a centralized internal platform that provides shared execution infrastructure for AI agents across engineering teams, replacing the previous pattern of team-specific, laptop-based AI integrations. Agents run inside Firecracker microVM sandboxes, which provide isolation and security for unattended task execution, and access external tools through a centralized MCP-based gateway that enforces permission controls and audit logging. Engineers trigger agents through GitHub, Slack, the command line, or scheduled jobs, depending on the task, and over 300 reusable playbooks standardize how recurring workflows are delegated and executed. The underlying foundation models powering Flux are not publicly disclosed. Human oversight varies by workflow: engineers initiate tasks and review agent-generated findings or code changes, but agents can investigate CI failures, run maintenance jobs, and modify code without continuous supervision once a task is delegated.

Workflow Impact: Before Flux, engineers conducted pull-request inspection, implemented fixes, triaged CI failures, and ran maintenance either manually or via fragmented, team-built integrations tied to developer sessions. After Flux, many of those manual steps are delegated to cloud-based agents that run asynchronously and in parallel, changing the execution model from interactive assistance to unattended task delegation. The workflow now depends on a centralized execution layer for recurring and investigative tasks, although explicit changes in release cycle time or throughput are not disclosed. Engineers remain responsible for reviewing and approving outputs, and a centralized platform function maintains the shared agent environments, playbooks, and access controls.

Claimed Impact: DoorDash reports 130,000 agent tasks executed per month, more than 25,000 automated code reviews completed weekly, and adoption of over 300 reusable playbooks, indicating broad operational deployment. The company states the code-review agent has high engineer acceptance and reports improvements in review latency and per-review cost, but precise performance figures and success-rate metrics are not disclosed. These figures are self-reported and presented as utilization and adoption metrics rather than independently verified productivity or ROI measurements.

Transformation Classification: Level 3

Flux centralizes execution, governance, and tooling for agentic engineering tasks across the organization, changing how work is initiated, executed, and reviewed rather than only improving individual productivity. The combination of sandboxed execution, an MCP gateway for tool access, standardized playbooks, and integration with GitHub/Slack/CLI represents an organizational-level platform change. Reported scale of usage indicates adoption beyond isolated experiments, supporting classification as an operating-model integration within engineering.

Strategic Signal: DoorDash’s Flux demonstrates a move from individual AI assistants toward a shared, governed execution layer that embeds agentic workflows into core engineering processes. The approach’s defensibility comes primarily from operational integration—playbooks, permissioned tool access, auditability, and sandbox infrastructure—rather than from disclosed proprietary foundation models. Engineering teams benefit most, gaining the ability to offload repeatable implementation, review, and maintenance work to asynchronous agents while focusing on oversight and task framing. Replicating the platform requires meaningful engineering investment to reproduce the governance, workflow library, and secure execution environment, suggesting moderate difficulty for peers to emulate.

Retrieve the case and more (including their sources) on the AI in USE website

AI Operating Model Case — Airbnb

Organization Type: Public

Industry: Travel & Tourism

Initiative starting date: Not disclosed
Evidence Quality Score: 3/5

Scope of AI in Development Lifecycle: AI in prd spec, AI in design, AI in engineering delivery, AI in testing, AI in post launch analysis

Description: Airbnb is restructuring its internal product and engineering operating model around AI-assisted prototyping, code generation, organizational knowledge retrieval, and asynchronous engineering agents. Traditional PRD→design→engineering handoffs are being compressed into prototype-first workflows where product managers, designers, and engineers collaborate directly around executable code. Its internal Everest knowledge graph and AirChat agent surface organizational and codebase context across teams, enabling engineers to work beyond previous specializations. AI coauthors approximately 60% of code, while monitoring-triggered agents investigate production alerts and propose remediation for human review.

Nature of Integration: Airbnb operates a hybrid internal AI layer that combines an organizational knowledge platform (Everest), an agent-based assistant (AirChat), AI coding copilots, and containerized asynchronous engineering agents triggered by monitoring events. Everest provides retrieval over organizational knowledge, codebase patterns, and cross-system relationships to ground AI outputs in proprietary context, while AirChat connects engineers to that knowledge through MCP-connected tooling. The architecture leverages LLMs, embeddings, and agentic workflows, though precise infrastructure details are not disclosed. Humans retain high responsibility: engineers validate AI-generated code, review architecture and tests, and approve pull requests proposed by agents, while some low-risk alert handling may proceed with reduced intervention.

Workflow Impact: Before, teams followed sequential PRD, design, and engineering handoffs with engineers relying on specialist codebase knowledge and manually investigating incidents. After the change, sequential documentation is compressed into prototype-first collaboration and AI now coauthors roughly 60% of code, reducing manual implementation work. Monitoring-triggered asynchronous agents handle initial incident investigation and prepare remediation proposals for human review, and engineers increasingly rely on Everest to retrieve implementation patterns instead of searching manually. Airbnb reports faster build–test–iterate cycles, a reported 1.6× pull-request throughput per average engineer, and role shifts toward supervising AI outputs and cross-functional prototype collaboration, though full company-wide adoption and AI’s isolated causal effect are not disclosed.

Claimed Impact: Airbnb reports that AI coauthors approximately 60% of its code and that average engineer pull-request throughput is about 1.6× prior levels. The company also reports up to a 60% reduction in concept-to-delivery time on key initiatives and nearly 80% more features and improvements shipped year over year in H1 2026, with an illustrative project comparison showing large differences in delivery time attributed to knowledge reuse via Everest. These figures are company-reported and appear in the CTO interview and official financial results, but the methodology is not published and the isolated contribution of AI versus other organizational changes is not independently validated.

Transformation Classification: Level 3

Airbnb has moved beyond individual developer tools to redesign coordination between product, design, and engineering by replacing sequential documentation with executable prototypes and by making institutional code knowledge retrievable across teams. The introduction of monitoring-triggered agents for investigation and remediation proposals further integrates AI into engineering operations. Together these changes restructure work assignment, supervision, and cross-functional collaboration, indicating an operating-model-level transformation rather than a simple workflow optimization.

Strategic Signal: Airbnb’s approach positions AI as a shared execution layer that spans specification, implementation, and incident triage, with engineering teams benefiting most from broader access to institutional knowledge and automated assistance. The model is moderately defensible: frontier models and agent frameworks are widely available, but Everest’s accumulated organizational context and its integrations create company-specific capabilities that are harder to replicate quickly. Replicating the prototype-first coordination and basic AI tooling is accessible to many firms, but reproducing the knowledge graph, cross-system retrieval, and embedded agent workflows requires substantial internal engineering and organizational change. The broader signal is that proprietary organizational context and workflow integration may matter more strategically than access to baseline coding models alone.

Retrieve the case and more (including their sources) on the AI in USE website

Thank you for being part of the AI in USE community! 🌟

🔗 Visit our website AI in USE to explore the full library of AI use cases and discover how AI is transforming industries across the globe:

Explore 100+ real-world use cases

✨ We hope these AI use cases spark inspiration!
💬 Let us know which one intrigued you the most—your feedback helps us grow!

Disclaimer: This content was (obviously 😉) built with the assistance of AI.

This week's cases answer that question at three noticeably different depths. HackerRank's Chakra removes the human interviewer entirely from first-round screening and keeps people involved only to review flagged evidence after the fact, which is as close as this week gets to a product owning a full outcome. Airbnb's support assistant offers something the quarterly report said was rare, a measured cost number rather than a belief, with support cost per booking down 16% year over year, matching the broader finding that AI's proven value is still landing on costs rather than revenue. TikTok's Shopping Assistant moves earlier in that same arc: it changes where a purchase decision happens, inside the content feed rather than on a separate storefront, but for now it comes with a market projection rather than a result tied back to the assistant itself, a reminder that embedding AI into a workflow and proving what it did are two separate steps.

HackerRank – Chakra Autonomous AI Interviewer

Organization Type: Scale-up
AI Purpose: Automate
Type of AI Model: Generative AI
AI Application Type: HR
Targeted Industry: Multi-industry
Target Group (AI User): HR teams, Recruiters

Use Case Description:
First-round technical interviews are expensive to run at scale: every candidate requires a live interviewer, scheduling coordination, real-time evaluation, and post-interview documentation. For companies hiring large volumes of software engineers or technical specialists, this bottleneck slows pipelines and consumes significant engineering and recruiter time. HackerRank built Chakra to conduct these interviews autonomously, without a human interviewer in the room. A recruiter provides a job description, and Chakra generates a structured interview plan, runs a voice-based conversational interview, adapts its questions based on candidate responses, and delivers a scored report with evidence tied to specific answers. Candidates complete coding exercises inside an integrated development environment, and the system also evaluates how they reason through problems and work alongside AI tools, not just whether their code is correct. Hiring teams receive a structured assessment report rather than raw notes, and review outcomes rather than conduct every screening session themselves. By October 2026, the product had run more than 500,000 interviews during its beta period, consolidating what had previously been three separate screening stages into a single AI-conducted session.

Key Features:

• Recruiters input a job description and receive a generated interview plan with competency sections, questions, and configurable scoring rubrics.

• A voice-based AI conducts the full interview without a live human present, asking follow-up questions and probing incomplete or shallow answers in real time.

• Candidates write and run code inside an integrated development environment during the interview, allowing the system to observe reasoning and process rather than only final output.

• The system evaluates how candidates define problems, direct AI coding tools, and verify or defend AI-generated answers, producing signals that correctness-only tests do not capture.

• A reporting agent scores candidate performance against predefined criteria, generates competency summaries, and links each score to supporting evidence from the interview transcript.

• Webcam, screen, and device signals flag potentially unauthorized assistance, with recorded evidence available for recruiter review before any interpretation.

Results:

• Conducted more than 500,000 AI interviews during the beta period, reported October 2026.

• Achieved an average candidate satisfaction rating of 4.8 out of 5, as reported by HackerRank.

• Reduced suspicious-activity flags by 70 to 80 percent compared to traditional assessments.

• Consolidated three separate screening stages (recruiter screen, take-home assessment, technical interview) into a single AI-conducted session.

Launch Date: February 2026

Retrieve the case and more (including their sources) on the AI in USE website

Airbnb – Inside-Out AI Transformation

Organization Type: Tech giant
AI Purpose: Augment
Type of AI Model: Generative AI, Unsupervised / Reinforcement learning
AI Application Type: Operations, User Experience
Targeted Industry: Travel & Tourism
Target Group (AI User): Consumers, Employees, Customer support agents, Hosts

Use Case Description:
Airbnb operates a global accommodation marketplace where the scale and diversity of available properties create significant friction for both guests and hosts. Travelers must navigate thousands of listings, translate personal preferences into rigid search filters, interpret lengthy reviews, and compare properties with different amenities and trade-offs before making a booking decision. After booking, routine questions, cancellations, and reservation changes generate substantial customer support demand, traditionally requiring costly human intervention. Hosts face a different challenge: creating compelling listings, optimizing prices, and understanding performance without necessarily having the time or expertise to do so effectively. To address these frictions, Airbnb is embedding AI directly into the guest and host experience. Travelers can search using natural language or voice, access AI-generated summaries of listings and reviews, and compare saved properties side by side. An AI-powered support interface answers reservation-specific questions and handles actions such as changing dates or canceling bookings without human intervention. Hosts use AI tools to create listings, identify pricing opportunities, and interpret earnings. The strategic objective is to reduce the effort required to discover and book accommodation, improve booking confidence and conversion, lower customer support costs, and make hosting more accessible and efficient at scale.

Key Features:

• Allows travelers to describe what they need in ordinary language or by voice, converting casual search inputs into relevant accommodation filters.

• Generates summaries of listings and reviews to highlight key property attributes.

• Compares saved homes side by side to support booking decisions.

• Answers reservation-specific questions in chat.

• Handles actions such as date changes and cancellations in chat.

• Helps hosts create listings, evaluate pricing opportunities and interpret earnings performance.

Results:

• Roughly 50% of support tickets were resolved by AI in October 2026.

• AI support reached 45% of support ticket resolution in Q2 2026.

• Support-related cost per booking decreased approximately 16% year over year in Q2 2026.

Launch Date: April 2025

Retrieve the case and more (including their sources) on the AI in USE website

TikTok – AI Shopping Assistant

Organization Type: Tech giant
AI Purpose: Augment
Type of AI Model: Generative AI
AI Application Type: User Experience
Targeted Industry: Retail & E-commerce, Media & Entertainment
Target Group (AI User): Consumers

Use Case Description:
TikTok launched Shopping Assistant, a conversational AI product embedded directly inside the TikTok app, to close the gap between product discovery and purchase. On social platforms, users regularly encounter products through video and creator content but then abandon the journey when they need to search elsewhere for sizing, specifications or availability. Shopping Assistant places a chat interface on the product page within TikTok’s in-app browser, where shoppers ask plain-language questions and receive answers drawn from merchant-supplied information, without leaving the app. The assistant tracks the thread of a conversation, so a shopper who asks about materials, then shipping time, then returns policy does not need to repeat context between questions. TikTok pairs this capability with Buy Direct, a separate one-click native checkout with saved payment details and in-app order tracking, giving merchants a continuous path from discovery to transaction entirely within TikTok. The strategic logic is straightforward: TikTok already generates high purchase intent through its recommendation engine, and Shopping Assistant converts that intent into sales by removing the information gaps that cause users to leave.

Key Features:

• Shoppers ask natural-language questions about product specifications, materials, sizing, availability and delivery directly on the product page inside TikTok’s in-app browser.

• The assistant retains conversational context across a session, so follow-up questions receive relevant, connected answers without the shopper repeating earlier information.

• All responses are grounded in merchant-supplied product data, reducing the risk of answers disconnected from the actual listing.

• From the assistant interface, shoppers continue to checkout on the merchant’s website while remaining inside TikTok’s in-app browser, preserving the session.

• Buy Direct enables one-click purchasing with saved payment methods and addresses, plus in-app order tracking, as a complementary native checkout path.

• Merchant integration is supported through partnerships with Shopify, Salesforce, Shoplazza and Stripe, with Buy Direct using a unified commerce protocol for interoperability.

Results:

• TikTok Shop is projected by EMARKETER to reach approximately $23.41 billion in US gross merchandise value in 2026, reflecting the scale of the broader commerce environment into which Shopping Assistant is now embedded.

Launch Date: October 2026

Retrieve the case and more (including their sources) on the AI in USE website

Thank you for being part of the AI in USE community! 🌟

🔗 Visit our website AI in USE to explore the full library of AI use cases and discover how AI is transforming industries across the globe:

Explore 100+ real-world use cases

✨ We hope these AI use cases spark inspiration!
💬 Let us know which one intrigued you the most—your feedback helps us grow!

Disclaimer: This content was (obviously 😉) built with the assistance of AI.

TypeSafe AI – Jev AI Decision Layer

Organization Type: Start-up
AI Purpose: Augment
Type of AI Model: Unsupervised / Reinforcement learning
AI Application Type: R&D / Product development
Targeted Industry: Software & Technology
Target Group (AI User): Software engineers, Data & analytics

Use Case Description:
AI agent pipelines typically route every decision through the same frontier language model used for generation, whether that decision is complex reasoning or a simple question like “which tool should run next?” or “is this action safe?” Sending those bounded questions to a full generative model adds latency, token cost, and unpredictable string output that downstream code must parse and validate. TypeSafe AI built Jev to sit inside these pipelines as a dedicated decision layer: engineers define the possible outputs in advance, supply application state and a set of bounded questions, and receive typed answers with calibrated probability scores rather than generated prose. Jev does not replace the frontier model doing actual reasoning or writing; it handles the much larger number of smaller semantic judgments surrounding those calls. Software engineers and AI platform teams at companies like Browserbase integrate Jev for tasks such as selecting browser actions, routing agent requests, running safety checks before tool execution, and scoring production traces at scale. In Browserbase’s Stagehand implementation, replacing generative model calls for bounded action selection reduced median latency from 1.97 seconds to 0.46 seconds. The strategic bet TypeSafe is making is that most agent compute today is misallocated: frontier models are doing classification work that a faster, cheaper, purpose-built decision model can handle more reliably.

Key Features:

• Accepts typed decision schemas before inference so outputs are always one of the predefined valid values, eliminating parse errors and schema violations in downstream code.

• Supports three decision primitives: choosing among predefined alternatives, scoring against an ordered rubric, and returning a calibrated probability for a yes/no proposition.

• Evaluates multiple independent questions over the same application state in a single call, reducing sequential model invocations and their cumulative latency.

• Attaches a confidence measure to each output so application code can route high-confidence decisions to automation, mid-confidence decisions to verification, and low-confidence decisions to a human or a stronger model.

• Integrates with agent frameworks including LangChain and Vercel AI Gateway, allowing teams to add Jev as a routing or guardrail step without rewriting existing pipelines.

• Classifies proposed agent tool calls before execution to flag potentially unsafe actions without requiring a separate generative model call.

• Scores or classifies production agent traces for evaluation, enabling teams to assess a much larger share of live traffic than sampling-based LLM-as-a-judge approaches permit.

Results:

• Browserbase reduced median browser automation action latency from 1.97 seconds to 0.46 seconds by routing bounded decisions through Jev with LLM fallback below a 0.7 confidence threshold.

• LangChain’s independent evaluation recorded a per-call cost of $0.00035 and an average latency of 0.44 seconds for agent trace classification, versus $28.17 total cost for the equivalent Claude Sonnet 4.6 run.

• LangChain measured Jev’s output variance at 92 to 913 times lower than GPT-5.6 and Claude Sonnet 4.6 configurations on repeated identical inputs, indicating substantially more consistent decisions.

• Jev reached nearly 13% of paid Vercel AI Gateway teams within 24 hours of launch, more than twice the first-day adoption rate of the GPT-5.6 family on the same platform.

• One independent developer classified 1,018 AI research papers into 24 categories for $0.08 in inference cost at a 256-millisecond median latency per paper.

• A Vercel engineer replacing a GPT-5.6 Luna safety classifier with Jev observed a 5 to 18 times reduction in latency on the same task.

Launch Date: September 2026

Retrieve the case and more (including their sources) on the AI in USE website

Anthropic – 950-Agent Genome-Mining Workflow

Organization Type: Scale-up
AI Purpose: Augment
Type of AI Model: Generative AI, Unsupervised / Reinforcement learning
AI Application Type: R&D / Product development
Targeted Industry: Biotechnology, Pharmaceuticals, Software & Technology
Target Group (AI User): Researchers

Use Case Description:
Genomic databases have grown far beyond what research teams can inspect manually, while conventional computational pipelines tend to search for patterns scientists already expect. Anthropic tested a different approach: using Claude agents as a high-throughput discovery layer. Over 21 hours, roughly 950 Claude agent sessions explored a database of about 1.9 billion protein clusters, screening more than 200,000 reverse-transcriptase candidates and reducing them to around 20 human-readable reports for expert review. The agents planned analyses, ran database queries, searched the literature, and even followed side investigations autonomously. One of those side investigations surfaced an unexpected repeat-array structure next to a known enzyme. Anthropic’s computational and molecular biologists reviewed the finding and validated it experimentally, confirming that the arrays are expressed as distinct RNAs. The key product insight is not that AI replaces scientists. It changes where their scarce time is spent: instead of weeks mining genomes manually, researchers can focus on judging a ranked shortlist and deciding what deserves an experiment. That points to a broader thesis for Claude Science: agentic AI as a high-throughput triage and exploration layer upstream of human scientific expertise.

Key Features:

• Agents searched a database containing roughly 1.9 billion protein clusters.

• The system planned analyses and executed follow-up work without requiring each step to be specified by hand.

• An unexpected repeat-array structure was surfaced during a side investigation.

• Claude measured the repeats, compared them with known systems, and searched prior literature before escalating the finding.

• More than 200,000 reverse-transcriptase candidates were reduced to roughly 3,500 systems and then to about 20 reports for expert review.

• Candidate evidence was packaged into human-readable artifacts for scientific review.

• Scientists performed the wet-lab validation on the shortlisted candidates.

Results:

• 949 agent sessions completed the campaign in approximately 21 hours.

• More than 200,000 reverse-transcriptase clusters were screened and reduced to roughly 3,500 candidate systems.

• Around 20 detailed reports were produced for human review.

• The campaign identified ART, a previously uncharacterized array-associated reverse-transcriptase system.

• In one dataset, array-derived RNAs represented up to 8% of phage RNA at 15 minutes post-infection.

Launch Date: September 2026

Retrieve the case and more (including their sources) on the AI in USE website

Meta – Muse Personal AI Agent

Organization Type: Tech giant
AI Purpose: Automate
Type of AI Model: Generative AI
AI Application Type: User Experience
Targeted Industry: Software & Technology, Retail & E-commerce, Financial Services
Target Group (AI User): Consumers

Use Case Description:
Consumers still spend significant time moving between apps, websites and services to research products, fill forms, book appointments, manage email and complete purchases. Most AI assistants can help with parts of those workflows, but stop before actually completing them. Muse is Meta’s consumer agent designed to close that gap. Users give it a goal, and the agent plans and executes the workflow across websites and connected services: browsing, comparing options, filling forms, managing communications and completing transactions. It operates inside its own cloud browser, so tasks can continue after the user closes the app, with Muse returning only when input or approval is required. Meta distributes Muse through iOS, Android, web and WhatsApp, with additional integrations planned across its device ecosystem. Connections to services including email, calendars and payment providers allow it to move beyond answering questions toward actually completing tasks. The broader product thesis is significant: if an agent captures the user’s intent and executes the workflow on their behalf, the agent—not the merchant, marketplace or individual app—can become the primary interface through which consumers interact with digital services. Muse is therefore less a better chatbot than an attempt to build an execution layer for everyday digital life.

Key Features:

• Executes multi-step tasks autonomously in a dedicated cloud environment, continuing in the background after the user closes the app.

• Navigates websites and fills forms through a built-in browser, reaching services that have no dedicated connector.

• Connects to email, calendars, shopping platforms and payment rails so the agent can pull information and act across services in a single workflow.

• Completes purchases end-to-end using Stripe Link, Shop Pay and PayPal, with one-time card numbers that keep the user’s underlying payment details away from the agent.

• Asks for explicit user approval before sensitive actions such as sending emails or finalising purchases, and exposes a full activity trail the user can review.

• Stores user preferences and goals so it can proactively suggest relevant tasks without requiring a fresh prompt each time.

• Produces task-appropriate outputs such as itineraries and dashboards rather than forcing every result into plain chat text.

• Surfaces suggested use cases proactively during onboarding to help users understand what the agent can do on their behalf.

• Runs inside WhatsApp and AI glasses in addition to its standalone app, placing the agent inside surfaces consumers already open many times a day.

Results:

• Reached more than 500,000 users and more than 250,000 daily active users within about one week of launch.

• Generated more than 2 million prompts shortly after launch.

• Reached more than 3.4 million downloads within the first three weeks.

• Reached the top of the free-app charts in the U.S. and Canada.

Launch Date: September 2026

Retrieve the case and more (including their sources) on the AI in USE website

Thank you for being part of the AI in USE community! 🌟

🔗 Visit our website AI in USE to explore the full library of AI use cases and discover how AI is transforming industries across the globe:

Explore 100+ real-world use cases

✨ We hope these AI use cases spark inspiration!
💬 Let us know which one intrigued you the most—your feedback helps us grow!

Disclaimer: This content was (obviously 😉) built with the assistance of AI.

What does it take to build real-world AI that people actually trust?

Benjamin’s path into AI did not begin with hype. It began with physics, natural science, and a PhD that trained him in analytical thinking, hypothesis-driven work, math, and programming. After academia, he wanted to apply those skills to real-world problems, and data science became the natural bridge.

That first chapter matters, because it explains why his perspective on generative AI is grounded rather than hype-driven. He didn’t make the shift toward GenAI because he sees mystique - under the hood, it’s matrix multiplication at massive scale - but because he sees where it becomes useful when paired with rigor, guardrails, and product thinking.

He first worked on internal sales and presales use cases at Criteo, before moving to Stuart to work on customer-facing problems. That transition shaped his view of AI in a fundamental way. Internal use cases and customer-facing AI may rely on similar underlying technologies, but they do not operate under the same constraints. Once AI is exposed directly to the user, the tolerance for approximation drops sharply. The expected level of robustness becomes much closer to software engineering: systems need to be reproducible, reliable, and controlled enough to deserve trust.

That is where the complexity of client-facing AI really begins. AI is probabilistic, not rule-based. So the challenge is not only to make it work once, in a promising demo or internal prototype. The challenge is to shape its behavior, put in guardrails, reduce failure modes, and make the output dependable enough for people to use confidently in real workflows. When the product touches the customer directly, “good enough” stops being enough. Trust becomes part of the product requirement.

That lens runs through the entire conversation with Benjamin: AI is not difficult only because of model quality. It is difficult because production-grade, client-facing AI has to meet a much higher bar.

🛠️ What he’s building

Today, Benjamin leads an AI Engines team at Qonto, where he works on customer-facing machine learning and generative AI solutions. The broader mission is to build the AI that puts customers’ finances on autopilot so SMEs, freelancers, and founders can spend less time managing finance and more time building their business.

For the AI Engines teams, that mission translates into a very concrete product goal: create user flows that feel seamless, with as few touchpoints as possible and errors approaching zero. AI comes into play where traditional programmatic solutions start to hit their limits.

What stood out to me is how disciplined Benjamin’s approach is. He does not start with AI for AI’s sake. He starts with the customer need.

From there, the sequence is clear: first identify the user problem, then ask whether AI is truly needed or whether a simpler heuristic would do, check the effort-versus-value tradeoff, and only then move into a proof of concept.

That POC step is central in how he works. In his view, quick POCs matter for two reasons. First, the technology evolves so quickly that what did not work one month ago may work now. Second, they are one of the fastest ways to get real user feedback, even through something as lightweight as a thumbs up or thumbs down signal. It is not prototyping for show. It is a way to test feasibility, reduce risk, and learn quickly before investing too heavily.

🧩 AI challenges & pain points

When asked about the unresolved AI challenge that keeps him thinking, Benjamin’s answer was not primarily about the frontier capabilities of the models themselves.

It was about adoption.

Not everyone adopts AI at the same pace. Some people are natural early adopters; others are much more hesitant. And one of the real challenges, in his view, is helping people understand that these tools can unlock productivity gains across tech and digital work - not just by accelerating execution, but by saving time and energy for deeper thinking.

He also pointed to a second challenge that follows naturally from that: prioritization. There are now so many possible AI opportunities that teams can quickly lose focus. The risk is not a lack of ideas. It is too many of them. That is why his framework matters so much: start with a concrete customer need, determine whether AI is truly necessary, weigh effort against value, then test fast.

Fast POCs matter here for another reason too. Benjamin noted that the technology evolves so quickly that something which did not work some months ago may now work. A POC is therefore not just a product tool, but a way to stay in sync with a fast-moving technical landscape. It also creates the shortest path to user feedback, which is essential when deciding what deserves to go further.

So the challenge is not simply “how do we build more AI?” It is: how do we help people adopt it, and how do we stay disciplined enough to choose the right things to build?

🚀 The use case

One project Benjamin’s team has been investing in is a document extraction solution built with the ambition to read, understand, and extract data from every document that enters Qonto’s platform.

It started with a very concrete problem. For SMEs, freelancers, and founders, documents are everywhere - and in finance workflows, those documents are often a source of critical information for downstream actions. A supplier invoice, for example, contains details like an IBAN or an amount that should not require manual copying and pasting. But also for more complicated documents like bank statements or company bylaws, extracting information like account balances, transaction details, or legal entity data is just as much a user problem worth solving, even if these documents are less standardized and frequent. The smoother that extraction can be, the greater the time and cognitive load customers’ save - which is where the product’s real value lies, beyond simplification alone.

At first glance, OCR might sound like a solved problem. Plenty of vendors offer it. But Benjamin’s explanation shows why “solved” is often misleading in product reality.

The real issue was the tradeoff between accuracy and latency.

Qonto needed both. Accuracy was critical because these are trust-sensitive workflows: if the extracted information is wrong, the experience stops feeling helpful and starts feeling risky. But low latency was also essential. In many workflows, waiting tens of seconds for a document to be analyzed isn’t acceptable to users - and expecting both perfect accuracy and near-instant latency at the same time isn’t realistic. At Qonto’s scale, that tradeoff has to be actively managed, not treated as marginal.

According to Benjamin, many market OCR solutions achieve strong accuracy by stacking complex mechanisms, but that often increases latency. Because this capability was so central to Qonto’s value proposition, the team chose to build the OCR system in-house. LLMs are a part of this strategy. The result is a system able to handle large volume while delivering both high accuracy and the right latency budget for the use case.

What I like about this example is that it reflects Benjamin’s broader lens on AI. He is not chasing novelty. He is looking for the point where AI can create a better product experience than deterministic systems or packaged tools.

📈 Trends & misconceptions

We also talked about what excites Benjamin most in AI right now.

The first theme is speed. He pointed out how much faster development has become, including in core ML engineering workflows. Tasks that once required manually plotting charts during exploration can now begin with a single prompt that generates multiple views, and even suggests new directions to investigate. The same dynamic applies more broadly to model optimization and ML Ops pipelines, which are becoming faster and more accessible than before.

The deeper point is not just time saved. It is leverage. AI lets people explore more, test more, and spend more of their energy on thinking rather than execution.

The second theme is how job descriptions are changing. Benjamin mentioned that anyone in the company today, regardless of role, can create a pull request. In his view, AI is pushing many roles toward a more full-stack reality, and companies that do not adapt risk falling behind. He expects most will move in that direction within a year.

The third is agentic engineering, where AI systems handle multi-step engineering tasks autonomously. He sees real promise there, but organized adoption and training are key, especially to grow best practices and avoid duplicated effort as everyone builds their own agents in parallel. What interests him is the idea of reusable skill palettes that reduce reinvention and make capabilities more portable across contexts.

💭 Reflections & future visions

To close, I asked Benjamin about the most important mindset for working in AI today.

His answer was simple: be curious. And try.

That simplicity is probably what makes it powerful. Things are changing quickly. AI-native workers will enter the market with very different reflexes, and waiting passively is not much of a strategy. Benjamin’s instinct is to lead, experiment, and learn through action rather than stand still.

What stayed with me after this conversation is the combination he embodies: scientific rigor, product clarity, and openness to new tools without naïveté about their limits.

His story is a reminder that the hardest part of AI is often not the model itself. It is everything around it: adoption, prioritization, usability, guardrails, and trust.

🔗 Follow Them

You can follow Benjamin on LinkedIn

Thank you for tuning in to this edition of the AI Builders Interview series! 🎙️

🔗 Dive deeper into the minds behind AI innovation—explore more conversations with the builders shaping the future, and check out 200+ real-world AI case studies in our growing library:

Discover more interviews of AI builders

✨ We hope this dialogue offered fresh insights and sparked new ideas!

🛠️ Know an AI builder doing great work? Nominate them for a future interview!

Disclaimer: This content was (obviously 😉) built with the assistance of AI.

Salesforce – AI Sales Workspace in Claude

Organization Type: Tech giant
AI Purpose: Augment
Type of AI Model: Generative AI
AI Application Type: Sales & Marketing
Targeted Industry: Multi-industry
Target Group (AI User): Sales reps, Operations teams

Use Case Description:
Salesforce and Anthropic built a joint product that embeds a full sales workspace directly inside Claude, with Salesforce serving as the governed system of record underneath. The core problem it addresses is fragmented selling context: account history, open opportunities, Slack threads, call transcripts, and emails all live in separate places, forcing sellers to spend hours manually assembling information before and after each meeting. Inside Claude, a seller states an outcome (”prep me for this call,” “what deals need attention this week”) and the product retrieves live Salesforce and Slack data, reasons across it, and returns a structured briefing, a pipeline view, or a drafted follow-up, all within a single interface. Any proposed change to a Salesforce record, whether updating a close date, logging a call, or adding a contact, is shown to the seller for approval before being written back, with Salesforce’s existing permission and audit controls applying throughout. At Legora, roughly one in five interactions with the system results in a CRM write-back action, suggesting sellers are treating the product as an operational interface rather than a search tool. Salesforce reports approximately 7,000 of its own sellers use the product internally, with GitLab, Siemens, Legora and Deloitte named as external pilots.

Key Features:

• Provides 37 prebuilt sales workflows for account research, meeting preparation, pipeline review, renewals, and CRM updates.

• Generates a personalized daily sales brief with meetings, at-risk deals, and unread threads.

• Combines Salesforce, Slack, and email context into a single meeting briefing.

• Drafts follow-up emails, Slack summaries, and CRM updates after calls.

• Lets users log calls, create tasks, and update records through natural language with approval before write-back.

• Shows pipeline and forecast views with deal risk and team summaries.

Results:

• Approximately 7,000 Salesforce internal sellers are reported as active users of the product.

• At Legora, nearly 20% of all tool interactions result in a CRM write-back action, indicating active use of the product as an operational interface rather than a read-only assistant.

• Automated call-note creation at Legora has eliminated manual note-writing by sales representatives after calls.

• Legora’s CFO reports that meeting preparation that previously took hours now takes seconds, based on live Salesforce data retrieved at query time.

Launch Date: September 2026

Retrieve the case and more (including their sources) on the AI in USE website

Transportation Security Administration (TSA) – Ace, AI Traveler Support Assistant

Organization Type: Public
AI Purpose: Augment
Type of AI Model: Generative AI
AI Application Type: Customer Care
Targeted Industry: Government / Public Administration, Travel & Tourism
Target Group (AI User): Customer support agents, Consumers

Use Case Description:
TSA handles millions of traveler inquiries each year, the vast majority of them repetitive questions about permitted items, accepted identification, screening procedures, and medical devices. Before Ace, those questions required either a contact-center agent or a traveler willing to search through TSA documentation, both slow under peak-travel conditions. TSA first consolidated its customer-service data and workflows into a unified platform, then deployed Ace on top of that foundation as a conversational assistant available through TSA.gov and the myTSA app. Ace answers traveler questions by drawing on TSA guidance and a library of more than 550 knowledge articles, rather than generating responses from general training data alone. When a question falls outside its defined scope, Ace transfers the conversation to a human agent along with an automatically generated summary of the interaction, so the agent picks up without having to reconstruct context. For TSA, the strategic logic is clear: scale self-service during demand spikes without increasing headcount, and preserve human capacity for the cases that actually require it.

Key Features:

• Travelers ask natural-language questions about screening rules and receive answers grounded in TSA guidance and a 550-plus-article knowledge base, without navigating documentation.

• The assistant operates around the clock across TSA.gov and the myTSA app, handling routine inquiries at volume during peak travel periods.

• Conversations that exceed the assistant’s defined scope are automatically transferred to a live TSA agent, keeping harder or sensitive cases in human hands.

• Each escalated conversation arrives with an AI-generated summary of the prior exchange, allowing the receiving agent to respond immediately without reconstructing context.

• Contact-center staff access a unified view of traveler interactions across email, phone, and social channels through the underlying Customer Service Managers App.

Results:

• Ace handles approximately 100,000 traveler conversations per month.

• 96% of routine inquiries are resolved without escalation to a human agent.

• Interaction cost is reduced by more than 90% compared to agent-handled contacts.

Launch Date: June 2026

Retrieve the case and more (including their sources) on the AI in USE website

Fountain 0 – AI-Native Film Production Engine

Organization Type: Start-up
AI Purpose: Augment
Type of AI Model: Generative AI
AI Application Type: R&D / Product development
Targeted Industry: Media & Entertainment
Target Group (AI User): Movie directors and producers

Use Case Description:
Fountain 0 is building a production engine that lets a writer-director work alone, or nearly alone, to produce a feature-length film without physical sets, cameras, or a conventional crew. The cost of making a visually ambitious narrative film has historically been prohibitive for independent creators: even modest productions require casting, locations, equipment, VFX vendors and scheduling coordination that push budgets into the millions. Fountain 0’s workflow replaces that physical execution layer with a proprietary orchestration system layered over generative video models, primarily Kling 3.0, which the company selected for its stronger character consistency and face rendering. The creator writes and directs as usual, then uses the engine to generate, review and iteratively regenerate individual scenes before assembling them into a finished cut. Ash Koosha wrote and directed Odysseus: The Fall this way, completing a 150-minute feature in roughly four months of part-time work at a reported out-of-pocket cost of approximately $25,000, a figure that reflects the workflow’s compression of the production stack rather than the replacement of creative authorship.

Key Features:

• Text-to-video scene generation replaces physical shooting, enabling locations, creatures and action sequences without sets, cameras or location permits.

• A proprietary production orchestration engine coordinates pre-production, scene generation and post-production so the creator directs iteratively within a single environment rather than operating disconnected tools.

• Iterative regeneration and shot selection allow the filmmaker to reject, revise and reassemble generated footage using conventional film grammar.

• Character consistency controls, built around Kling 3.0’s rendering capabilities, address one of long-form generative video’s hardest problems: maintaining a recognizable character across an entire feature.

• A licensed digital likeness system allows real individuals to appear in productions under separate, project-specific agreements, with 13 licensed likenesses incorporated into Odysseus: The Fall.

• A direct-to-consumer storefront distributes the finished film without theatrical or streaming intermediaries, at $9.99 per rental with 4K UHD delivery.

• A likeness directory, currently in development, allows individuals to register and be matched to productions, with actual likeness use governed by a separate licensing contract.

Results:

• Odysseus: The Fall was completed in approximately four months of part-time work by a single principal creator, against a reported out-of-pocket production cost of approximately $25,000.

• The finished film runs 150 minutes and is commercially distributed in 4K UHD, demonstrating that the workflow can produce long-form output rather than short clips.

• 13 real individuals licensed their likenesses for inclusion in the released version.

• The company’s first feature, Dreams of Violets, produced for approximately $2,000, was officially selected for Tribeca 2026, representing the first fully AI-generated live-action feature accepted into that festival’s program.

Launch Date: September 2026

Retrieve the case and more (including their sources) on the AI in USE website

Thank you for being part of the AI in USE community! 🌟

🔗 Visit our website AI in USE to explore the full library of AI use cases and discover how AI is transforming industries across the globe:

Explore 100+ real-world use cases

✨ We hope these AI use cases spark inspiration!
💬 Let us know which one intrigued you the most—your feedback helps us grow!

Disclaimer: This content was (obviously 😉) built with the assistance of AI.

For this issue, I went through 38 reports and working papers published in Q2 2026 to find where the actual consensus sits.

My conclusion is simple: the story has moved from whether the technology works to whether the organization around it has done the harder, less visible work. The technology keeps clearing benchmarks. What separates the companies capturing real value from the rest is decision rights, patient capital, workforce enablement, and governance built into the system rather than bolted on.

Executive summary

• The gap, not the growth rate: 69% of CEOs say AI is already changing their core business, but only 10% credit advanced agentic systems as the primary driver of revenue growth (IBM’s 2026 CEO Study). That gap is real. It’s narrower than “AI isn’t delivering,” though: this quarter’s evidence shows AI’s realized value landing on the cost side, not the top line.

• Winning now comes with financial-market evidence: AI-first firms report a 17% revenue-growth premium (IBM). In insurance specifically, top performers show 21% revenue growth and 51% greater share-price gains over three years (Capgemini). Different populations, not one number, but the direction repeats.

• The named blocker has shifted: since 2023 the standing barrier was data readiness. This quarter’s reports increasingly point to “org debt” (legacy roles, siloed decision rights, unclear accountability) as the harder constraint once basic data access stops being the excuse.

• Organizational absorption, not tool access, separates winners: multi-year investment horizons instead of per-pilot ROI tests, and roughly 14% of AI budget on workforce enablement versus 8% among peers (BCG).

• Physical infrastructure is now a headline, not a footnote: 2026 AI infrastructure capex is projected near $700B, up from roughly $410B in 2025, with global data-center electricity demand approaching 1,000 TWh by 2030 (WEF).

• Model convergence is now numeric: the gap between Anthropic’s leading model and the top Chinese model (DeepSeek-R1) narrowed to roughly 2.7% by March 2026 (Stanford AI Index). Hallucination rates across top models still range from 22% to 94%, though. Convergence and reliability are not the same axis.

• The open question since 2022 is still open: whether AI’s returns will broaden or concentrate. This quarter’s evidence sharpens the case for concentration, without proving it’s permanent.

What changed this quarter

Three things moved between Q1 and Q2 2026, and it is worth separating a sharper description of an old problem from a genuinely new one.

The maturity gap now has a sharper number attached. Where Q1 2026 measured it as 38% “scaled” against 11% “top maturity,” this quarter frames it as belief against attribution: 69% of CEOs say AI is changing their business, but only 10% credit it for their revenue growth. That word matters. The same material also shows an 11% operating-margin gain for companies that pair AI with change management, and only 22% of organizations even measuring AI’s revenue impact at all, against much higher shares tracking productivity and time savings. Automation usually moves the bottom line before the top line. So the 10% figure may say less about AI underperforming than about which line this quarter’s reports are actually measuring. The underlying gap between belief and results is not new; it traces back to the 2022 “Achievers” data.

This is also the first quarter to tie AI maturity to an adopting company’s own share price. Financial-market figures aren’t new to this series: 2022 tracked AI-company valuations recovering, and 2024 noted NVIDIA’s market cap and public AI companies’ combined enterprise value. Those numbers were always about the AI industry’s own valuation, though. Every prior maturity-premium claim, from 2022’s Achievers to 2024’s Pioneers, used revenue growth or EBIT as the outcome metric. Using an adopting company’s stock performance as the return on its own AI maturity is new. It’s a meaningful upgrade in evidence, even if it currently comes from a single sector (insurance) rather than a broad cross-industry sample.

Physical infrastructure also moved up the agenda. Energy, water, minerals, and land used to sit as a secondary constraint alongside compute. Now they form a strategic category of their own, with a named framework (“Nexus Strategy”) and real capex figures behind it. This is an escalation of a trend visible since the 2025 Q4 “power over chips” framing, but the scale and specificity are new this quarter.

What is continuation vs what is genuinely new

Several threads here are simply continuing. The maturity gap has been visible since 2022, when only 12% of firms counted as “Achievers.” It hardened into the GenAI Divide (5% versus 95%) by 2025, then became 38% scaled against 11% top-maturity in Q1 2026, and now shows up as this quarter’s 69%/10% revenue-attribution gap. The data-readiness barrier is the same story: lakehouses in 2023, 9% full data access in 2025, 18% high readiness in Q1 2026, and now just 4% “fully prepared” for unstructured data. That barrier has simply been renamed and broadened under the “org debt” label, from a data problem into decision rights, legacy roles, and accountability. The idea that operating model matters more than technology is also old, running from the 10-20-70 rule in 2024 through the Agentic Organization in 2025 to orchestration in Q1 2026. Multi-year investment horizons were Q1 2026’s own headline finding, not a Q2 development. And governance embedded directly into agent workflows has a history too: 2025 sketched it as “guardrail” and “critic” agents, Q1 2026’s dominant framing tilted toward “read-only” AI instead, and this quarter swings back toward embedding controls in the system as agent decision volume scales.

Three things are genuinely new this quarter. Winners now show a specific budget split: roughly 14% on workforce enablement versus 8% among peers, where prior quarters only offered a broad “invest more, invest longer” framing. There’s a numeric figure for geopolitical convergence (2.7%) where earlier quarters gave only qualitative claims about US-China standing. And an adopting company’s own share price, not revenue growth or EBIT, gets used for the first time to measure the AI-maturity premium.

Some claims just repeat. “Data is the top barrier” has been said, almost word for word, since 2023. General statements that AI is a strategic imperative add little. The skills gap got relabeled a “knowledge gap” this year, but without much new evidence behind the rebrand.

Trend 1: AI strategy and operating models

Signal: the defining strategic question this quarter is no longer “which AI use cases should we launch” but “what should the system be allowed to do.” That reframes strategy as decision architecture: what can be delegated, what is reversible, what requires approval, what confidence level is enough, who owns an automated decision, and what happens when the system hits an exception. CEOs still rank productivity and profitability as the top priority. The “rewiring the C-suite” language now comes with concrete mechanics behind it, too. Buy-vs-build is hardening into a simple rule: buy commodity AI, build or fine-tune anything meant to be a moat. And 45% of executives report using a multi-vendor strategy specifically to avoid lock-in (IBM).

Historical perspective: 2022 already showed no consensus on centralized-vs-decentralized structure. By 2024 the field had its first real playbook, the 10-20-70 rule. 2025 introduced the Agentic Organization and the Minimum Viable Organization, then Q1 2026 reframed AI as infrastructure and made sovereignty a design constraint. Q2 2026 does not replace that arc. It adds capital allocation and decision rights as concrete, measurable levers.

Limit: the “top quartile” and “trailblazer” framings recur across IBM, Capgemini, and others, but they describe different populations and thresholds. Treating them as one global “10% of firms are winning” statistic flattens data that does not support it.

Product implication: workflow redesign now needs a decision-rights layer built in from the start, not bolted on after an agent misfires.

Trend 2: Adoption, deployment, and ROI

Signal: broad adoption is now near-universal: 88% in the Stanford AI Index. That sits uneasily next to the 10% figure on agentic systems as a primary driver of revenue growth. In financial services specifically, roughly 50-55% are piloting agentic systems, yet only about 40% report profitability gains, and 60-70% still say they cannot cleanly define or measure AI’s value (CCAF). Where value shows up, usage depth separates winners from the rest, not access: 73% six-month usage among top performers versus 36% for the median, alongside a workforce-enablement gap of 14% of AI budget among winners versus 8% among peers (BCG).

Historical perspective: this pattern has held since 2022, sharpening into a longer chain each year: interest outruns adoption, adoption outruns integration, integration outruns value capture. 2023 saw universal experimentation but only 28% enterprise-scale. Then came “one in four cracked the code” in 2024, the 5%/95% GenAI Divide in 2025, and 38% scaled against 11% mature in Q1 2026. Q2 2026’s contribution is definitional: it separates “adoption” from “value capture” more cleanly than any prior quarter, and that matters because the two get conflated constantly in headline statistics.

Interpretation: the winners’ pattern (deeper usage, heavier enablement spend) suggests AI advantage may increasingly compound through organizational learning rather than model access. That remains a hypothesis, not a proven causal relationship. High-performing firms may simply have had stronger data, management, and digital foundations before they adopted AI at all.

Product implication: adoption rate is a vanity metric on its own. Usage depth and value attribution are the numbers worth building dashboards around.

Trend 3: Technology, product, and technical evolution

Signal: SWE-bench Verified is a widely used test of whether an AI system can independently fix real software bugs pulled from live open-source codebases, a proxy for real-world coding ability. Performance on it is now reported near 100% of the human baseline, a benchmark-saturation moment for coding tasks. Hallucination rates tell a different story, though: across 26 top models they still range from 22% to 94% (Stanford). Capability and reliability are visibly not the same axis. Looking toward 2030, foundation-model-only strategies are shrinking, from roughly 39% to 13%, while hybrid strategies (combining large models with fine-tuned small and domain-specific models) rise toward roughly half of organizations. Governance is moving into the architecture itself. Because agents now execute large volumes of decisions, control increasingly means pre-defined boundaries built into the system, not a human sign-off step.

Historical perspective: the technical story since 2022 has been one of successive abstraction layers. 2022’s frontier was general-purpose architectures, diffusion, and scaling laws. 2023 was about production deployment, data access, and early copilots. 2024 shifted toward production value, hosted LLMs, and the economics of deployment. 2025 brought agentic architectures, reasoning, and longer-horizon task execution. Q1 2026 shifted the center of gravity again, toward orchestration over model choice. Q2 2026 continues that same line and adds a physical dimension: compute capacity has grown roughly 3.3x annually since 2022, and the infrastructure now required to sustain it (power, water, minerals) is itself becoming a technical design constraint, not just a cost line.

Caution: this quarter’s material cites water-use estimates at different scopes (a single model’s annual footprint versus the whole industry’s 2030 projection) that are easy to mistake for one comparable number. They are not. They measure different things.

Product implication: choosing a frontier model is no longer the differentiating decision. Verification loops, observability, and governance embedded in the architecture are.

Product perspective

Four things stand out for product teams this quarter.

The central design question has shifted. It’s no longer “where can we add a copilot” but “can the product take responsibility for a meaningful part of an outcome”: collecting information, identifying exceptions, taking permitted actions, updating systems, escalating only when necessary. That means designing five things together, not sequentially: workflow (the actual end-to-end outcome), context (what data and state the system needs), authority (what AI is allowed to decide and execute), control (how confidence, permissions, exceptions, and human intervention are designed), and learning (whether each execution makes the system or organization better).

Operating models are visibly maturing faster than many user-facing products. Some organizations are professionalizing governance and capital allocation while the actual product experience for end users barely changes. That’s the same gap Q1 2026 flagged, now with a financial number attached to it.

The reliability gap deserves more product attention than the capability race. Hallucination rates are still wide and unresolved even as benchmarks saturate. The differentiating engineering work sits inside the Control layer above: verification, monitoring, and embedded guardrails, not swapping the underlying model.

Experimentation and productization are still routinely confused, and this quarter names the gap precisely. Experimentation means people are using AI. Productization means AI is embedded into a repeatable workflow with ownership, integrations, controls, measurement, and economics that survive the pilot. The clearest version of that confusion this quarter is the 88%-vs-10% gap itself, and its budget-line cousin: buying tools without funding workforce enablement (8% vs 14% of spend). Access without absorption is not productization.

Bottom line

This quarter’s most likely meaning: the AI story is less about whether the technology works (on narrow, benchmarked tasks, it increasingly does) and more about whether an organization has done the unglamorous, non-technical work. That means decision rights, patient capital, workforce enablement, and governance built into the system rather than bolted on. Q2 2026 gives that divide a price tag it did not clearly have before.

What remains uncertain is whether that price tag reflects a temporary early-mover advantage that will diffuse over time, the way cloud and ERP adoption cycles did, or a structurally compounding one. This quarter’s reports use the language of self-reinforcing “flywheels,” where proprietary data sharpens agents and agents sharpen people. The reports do not settle which it is. That question has been open since 2022.

Three things are worth watching. Does the financial-market premium (share price, revenue growth) hold up outside single-sector studies like insurance? Does “org debt” become something organizations actually measure, or does it stay a label? And do the physical infrastructure figures (capex, power, water) start functioning as a hard constraint on how fast the winners can pull further ahead?

Sources

Based on 38 reports and working papers published primarily in Q2 2026 (ranging from April to June 2026) by leading global organizations and strategic consultancies—including the World Economic Forum (WEF), International Monetary Fund (IMF), OECD, IBM Institute for Business Value, Stanford Institute for Human-Centered AI (HAI), Bain & Company, Boston Consulting Group (BCG), KPMG, McKinsey & Company, PwC, and Capgemini—these sources provide a rigorous analysis of the global pivot from generative experimentation to industrialized agentic AI.

The sources document a widening “performance chasm,” where a top quartile of “AI-first” organizations is capturing a 17% to 21% revenue growth premium by successfully “rewiring” their C-suite and dismantling functional silos. Key focus areas include the emergence of autonomous agentic workflows in government and finance, the strategic management of the AI-Energy-Water-Minerals-Land nexus, and the rise of AI Sovereignty as a core tenant of both statecraft and corporate strategy. These reports highlight that while technical capabilities continue to accelerate at 3.3x annually, the primary bottleneck to value realization is now “org debt”—the accumulated complexity of legacy processes and fragmented data foundations.

Sample source links and references from these reports include:

• Stanford HAI AI Index 2026 Report: https://hai.stanford.edu/ai-index/2026-ai-index-report

• IBM 2026 CEO Study: Rewiring the C-suite: https://ibm.biz/ceo-2026

• WEF Making Agentic AI Work for Government: https://www.weforum.org/publications/making-agentic-ai-work-for-government/

• Bain & Company Proprietary Intelligence: https://www.bain.com/insights/proprietary-intelligence-how-to-win-with-ai/

• IMF Note: Global Economic and Financial Implications of AI: https://www.imf.org/en/publications/policy-papers/issues/2026/05/global-economic-and-financial-implications-of-ai

• McKinsey State of AI Trust in 2026: https://www.mckinsey.com/capabilities/tech-and-ai/our-insights/tech-forward/state-of-ai-trust-in-2026-shifting-to-the-agentic-era

• Capgemini World Property and Casualty Insurance Report 2026: https://www.capgemini.com/insights/research-library/world-property-and-casualty-insurance-report-2026/

Thank you for being part of the AI in USE community! 🌟

🔗 Visit our website AI in USE to explore the full library of AI use cases and discover how AI is transforming industries across the globe:

Explore 100+ real-world use cases

✨ We hope these AI use cases spark inspiration!

💬 Let us know which one intrigued you the most—your feedback helps us grow!

Disclaimer: This content was (obviously 😉) built with the assistance of AI.

When Sara Maldon joined Make to lead AI transformation, there was no playbook waiting for her.

It was still early in the generative AI wave. ChatGPT had created huge curiosity, but companies were also dealing with security, compliance, procurement and intellectual property questions. At Make, the answer at the time was clear: AI was prohibited.

Sara’s task was not just to introduce new tools. It was to help the company understand how AI could be used safely, practically and meaningfully across the organisation.

Her path into that role was not conventional. Sara started in law, a field she loved intellectually but quickly realised was not the right industry for her. She describes herself as “a builder by heart”: someone who likes moving fast, changing things and rebuilding them better.

That instinct led her to make what she calls a 180-degree career change. She retrained as a software developer, then moved through cybersecurity, data projects and technical project management before landing in AI transformation.

Looking back, the path makes sense. Sara’s work today is not only about technology. It is about helping people cross the gap between interest and real adoption.

🛠️ What They’re Building

At Make, Sara leads Business Automation & AI. Her work has two sides: enabling employees to become confident AI builders, and delivering AI projects that create measurable business impact.

The first months were very pragmatic. Before launching ambitious use cases, the company needed safe tools, clear procurement, compliance answers and a strong rationale for why AI investment made sense.

Only then did Sara focus on early projects. Rather than trying to transform every function at once, she looked for low-hanging fruit aligned with Make’s strategy, especially around marketing operations and process-heavy use cases.

Two years later, the results are significant. Sara shared that 96% of employees have built their own AI agents, more than 400 agents are running in production, and last year alone her team delivered 34 AI projects, 23 of which were considered successful.

What makes this especially interesting is the operating model behind the results. Sara’s team is decentralised: its members spend most of their time embedded in business functions, building AI and automation solutions directly, while also dedicating part of their capacity to coaching other employees. The team is responsible for roughly 30% of the company’s agents and automations, but its coaching work helps enable the remaining 70% built across the organisation.

🧩 AI Challenges & Pain Points

One of Sara’s biggest lessons is that enthusiasm does not create adoption by itself.

Employees may be curious, motivated and full of ideas, but daily work quickly takes over. Without support, even promising experiments can stall. People need time to build, someone to answer questions, and a structure that keeps momentum alive.

As Sara put it: “Interest alone is not going to get you there.”

That is why she believes AI transformation cannot be treated as a side job. If no one truly owns the enablement, adoption often remains limited to a few enthusiastic champions. To scale, companies need dedicated people creating visibility, removing blockers and helping teams continue after the initial excitement fades.

Another surprise was the role of training. Early on, Make relied more on external AI experts. Over time, internal enablement became much more effective. Employees learned faster from colleagues solving real problems with the same tools, systems and constraints.

Sara also noticed something unexpected about agents. She distinguishes between workflows, where every step is predefined, and agents, where the builder defines a goal and gives the system tools to complete it. She expected agents to be harder for beginners, but often found the opposite. New builders could think naturally in outcomes, while experienced automation builders sometimes struggled because they were used to controlling every step.

🚀 The Use Case

One use case Sara shared comes from sales, where valuable information often disappears after a customer call. Sales representatives still have to update the CRM, write follow-ups, escalate technical questions and share product feedback, which means time that could be spent selling is instead spent on administration.

Make built what it calls the Post-call Hero: an AI agent that extracts information from sales conversations and orchestrates the next actions. Depending on what was discussed, it can prepare CRM updates, draft a follow-up email, escalate a technical question, provide coaching to the account manager or involve the right internal stakeholders. According to the example Sara shared, this saved around four hours per sales representative each week and was associated with a 32% increase in sales revenue.

The same source data is used by a second system focused on deal losses. Previously, 79% of lost deals had no loss reason recorded in the CRM. Even when a reason was entered, the team found that in 86% of cases it differed from what the call transcripts indicated had actually happened. A deal marked as lost because of price, for example, might instead have failed because no clear use case was established, the economic buyer was absent, or an important product capability was missing.

A multi-agent system now analyses the transcripts, identifies the likely reasons behind the loss and updates the records. The result is not only less administrative work for salespeople, but also more reliable information for sales leaders and product teams. It turns customer conversations into structured input for coaching, forecasting and roadmap decisions.

💭 Reflections & Future Visions

When asked what matters most for working in AI today, Sara did not start with technical skills.

She talked about agency.

For her, high-agency people are curious, resilient and courageous. They want to understand why things happen, they keep going when they hear “no”, and they are willing to experiment before everything is perfectly clear.

She also believes more people will need management skills, even if they never manage people. As AI agents become part of everyday work, professionals will need to delegate clearly, break problems into tasks, coordinate several streams of work and evaluate outputs. These used to be manager skills. Increasingly, they will become everyday AI skills.

Sara’s story is a reminder that the hard part of AI transformation is rarely the technology alone. The harder challenge is building the habits, trust and operating model that allow people to use it well.

🔗 Follow Sara

• LinkedIn

Thank you for tuning in to this edition of the AI Builders Interview series! 🎙️

🔗 Dive deeper into the minds behind AI innovation—explore more conversations with the builders shaping the future, and check out 150+ real-world AI case studies in our growing library:

Discover more interviews of AI builders

✨ We hope this dialogue offered fresh insights and sparked new ideas!

🛠️ Know an AI builder doing great work? Nominate them for a future interview!

Disclaimer: This content was (obviously 😉) built with the assistance of AI.

Florian Bruniaux’s move back into hands-on engineering was less strategic than personal. After more than a decade in senior leadership roles (CTO, CTPO, VP Engineering, Engineering Manager), he realized he was heading toward burnout, with management crowding out the actual building. So he pulled back. Today he works as Founding AI Engineer at Méthode Aristote, where he develops EdTech products for adult education and children’s learning, with AI central to both what gets built and how the team builds it.

Over the last eight months, Florian has gone deep into AI-assisted development, spending two to three hours a day just tracking what is changing. He is careful to distinguish between “vibe coding” (letting the machine drive, accepting whatever it produces) and what he actually does: using AI to move faster while keeping full technical ownership of architecture, quality, and long-term consequences. That difference shapes everything about how he works.

His background gives him an unusual vantage point for this moment. More than 12 years across SaaS, data, AI, e-commerce, and EdTech means he has seen enough cycles to resist hype and enough hands-on experience to recognize what AI actually changes versus what it only appears to change.

🛠️ What He’s Building

At Méthode Aristote, Florian works as principal engineer on a platform combining personalized academic support, qualified tutors, and AI tools including an educational chatbot and automatic report generation. The goal, in his words, is to build faster and create a “wow” effect for users. What makes the story interesting is less the product itself and more who is now involved in building it.

The product team currently pushes around ten pull requests per day. QA pushes two or three. Florian’s boss, a former teacher with no engineering background, contributes directly. The quality is not perfect, but it is far from catastrophic, which is precisely the point: development is no longer a space only engineers occupy, and that changes what engineers are actually for.

Florian describes this shift as “harness engineering.” Developers are no longer primarily feature builders. They are becoming the people who design the tools, tests, rules, and technical building blocks that allow product and QA to contribute safely without accumulating invisible debt. The bottleneck has moved away from implementation and toward the adjacent constraints: cybersecurity, product clarity, design governance, QA capacity.

Design is one concrete example. When AI generates interfaces without human review, everything converges toward the same aesthetic. The process that actually works, according to Florian, is a tighter loop between prototype, design system, and prototype again, with a human reviewing output against explicit rules rather than simply accepting what Claude produces.

The pace has also changed the meeting structure. Fewer retrospectives and dailies, more direct co-working sessions where product people are inside the IDE and specifications evolve in conversation with the code. A sprint that used to take two weeks can now compress into a single day.

Florian keeps building outside of work too. One idea can become a working proof of concept in an evening, with the constraints becoming very physical: how much the Mac can run and how much sleep he can give up. One of those evening projects is RTK, his first open-source tool, a Rust CLI proxy that reduces LLM token usage by 60 to 90%. He is not only using AI tools. He is building tooling for the people who are.

🚀 AI Challenges & Pain Points

The most consequential challenge Florian raised is not technical but organizational: what happens when more people can build? If product teams can ship code and QA can contribute directly, the developer’s responsibility shifts from writing features to defining what others can safely change. That raises hard questions about quality ownership, codebase governance, and who is accountable when something works on the surface but accumulates invisible fragility underneath.

Speed creates a second kind of pressure. Teams need a completely new reference for what is normal, and once people are accustomed to AI-assisted velocity, they can become disconnected from organizations that are not working the same way. Expectations shift in both directions, sometimes faster than the systems around them.

Florian is equally specific about anticipation. AI helps generate the next feature but does not automatically think through what it costs six months later. Technical debt, infrastructure costs, GitHub usage, QA load, and review complexity all tend to grow in proportion to output volume, and most teams do not notice until the accumulation is already significant.

The last challenge is harder to categorize. When everything becomes possible, everything becomes tempting. Florian describes a kind of cognitive pressure that comes with constant productivity: the excitement of building, combined with the awareness that anything could become a project tonight. There is something almost addictive about that state, and it requires deliberate restraint that does not come naturally when the tools are this fast.

💭 Reflection

When asked about the mindset that matters most right now, Florian did not offer a productivity framework. He raised a question he does not have a clean answer to: with this level of assistance, do developers risk losing the skills they are no longer practicing?

It is not a rhetorical move. He sits with it as an open problem. AI compresses months of work into days, brings implementation closer to product, and lets smaller teams operate at a scale that used to require much larger ones. What is less clear is whether faster output leads to deeper understanding or substitutes for it, whether developers who rely on AI for architecture decisions are building intuition or delegating it. The answer probably depends on how deliberately they engage with what the machine produces, rather than how much of it they generate.

That is the tension Florian keeps returning to. The developer of the AI era may not be the only person writing code anymore, but they may need to own more of what surrounds it: the guardrails, the architectural decisions, the quality bar, and the judgment calls that no tool makes on their behalf.

🔗 Follow Florian

You can follow Florian on LinkedIn and explore his work through his personal website, where he shares open-source projects around Claude Code, RTK, developer productivity, and AI-assisted workflows, and of course on his GitHub.

Thank you for tuning in to this edition of the AI Builders Interview series! 🎙️

🔗 Dive deeper into the minds behind AI innovation—explore more conversations with the builders shaping the future, and check out 150+ real-world AI case studies in our growing library:

Discover more interviews of AI builders

✨ We hope this dialogue offered fresh insights and sparked new ideas!

🛠️ Know an AI builder doing great work? Nominate them for a future interview!

Disclaimer: This content was (obviously 😉) built with the assistance of AI.

The three products in this edition share something that most AI tooling still gets wrong: they were designed around a specific moment of human judgment, not around the idea of removing it. Lithosquare lets geologists interrogate AI-ranked targets before any drill is funded. Pivot flags the purchase, benchmarks the supplier, then surfaces the decision for the buyer. Claude Tag executes asynchronously and posts results back into the shared thread where the team can inspect, redirect, or continue the work. In each case, the product earns adoption by being explicit about what it decides and what it defers — and that clarity is what makes it deployable in high-stakes, multi-stakeholder environments. The pattern worth watching is not that AI is doing more. It is that the products gaining real traction are the ones that have mapped the workflow carefully enough to know where human judgment is irreplaceable, and designed the interface around that point rather than around the model’s capabilities.

🪨 Lithosquare — compresses months of geological data processing into days, so geoscientists spend their time on drilling decisions, not data cleaning.

💸 Pivot — turns procurement into a structured system of record, with AI that benchmarks suppliers and routes decisions without replacing the buyer’s call.

💬 Anthropic (Claude Tag) — moves AI from private assistant to shared team agent inside Slack, where any teammate can inspect, correct, or continue the work in the same thread.

Lithosquare – Geology AI for Critical Mineral Exploration

Organization Type: Start-up
AI Purpose: Augment
Type of AI Model: Supervised learning, Generative AI
AI Application Type: R&D / Product development
Targeted Industry: Mining
Target Group (AI User): Researchers, Data & analytics

Use Case Description:
Lithosquare is a Paris-based company that applies AI to one of the most expensive and uncertain phases of mining: finding where valuable mineral deposits actually are. Exploration teams traditionally spend the majority of their time assembling and cleaning fragmented geological, geochemical and geophysical datasets before any interpretation can begin, a process that stretches across months and leaves geologists with little time for the judgment calls that matter. Lithosquare’s platform unifies these scattered data sources into a single geological framework, then ranks exploration targets by prospectivity and sequences fieldwork to reduce unnecessary drilling. Geoscientists interact directly with the system, interrogating targets and validating AI-generated scenarios before any capital is committed. The platform has been deployed in active programs with Aterian across Morocco and Botswana, and with Eramet and BRGM for critical-metals discovery in Africa. For exploration teams managing large, multi-country portfolios, the practical effect is that months of data processing compress into days, freeing geologists to focus on field interpretation and drilling decisions.

Key Features:

• Consolidates geological, geochemical, geophysical and public datasets into a unified framework with traceable exploration reasoning.

• Builds dynamic 3D geological scenarios to identify where uncertainty is highest and where sampling or drilling will resolve it most efficiently.

• Updates active geological scenarios continuously as new drill intercepts arrive, narrowing uncertainty in real time.

• Ranks exploration targets by prospectivity across large multi-country portfolios, surfacing the highest-value opportunities first.

• Provides a human-in-the-loop validation layer where geoscientists review and interrogate AI-ranked targets before capital is deployed.

Results:

• Reduced exploration analysis timelines from months to days across active programs, according to company and investor reporting.

• Identified 8 high-priority targets across Morocco and Botswana in the Aterian collaboration, completing the target-selection phase by April 2026.

• Raised a $25M seed round in May 2026, led by World Fund and Kindred, signaling institutional confidence in the platform’s commercial traction.

Launch Date: January 2024

Retrieve the case and more (including their sources) on the AI in USE website

Pivot – AI Procurement Operating System

Organization Type: Start-up
AI Purpose: Augment
Type of AI Model: Generative AI, Unsupervised / Reinforcement learning
AI Application Type: Operations
Targeted Industry: Financial Services, Healthcare, Software & Technology
Target Group (AI User): Operations teams, Finance teams

Use Case Description:
Pivot is a procurement operating system for enterprise teams that need one place to manage sourcing, approvals, purchasing, invoices, payments, budgets, expenses, reporting, and ERP integrations. It replaces fragmented procurement work spread across email, spreadsheets, and legacy tools with a structured system of record. Procurement and finance teams use the product to ask spend and budget questions in plain English, submit requests more easily, and automate tasks such as contract field extraction, vendor benchmarking, and stakeholder routing. The platform also supports custom procurement agents for customer workflows, with ROI tracked for each deployment. This gives teams faster decisions, better spend visibility before month close, and a more auditable procurement process.

Key Features:

• Natural-language spend analytics lets procurement and finance teams query budgets, contracts, and savings in plain English without writing reports.

• Smart contract extraction automatically pulls key fields from a contract repository, trained on each customer’s own contract history.

• AI-driven intake adapts the request submission form to the specific request type, reducing manual errors and back-and-forth with requesters.

• The Intelligent Benchmarking Agent places every purchase request under automated negotiation and benchmarks supplier pricing to identify additional savings.

• Custom procurement agents are deployed per customer workflow, with ROI tracked per agent deployment.

• ERP and financial system integrations give the AI layer a structured, auditable data foundation rather than relying on fragmented legacy inputs.

Results:

• Lemonade’s Intelligent Benchmarking Agent covers 100% of purchase requests and delivers 10% additional savings per deployment.

• Pivot processes more than $3 billion in invoices annually across its customer base.

• The platform operates across 25+ countries with more than 2,000 vendors onboarded.

• Pivot raised $40 million in Series B funding in 2026, bringing total funding to $70 million since its 2023 founding.

Launch Date: January 2023

Retrieve the case and more (including their sources) on the AI in USE website

Anthropic – Claude Tag, Slack-Native Team AI Agent

Organization Type: Start-up
AI Purpose: Augment
Type of AI Model: Generative AI
AI Application Type: Operations
Targeted Industry: Multi-industry
Target Group (AI User): Software engineers, Product managers, Operations teams, Customer support agents

Use Case Description:
Anthropic built Claude Tag to turn Slack into a shared work surface for enterprise teams. The product addresses a common problem in knowledge work: important tasks get discussed in channels, but the context is scattered across threads, documents, code repositories, and other tools. With Claude Tag, users mention @Claude in Slack to delegate work in the same conversation where the request started, and the agent reads the local context, uses approved tools and data sources, and posts the result back in the thread. Product, engineering, support, sales, and operations teams use it to reduce repeated context-setting, manual triage, status chasing, and cross-channel information retrieval. Anthropic positions it as a shift from private AI assistants to a shared, multiplayer workflow inside the tools teams already use.

Key Features:

• Users mention @Claude in any Slack channel or thread to assign work directly, without leaving the conversation.

• The same Claude instance is visible to everyone in the channel, so teammates can inspect, correct, or continue a task in the shared thread.

• Claude retains context from the channels it is permitted to access, reducing the need to re-explain team-specific background on every request.

• Admins configure exactly which Slack channels, tools, repositories, documents, and data sources Claude can reach, with access scoped per channel or workspace to limit data exposure across functions.

• Claude executes tasks asynchronously, posting results when complete and scheduling follow-up work without requiring the user to stay in the conversation.

• When enabled, Claude proactively flags unresolved threads or posts updates when background tasks finish.

• Engineering teams route coding tasks from Slack directly to Claude Code, making the Slack channel a delegation point for software development work.

Results:

• 65% of Anthropic’s own product team code is generated through its internal Claude Tag deployment, as reported by Anthropic at launch.

• Teams reduced repeated context-setting across support ticket triage, bug root cause analysis, and product metrics retrieval by routing those tasks to Claude inside existing Slack threads.

• Asynchronous task execution allows team members to delegate work and move on immediately, recovering time previously spent on manual status tracking and cross-channel information retrieval.

Launch Date: June 2026

Retrieve the case and more (including their sources) on the AI in USE website

Thank you for being part of the AI in USE community! 🌟

🔗 Visit our website AI in USE to explore the full library of AI use cases and discover how AI is transforming industries across the globe:

Explore 100+ real-world use cases

✨ We hope these AI use cases spark inspiration!
💬 Let us know which one intrigued you the most—your feedback helps us grow!

Disclaimer: This content was (obviously 😉) built with the assistance of AI.

The three products in this edition share a detail that is easy to overlook: none of them are designed to finish the job. Webflow generates a structured site draft, not a published website. FloQast runs reconciliation workflows through a sandbox before anything touches the general ledger. CoCounsel returns citation-backed research and first-draft documents, but the lawyer signs off. That restraint is not a limitation — it is the product decision that makes adoption possible in the first place, particularly in domains where a wrong output carries real professional or financial consequences. What these teams understood is that AI earns operating room by being legible and reversible, not by being autonomous.

The sharper lesson buried in these cases is about where friction gets moved, not eliminated. Each product removes the setup friction that used to eat hours before real work could begin — the blank canvas, the IT ticket, the manual database query. But each product reintroduces human judgment at the point where the stakes actually matter. That sequencing is deliberate, and it is what separates products that teams adopt from those that get piloted once and quietly shelved.

🌐 Webflow — 60,000 sites generated in under a year, not because the AI builds great websites, but because it removes the fifteen decisions that used to block designers from starting

⚖️ FloQast — accounting teams running their own automation with a sandbox, inline validation, and sign-off checkpoints, meaning finance no longer waits on engineering to move

🔍 Thomson Reuters CoCounsel — small law firms handling 77% more matters per month, not by replacing lawyers, but by eliminating the hours spent constructing the research before any legal thinking begins

Webflow – AI Site Builder

Organization Type: Scale-up
AI Purpose: Augment
Type of AI Model: Generative AI
AI Application Type: User Experience
Targeted Industry: Multi-industry
Target Group (AI User): Designers, Marketers, Employees

Use Case Description:
Building a new website has always started with the same friction: blank pages, structural decisions made too early, and a long gap between a rough idea and something editable in a real production environment. Webflow’s AI Site Builder closes that gap by turning a short text description into a multi-page, responsive website draft, complete with a foundational design system covering typography, colors, buttons, and layout, generated directly inside Webflow’s visual builder rather than as a separate export. Designers, marketers, and non-technical founders use it to skip the setup phase and start from a structured, customizable draft instead of a blank canvas or a rigid template. The generated site connects immediately to Webflow’s CMS, hosting, localization, analytics, and optimization tools, so the output lives inside the same workflow teams already use. The practical result is a faster first draft, not a finished product: independent reviewers consistently describe the output as a solid scaffold that still requires human judgment for brand alignment, SEO, accessibility, and design polish.

Key Features:

• Generates a multi-page responsive website from a short text prompt, producing up to five structured pages at once.

• Creates a foundational design system with reusable styles for colors, typography, buttons, and layout spacing.

• Includes animation presets built on GSAP, selectable during the generation flow.

• Allows section-level editing through an AI Assistant that generates navbars, footers, hero sections, testimonials, and contextual copy inside the visual builder.

• Connects generated sites directly to Webflow’s CMS, localization, analytics, and optimization tools without requiring export or migration.

• Gives workspace administrators the ability to enable or disable Webflow AI features across a team, supporting enterprise governance requirements.

Results:

• More than 60,000 websites published using AI Site Builder within the first year of the beta release.

• Reduces time from brief to first editable draft to minutes, according to Webflow’s own product documentation, with external reviewers confirming meaningful acceleration at the scaffolding stage.

• Extends Webflow’s addressable user base by making structured site creation accessible to non-technical founders and marketers who previously lacked the skills to set up a Webflow project from scratch.

Launch Date: February 2025

Retrieve the case and more (including their sources) on the AI in USE website

FloQast – Visual Agent Builder

Organization Type: Scale-up
AI Purpose: Augment
Type of AI Model: Generative AI, Unsupervised / Reinforcement learning
AI Application Type: Operations
Targeted Industry: Financial Services, Software & Technology
Target Group (AI User): Finance teams

Use Case Description:
FloQast’s Visual Agent Builder gives corporate accounting teams a way to build and manage AI-driven close workflows without relying on engineers. The product addresses a clear operational problem: finance teams handle deadline-driven reconciliations, journal entries, and audit documentation, but traditional automation often sits behind IT tickets and technical bottlenecks. Here, accountants use a visual canvas and natural-language agent builder to create, edit, and inspect workflows themselves, with reusable accounting actions, validation checks, sandbox testing, and human review built into the process. The AI value is in making close automation easier for accountants to operate directly while keeping workflows legible, auditable, and controlled. This reduces dependence on engineering support and gives accounting teams more direct control over how automation runs in their close process. It also strengthens governance by embedding review and auditability into the workflow rather than treating them as separate steps.

Key Features:

• Drag-and-drop canvas lets accountants build accounting agent workflows as visual process maps without writing code.

• Skills library provides reusable accounting actions including ERP connectors, journal entry creation, trial balance fetch, reconciliations, and input/output mapping.

• Inline validation detects missing inputs or broken connections in the workflow before any action reaches the general ledger.

• Sandbox testing environment runs workflows against real data in isolation before production posting.

• Natural-language agent creation allows accountants to describe a workflow in plain language and generate the corresponding agent configuration.

• Human-in-the-loop review checkpoints are embedded directly inside workflows to preserve auditability and sign-off controls.

• AI Variance Analysis flags general ledger anomalies, explains variances, and links them back to source transactions.

Results:

• 38% reduction in reconciliation time reported across FloQast customer surveys.

• 23% reduction in time spent on audit processes reported across FloQast customer surveys.

• 20% reduction in discrepancies identified by auditors reported across FloQast customer surveys.

• 39% increase in close-data accuracy reported across FloQast customer surveys.

• 27 hours saved per month per team reported across FloQast customer surveys.

Launch Date: March 2026

Retrieve the case and more (including their sources) on the AI in USE website

Thomson Reuters – CoCounsel Legal AI Assistant

Organization Type: Corporate
AI Purpose: Augment
Type of AI Model: Generative AI
AI Application Type: User Experience
Targeted Industry: Legal, Financial Services, Education
Target Group (AI User): Legal teams, Researchers, Employees

Use Case Description:
Legal work runs on research, document review, and drafting, all of which traditionally require lawyers to manually query databases, read through large document sets, and assemble arguments from scratch. Thomson Reuters built CoCounsel Legal as an AI assistant embedded directly into Westlaw, Practical Law, and connected legal workflows to let lawyers delegate structured tasks rather than execute them step by step. A lawyer assigns a research question or a set of documents, and CoCounsel plans and runs the work autonomously, returning citation-backed answers, extracted clauses, issue tables, or draft correspondence. The product is used by law firm attorneys, in-house legal teams, litigators, and legal operations professionals across both private practice and corporate legal departments. Small firms using CoCounsel reported increasing their average monthly matter capacity from 10.5 to 18.6, and 82% reported measurable time savings. By grounding every output in Westlaw’s proprietary content and enforcing human review as the final step, Thomson Reuters positions CoCounsel as a defensible productivity layer rather than an autonomous replacement for legal judgment.

Key Features:

• Deep Research plans and executes multi-step legal research grounded in Westlaw and Practical Law, returning citation-backed answers without manual query iteration.

• Document review extracts issues, builds structured analysis tables from large document sets, and flags relevant passages for lawyer review.

• AI-assisted drafting generates first-draft legal documents and correspondence from lawyer instructions, reducing initial writing effort.

• Contract analysis extracts specific clauses and checks documents against internal policies or standard playbooks.

• Tabular Analysis organizes document review findings into structured tables, supporting litigation and due diligence workflows.

• The assistant is embedded across Westlaw, Practical Law, HighQ, Document Intelligence, and Microsoft Teams, so lawyers work inside their existing environments without switching tools.

• All model interactions run with zero-retention API controls, meaning client prompt content is not stored or used for model training.

Results:

• Reached 1 million professional users globally as of February 2026, with Thomson Reuters shares rising more than 11% on the announcement.

• Small-firm respondents reported average monthly matter capacity increasing from 10.5 to 18.6, a 77% increase.

• 82% of surveyed small-firm users reported measurable time savings on core legal tasks.

• 70% said CoCounsel enabled higher caseload or revenue capacity.

• Users reported spending up to one-third less time on research, document review, and drafting.

• G2 reviewers gave CoCounsel Legal a 4.5 out of 5 rating across 285 reviews, with time savings and usability as the most cited strengths.

Launch Date: March 2023

Retrieve the case and more (including their sources) on the AI in USE website

Thank you for being part of the AI in USE community! 🌟

🔗 Visit our website AI in USE to explore the full library of AI use cases and discover how AI is transforming industries across the globe:

Explore 100+ real-world use cases

✨ We hope these AI use cases spark inspiration!
💬 Let us know which one intrigued you the most—your feedback helps us grow!

Disclaimer: This content was (obviously 😉) built with the assistance of AI.

The three products in this edition all cross the same threshold: AI that does work, not just AI that informs it. Notion runs multi-step workspace tasks for up to 20 minutes before asking a human to review. JPMorgan’s agents execute across internal systems for up to two hours. Waiv rules out 70% of patients from a diagnostic test without a specialist ever looking at the case. None of these are chatbots dressed up as agents. What they share is a precise, deliberate answer to a question most teams still avoid: at what point should the system stop and ask? The real product design problem here is not the model, and it is not the interface — it is the handoff. Where each of these teams drew that line, and how they made it legible to the people in the loop, is what separates a deployed product from a liability.

🗂️ Notion: a workspace agent that plans before it acts, showing users the proposed sequence of changes before executing anything across their pages and databases

🏦 JPMorgan Chase: 200,000 employees onboarded to governed AI in eight months, with autonomous agents operating up to two hours inside controlled, auditable boundaries

🔬 Waiv: a diagnostic AI that earns clinical trust not by replacing pathologists, but by eliminating 70% of unnecessary tests with a 99% negative predictive value they can stand behind

Notion – AI Workspace Agent

Organization Type: Scale-up
AI Purpose: Automate
Type of AI Model: Generative AI
AI Application Type: Operations
Targeted Industry: Software & Technology, Multi-industry
Target Group (AI User): Product managers, Operations teams, Marketers, Software engineers

Use Case Description:
Notion is a workspace platform combining documents, databases, projects, and calendar into a single environment used by over 100 million people. The core problem it addresses is execution overhead: knowledge workers spend significant time on repetitive coordination tasks such as updating trackers, creating project docs, synthesizing inputs from Slack or Google Drive, and keeping pages in sync across teams. With Notion 3.0, the company rebuilt its AI layer into an agent that performs multi-step workspace tasks autonomously, not just drafts or summarizes on request. A user assigns a goal, the agent plans the required actions, executes them across pages and databases for up to 20 minutes, and presents a review step before finalizing changes. Custom Agents extend this further, triggering automatically on schedules or database events to file bugs, post reports, or update records without manual initiation. The product targets teams already embedded in Notion who want to reduce the manual work between receiving information and acting on it.

Key Features:

• Personal Agent performs multi-step tasks including creating pages, editing databases, and generating project documentation directly inside the workspace.

• Plan mode shows users the proposed sequence of changes before execution, allowing review and approval before any edits are applied.

• PDF-to-workflow extraction pulls contracts or research documents into structured trackers with tasks, dates, owners, and citations.

• Cross-app synthesis pulls content from Slack, Google Drive, email, and the web to produce reports, playbooks, or feedback summaries inside Notion.

• Custom Agents run autonomously on schedules or database triggers, with their own permission scope separate from the user’s personal access rights.

• Multi-model routing selects between Claude, GPT, Gemini, Grok, or Notion’s own routing layer based on the task’s speed and cost requirements.

Results:

• Vector search infrastructure scaled 10x over two years while infrastructure cost dropped 90%.

• Agents execute continuous multi-step workflows lasting up to 20 minutes across hundreds of pages and databases within a single run.

Launch Date: September 2025

Retrieve the case and more (including their sources) on the AI in USE website

JPMorgan Chase – Enterprise AI Agent Platform

Organization Type: Corporate
AI Purpose: Augment
Type of AI Model: Generative AI, Unsupervised / Reinforcement learning
AI Application Type: Operations
Targeted Industry: Banking, Financial Services
Target Group (AI User): Software engineers, Operations teams, Finance teams

Use Case Description:
JPMorgan Chase operates one of the largest internal AI deployments in financial services, built around a platform called LLM Suite that gives more than 200,000 employees secure access to generative AI for drafting, summarization, research, and workflow automation. The core business problem is straightforward: bankers, engineers, and operations staff spend large portions of their day on manual, repetitive work that delays client-facing activity and limits how many relationships each person can manage. The platform addresses this by embedding AI directly into existing workflows rather than offering a standalone tool, with capabilities ranging from overnight market screening for private bankers to coding assistance for engineers across the capital markets division. A separate transaction screening system reviews compliance cases at more than twice the previous volume while cutting manual operator checks by half, which directly reduces client delays on payment processing. The next step already underway is transitioning from single-task AI responses to long-running autonomous agents that can execute multi-step workflows across internal systems for up to two hours before requiring human review, a shift that moves AI from answering questions to completing work. Governance is the defining constraint: every capability operates inside a controlled environment with model risk oversight, access controls, and auditability requirements built into the deployment architecture.

Key Features:

• LLM Suite provides 200,000 employees with secure, governed access to large language models for drafting, summarization, ideation, and workflow integration.

• Private banking intelligence runs overnight screens of market activity, client positions, and research so bankers arrive at client conversations with structured, ready-to-use briefings.

• Transaction screening AI reviews compliance cases at more than double the previous volume while cutting the number of manual operator checks by half.

• AI coding assistants are used by more than 90% of engineers in the capital markets division, integrated directly into their development environment.

• Long-running autonomous agents execute multi-step workflows across internal software for up to two hours, handling tasks that previously required continuous human coordination.

• An employee assistant allows staff to get help and take action across firm systems through a single personalized AI interface.

Results:

• 200,000 employees onboarded to LLM Suite within eight months of launch.

• 65,000 active users in the capital markets division alone, with coding assistants adopted by more than 90% of its engineers.

• Engineering productivity improved by 10 to 20% from AI coding assistance, reported by JPMorgan’s CIO.

• Transaction screening volume more than doubled while manual operator checks were reduced by half.

• Private banking gross sales increased by 20%, attributed to AI tools screening markets and client positions overnight.

• JPMorgan estimates 450 active AI use cases with a path to 1,000, carrying an expected value of $1 billion to $1.5 billion.

Launch Date: June 2024

Retrieve the case and more (including their sources) on the AI in USE website

Waiv – AI Precision Testing for Oncology Diagnostics

Organization Type: Start-up
AI Purpose: Augment
Type of AI Model: Computer vision, Supervised learning
AI Application Type: R&D / Product development, User Experience
Targeted Industry: Healthcare, Pharmaceuticals, Biotechnology
Target Group (AI User): Healthcare professionals, Researchers

Use Case Description:
COMMENT[too many acronym, not understandable for “le commun des mortels”] Waiv, spun out of Owkin in March 2026 with $33 million in backing, builds AI-powered diagnostic tests that analyse routine digitised cancer pathology slides to surface biomarker signals, relapse-risk scores, and treatment-stratification information that would otherwise require more time, cost, and specialist effort to produce. In oncology labs today, biomarker testing such as MSI status or BRCA screening typically involves expensive reflex testing on every eligible patient, regardless of likelihood, creating significant workload and cost pressure on pathology teams. Waiv’s tests run on digitised H&E whole-slide images already captured as part of standard workflows, using deep learning to identify visual patterns that correlate with molecular or clinical outcomes. Pathologists use the system to triage patients before committing to downstream molecular tests, while oncologists receive structured risk signals to inform treatment decisions, and biopharma R&D teams use it to discover or validate biomarkers in clinical trial contexts. The flagship product, MSIntuit CRC, received CE marking under IVDR in April 2026 and can rule out approximately 70% of colorectal cancer patients from unnecessary reflex MSI testing, with a negative predictive value of around 99%. Waiv integrates its tests into existing digital pathology infrastructure through its Destra platform, which is compatible with systems from Roche, Proscia, Sectra, and Tribun Health.

Key Features:

• MSIntuit CRC pre-screens colorectal cancer patients for microsatellite instability from digitised H&E slides, eliminating the need for reflex molecular testing in the majority of cases.

• RlapsRisk BC combines pathology slide analysis with clinical variables such as age, lymph node status, and tumour size to generate a relapse-risk profile for breast cancer patients.

• BRCAura RUO screens for germline BRCA mutation likelihood from H&E breast cancer slides without requiring a dedicated genetic test.

• TLS Detect identifies tertiary lymphoid structures in pathology slides and delivers heatmap and PDF outputs formatted for research workflows.

• The Destra platform integrates all Waiv tests directly into existing lab digital pathology systems, allowing pathologists to access AI-generated results without changing their core tooling.

Results:

• MSIntuit CRC achieves a negative predictive value of approximately 99%, ruling out up to 70% of colorectal cancer patients from reflex MMR/MSI testing.

• Estimated workload reduction reaches 50% for biopsies and 70% for surgical resections in labs using MSIntuit CRC.

• Payer cost savings are estimated at 31 to 57% per patient screened, based on modelling from clinical validation data.

• MSIntuit CRC v2 was validated across more than 1,500 biopsies and 500 resections spanning five external cohorts.

• BRCAura achieves a mean AUC of 0.80, with sensitivity of 0.93 in external validation cohorts.

• TLS Detect reaches 97% sensitivity for tertiary lymphoid structure detection in externally validated research settings.

Launch Date: January 2021

Retrieve the case and more (including their sources) on the AI in USE website

Thank you for being part of the AI in USE community! 🌟

🔗 Visit our website AI in USE to explore the full library of AI use cases and discover how AI is transforming industries across the globe:

Explore 100+ real-world use cases

✨ We hope these AI use cases spark inspiration!
💬 Let us know which one intrigued you the most—your feedback helps us grow!

Disclaimer: This content was (obviously 😉) built with the assistance of AI.

What does it take to build AI that works outside the demo?

For Julien Guitard, the answer did not start with generative AI, agents, or foundation models. It started much earlier — with statistics, public policy, and a deceptively difficult question: how do you know whether something actually worked?

Julien studied at École Polytechnique and ENSAE, then completed a PhD in economics and econometrics, working on applied statistics, labor-market data, causality, and policy evaluation. His early work focused on measuring the impact of public policies: employment programs, labor-market interventions, and complex economic systems where the goal was not just to predict the future, but to understand what would have happened otherwise.

That counterfactual mindset stayed with him.

Before AI became the language of every boardroom, Julien was already working on the questions that now sit at the center of serious AI adoption: measurement, explainability, data quality, operational constraints, and business impact.

After roles in public policy, the French Treasury, strategy consulting, and data consulting, Julien joined Lucky Cart , where the scale changed dramatically. He moved from policy datasets and econometric models to hundred of millions of retail transactions, machine learning pipelines, deep learning models, MLOps, and large-scale optimization.

Today, Julien is the founder of vertee.ai, where he works as a fractional Chief Data & AI Officer for mid-sized companies, especially technology and data-driven businesses. His work sits at the intersection of AI strategy, data architecture, implementation, and organizational design.

And his point of view is refreshingly grounded: generative AI is powerful, but it does not make the old problems disappear.

Performance measurement still matters. A/B testing still matters. Explainability still matters. And in many cases, traditional machine learning is still the better tool.

🛠️ What He’s Building

Over the last six months, Julien has been building his own advisory and implementation practice through vertee.ai.

His clients are often fast-growing mid-sized companies. Some already have data, but it is not well-managed, structured, or leveraged. Others are excited about AI, but have not yet asked the deeper questions about their data foundations, technical stack, operating model, or measurement framework.

That gap is central to his thesis.

Many companies are under pressure to improve profitability and efficiency, while also feeling the pull of the generative and agentic AI hype cycle. But the reality inside organizations is often less glamorous: fragmented systems, unclear data models, manual processes, immature analytics, and teams that do not yet share a common language around data and AI.

Julien’s view is that companies should not treat strategy, team, and data as separate sequential steps.

Your strategy shows up in your operating model.
Your operating model shows up in your team topology.
Your team topology shows up in your data stack.
And your data stack eventually reveals what your business actually understands about itself.

At vertee.ai, Julien helps companies close that alignment gap across business, technology, data, and people. Sometimes that means designing a data model. Sometimes it means auditing a codebase. Sometimes it means building a machine learning or agent-based recommendation stack. Sometimes it means helping leadership teams define OKRs, governance, and a roadmap that connects ambition to execution.

The common thread: AI is not treated as a standalone layer. It is treated as part of the company’s operating system.

🚀 The Use Cases

One project Julien discussed was the AI/ML factory at scale he helped build at Lucky Cart.

The context was large-scale retail personalization. Lucky Cart needed to optimize promotional campaigns across massive volumes of transaction and consumption data. Manual campaign optimization could not scale to millions of users or the operational complexity of daily decision-making.

The goal was to automate opportunity identification using predictive modeling.

Julien and the team developed a machine learning pipeline factory using tools such as TensorFlow, Keras, Vertex AI, BigQuery, and custom loss functions adapted to retail-specific performance definitions. The system predicted user behavior patterns and helped automate optimal promotional strategies.

The result: roughly 80% of optimization decisions were automated, while campaign performance improved and customer success teams gained time back.

What makes this case interesting is not just the automation. It is the combination of prediction and measurement.

Julien’s background in causal inference and A/B testing shaped how he approached the problem. Retail data is noisy, unbalanced, and context-heavy. You cannot simply build a model and assume that a higher prediction score equals better business performance. You need measurement frameworks, benchmarks, and a way to connect model outputs to actual outcomes.

Another use case came from data infrastructure.

At Lucky Cart, the company had fragmented data systems across several retail clients and hundreds of brand clients. Commerce data formats were inconsistent, which slowed down API integrations and analysis. The task was to standardize the data architecture so the company could enable faster analytics across client operations.

Julien described this as building the kind of transformation layer that later became natural in the modern data stack: in-house pipelines first, then DBT-based transformations, with Airflow orchestration and unified commerce schemas.

The result was a reduction in data processing time from days to hours, making it possible to deliver faster business insights to clients.

The lesson: before AI can scale, data has to become legible.

A third use case came from a more recent codebase audit.

In that project, the issue was not a lack of AI ambition. It was heterogeneity of coding practices and architectural complexity. AI was used to identify patterns in the code, but Julien emphasized that the process required formal controls and human verification.

The workflow was not “AI replaces the expert.”

It was closer to:

Human → AI → Human

A human frames the problem.
AI accelerates exploration.
A human verifies, interprets, and reappropriates the output.

That pattern captures a lot of Julien’s philosophy. AI can move faster than humans across large bodies of information, but the final value still depends on judgment, structure, and accountability.

📈 Trends & Reflections

One of Julien’s funniest observations about the current AI moment is also one of the most revealing:

“Everything becomes a .md.”

Behind the joke is a deeper shift. More and more of company knowledge is becoming structured, written, reusable, and machine-readable: specs, prompts, workflows, documentation, agent outputs, customer feedback, internal reasoning. The small command line, the simple markdown file, the structured brief — these are becoming new interfaces between humans, teams, and AI systems.

But for Julien, the bigger shift is that AI is moving from a CTO topic to a COO topic.

It is no longer only about models, infrastructure, or engineering productivity. AI is entering workflows: how teams coordinate, how decisions are made, how customer feedback is processed, how operations are documented, and how execution is governed.

That is also where he sees a lot of friction.

Across companies, Julien observes strong heterogeneity in governance, management practices, and operational maturity. Some teams are ready to plug AI into clear processes. Others first need to clarify ownership, workflows, data structures, and decision rights.

This is why he is especially interested in structured generation: not just asking AI to produce text, but asking agents to produce outputs that follow a clear format, can be verified, reused, and inserted into real business workflows.

The promise is not just “AI answers faster.”

It is AI helping organizations turn messy inputs — codebases, customer feedback, internal docs, operational signals — into structured outputs that humans can check, own, and act on.

🔗 Follow Julien

You can follow Julien Guitard on LinkedIn

Thank you for tuning in to this edition of the AI Builders Interview series! 🎙️

🔗 Dive deeper into the minds behind AI innovation—explore more conversations with the builders shaping the future, and check out 150+ real-world AI case studies in our growing library:

Discover more interviews of AI builders

✨ We hope this dialogue offered fresh insights and sparked new ideas!

🛠️ Know an AI builder doing great work? Nominate them for a future interview!

Disclaimer: This content was (obviously 😉) built with the assistance of AI.

A clear pattern shows up across this week’s use cases, the winning products are not chasing full autonomy, they are removing the tedious steps that block progress while keeping the human, or the device, in charge of the call. The practical lesson for builders is that “end-to-end automation” is often the wrong goal, because the hard part is earning permission to act in workflows where trust, compliance, or quality is on the line. Design for assisted throughput: capture the input, do the legwork, write back into the system of record, and make review fast enough that people actually do it.

Adecco Group – AI Candidate Pre-screening Agent

Organization Type: Corporate
AI Purpose: Augment
Type of AI Model: Generative AI
AI Application Type: HR
Targeted Industry: Business Services
Target Group (AI User): HR teams, Consumers

Use Case Description:
Adecco Group uses an AI pre-screening agent to handle first-contact screening for high-volume recruitment and temporary staffing roles. The business problem is that recruiters only engaged a small share of applicants and missed many candidates who respond outside working hours, which slowed time-to-fill and increased repetitive workload. The product reaches out to candidates by SMS, chat, and voice, asks role-specific screening questions, captures answers, and writes the interaction back into the recruiting record. Recruiters review an AI-generated summary and full transcript and then decide next steps, with candidates not automatically advanced or rejected solely by the system. This shifts recruiters from repetitive early-stage outreach to qualification decisions and relationship work, while expanding screening coverage to all applicants.

Key Features:

• Generates job descriptions and role-specific screening questions from vacancy details.

• Checks whether a job record contains enough information before activating the screening agent.

• Runs discrimination and compliance checks on screening parameters before candidate outreach.

• Contacts candidates via SMS, chat, and voice so screening can be completed outside office hours.

• Conducts an automated pre-screening conversation, including handling common candidate questions.

• Logs candidate answers directly into Salesforce application records.

• Produces a recruiter-ready summary and provides the full transcript for review.

• Requires recruiter review and decisioning before progressing candidates.

Results:

• Increased pre-screening reach from about 10% of candidates to 100% in the UK deployment.

• Saved up to 15 minutes per interview for recruiters.

• Reported 15% time savings from Agentforce 360 usage.

• Achieved 4.6/5 candidate satisfaction in the first-wave rollout.

• Shifted screening to off-hours, with 57% of conversations happening outside standard working hours.

• Completed more than 100,000 agent-led screening conversations by February 2026.

Launch Date: January 2025

Retrieve the case and more (including their sources) on the AI in USE website

Google – Android Fake Call Detection for Impersonation Scams

Organization Type: Tech giant
AI Purpose: Augment
Type of AI Model: Generative AI
AI Application Type: Risk management
Targeted Industry: Telecommunications
Target Group (AI User): Consumers

Use Case Description:
Google adds fake call detection to Phone by Google to protect Android users from impersonation scams where attackers spoof caller ID and use AI voice cloning to sound like a trusted contact. The problem is that traditional caller-ID signals are easy to fake, so users answer high-risk calls believing they are real. The product verifies whether an incoming call actually originates from the contact’s device using an RCS-based device confirmation flow, and it escalates to an explicit double-check when the silent verification signal is missing. Android users see an on-screen warning when verification fails or when the real device indicates it is not calling. This reduces successful social-engineering fraud attempts and improves user trust in calls that are verifiably authentic.

Key Features:

• Silently verifies whether an incoming call is coming from the contact’s real device.

• Triggers a secondary confirmation check when the initial verification signal is missing.

• Displays an on-screen warning when verification indicates the call is not from the real device.

• Runs automatically in the background to reduce reliance on user judgment during a call.

• Uses an RCS-based approach designed for interoperability across supported Android calling stacks.

Results:

• Addressed a fraud category tied to FTC-reported impersonation scam losses of $2.95B in 2024.

• Aligned protections to a rising threat category reflected in FBI IC3 reporting of nearly $893M in 2025 losses tied to AI-related scam complaints.

Launch Date: June 2026

Retrieve the case and more (including their sources) on the AI in USE website

Figma – Prompt-to-Edit Natural-Language Design Editing

Organization Type: Scale-up
AI Purpose: Augment
Type of AI Model: Generative AI
AI Application Type: R&D / Product development
Targeted Industry: Software & Technology
Target Group (AI User): Designers, Product managers

Use Case Description:
Figma adds “Prompt to Edit” so teams can make UI changes by describing the intended update in plain language instead of hunting through menus and properties. The problem is that design work includes many repetitive edits and bulk changes across frames, which slows iteration and makes small updates costly. In Figma Design, users select specific layers and enter a prompt to apply changes such as layout tweaks, style updates, copy revisions, or generating variants like light and dark modes. Designers and product managers use the capability directly while building and refining screens, especially when they need to repeat similar changes across many frames. This reduces time spent on mechanical adjustments and speeds up creating and comparing design options, while keeping humans responsible for review and final quality.

Key Features:

• Edits selected layers based on a natural-language instruction entered by the user.

• Applies bulk changes across multiple frames to repeat layout, copy, or styling updates.

• Replaces or generates content to help complete mockups and create variants faster.

• Generates light and dark mode variants from an existing design using a prompt-driven workflow.

• Creates first-draft components from prompts to accelerate early-stage design exploration.

Results:

• Reduced manual effort for repetitive UI edits by shifting changes from property-by-property adjustments to prompt-driven updates.

• Shortened iteration cycles for producing design variants across multiple frames.

• Increased the speed of completing mockups by generating and replacing placeholder content within the design workflow.

• Expanded editing accessibility for non-specialists by allowing product managers to request precise changes without deep tool navigation.

Launch Date: September 2025

Retrieve the case and more (including their sources) on the AI in USE website

Thank you for being part of the AI in USE community! 🌟

🔗 Visit our website AI in USE to explore the full library of AI use cases and discover how AI is transforming industries across the globe:

Explore 100+ real-world use cases

✨ We hope these AI use cases spark inspiration!
💬 Let us know which one intrigued you the most—your feedback helps us grow!

Disclaimer: This content was (obviously 😉) built with the assistance of AI.

What does AI really change inside a company?

Not just the product. Not just the roadmap. Often, the first real shift happens in the invisible work teams do every day: writing specs, synthesizing feedback, preparing workshops, structuring discovery, documenting decisions, and turning scattered inputs into something others can act on.

Meet Marion Jachimski, Senior Product Manager in a fintech SaaS company. Her work sits at the intersection of product, operations, and AI-assisted workflows. Alongside her day job, she also teaches practical AI usage for product professionals through Noé and AI Discipline, where she designs hands-on trainings around prompting, AI agents, and AI-assisted product practices.

This interview was a little different from the usual format. We spent less time talking about “AI inside the product” and more time on something just as important: how AI is reshaping the craft of product management itself. And Marion’s perspective is especially useful because it comes from practice, not theory.

Her journey into AI started with a concrete need. At a previous company, the team wanted to launch an AI product, but she could not find the right internal resources at the time. So she started learning by doing: documenting herself, experimenting with prompting, testing use cases like verbatim analysis and categorization, and gradually turning that curiosity into a working method.

That method is now central to how she works: start from a real operational pain point, structure the input, build a repeatable AI-assisted workflow, then keep the human in charge of review and arbitration.

🛠️ What They’re Building

Today, Marion’s focus is split across two complementary worlds.

One of her main battlefields is product documentation and specification work. She started with an assistant to help draft specs — epics, features, and user stories. Then she moved toward a more structured delivery workflow built around a chain of three agents: one to generate specifications, one to structure the output, and one to turn it into tickets. Her role remains central in the process: she reviews, challenges, and arbitrates what should actually move forward.

A concrete example: instead of manually drafting a full specification from scratch for a relatively simple operational need — such as automating customer reminders in a billing workflow — she can start from a short description of the business objective. From there, the AI workflow combines product context, existing system constraints, and codebase knowledge to build a structured first version of the specification. It generates user stories, acceptance criteria, edge cases, delivery structure, and implementation considerations, while also surfacing tradeoffs, identifying unclear assumptions, and asking for missing decisions when needed. Rather than acting as a simple spec generator, the system behaves more like a product copilot connected to the delivery reality. What previously could require most of a working day can now become a strong, execution-ready first draft in roughly one hour — representing a three- to fivefold gain in speed.

But that progress also exposed a more structural issue: in startup environments, technical debt and missing documentation quickly become execution bottlenecks. Rather than treating documentation as a separate cleanup effort, she used AI to progressively reconstruct a living knowledge base from historical tickets, shipped features, legacy decisions, and existing code — including systems developed before she joined the company. What started as a manually generated first version evolved into a documentation layer continuously connected to the production codebase. The system now automatically updates parts of the documentation as new features are shipped and production changes are deployed, allowing the knowledge base to stay aligned with the actual state of the product. That documentation foundation has become a core context layer powering the different AI agents used across discovery, specification, and delivery workflows.

For Marion, success shows up in both speed and quality.

On the speed side, AI is a major accelerator. Specs and PRDs can be produced three to five times faster. Discovery work can also be dramatically accelerated: in some cases, roughly 50% of the work can be prepared in one hour.

On the quality side, the system acts as a structured memory and reasoning layer. It reduces the risk of overlooked details, produces more synthetic and harmonized documentation, and keeps outputs aligned with consistent standards and templates. Because the workflows are connected to the actual codebase and production reality, the generated specifications stay much closer to the real product context rather than becoming disconnected documentation artifacts. That consistency matters for the people consuming the work downstream — especially developers — because specifications are not only clearer and easier to navigate, but also more grounded in the actual implementation and technical constraints of the product.

🧩 AI Challenges & Pain Points

One challenge came through clearly throughout the conversation: AI adoption only sticks when the surrounding work is structured enough to support it. In practice, that means standards, documentation, checkpoints, and human review. Without that, AI may generate output quickly, but not reliably.

Marion described the difference very concretely.

Without structure, a quick prompt with little context often produces hallucinations, vague output, or something that requires several rounds of correction. The apparent speed gain disappears because the user spends too much time re-explaining, correcting, and iterating.

With structure, the dynamic changes. A shared documentation base, dedicated agents, clear templates, and a defined review process make the output much more directly usable. The human still reviews and arbitrates, but the starting point is stronger, more consistent, and closer to something the team can act on.

That is why Marion keeps a human in the loop at several stages. Her agents are not meant to replace product thinking. They are designed to create strong first drafts, reduce cognitive load, and reinforce existing quality standards. The outputs follow product conventions such as INVEST for user stories and Gherkin for acceptance criteria, which helps keep them usable and consistent.

She also highlighted the limits of the approach.

The initial setup takes time. Building the agents, refining prompts, testing outputs, and improving the workflow requires iteration — almost like building a product. The quality of the result also depends heavily on the quality of the sources. If the documentation base is incomplete, badly structured, or outdated, the agents can reproduce those weaknesses.

And there is a clear boundary: Marion does not use AI to make the decision.

AI helps prepare, synthesize, structure, and accelerate. But decisions remain human, especially when they involve prioritization, trade-offs, customer understanding, or internal alignment.

She also highlighted more tactical risks across different use cases. In transcript analysis, AI can overemphasize edge cases if you do not challenge it on frequency. In benchmarking, it is essential to ask for sources so claims can be verified. In NPS analysis, categories need to be intentionally defined, otherwise the model may generate buckets that are either too broad or too narrow to be useful. Across all of these examples, the same lesson returns: enthusiasm is useful, but discipline is what makes AI dependable.

🚀 The Use Case

Rather than one isolated flagship project, Marion shared a broader family of use cases that together show how AI is transforming product work.

The clearest example is her AI-assisted specification workflow. It began with a simple productivity question: could AI help reduce the time spent writing and structuring specs? That led to an assistant, then to a more integrated agent, and then to a much deeper realization: the value of the system depends on the quality of the context around it.

Missing documentation and scattered knowledge were limiting the usefulness of the workflow, so she used AI to help reconstruct a shared knowledge base from existing materials. This base is now used by her agents to generate better outputs. Over time, the system evolved beyond static documentation: parts of the knowledge base are now continuously updated from production and codebase changes, helping the agents stay aligned with the actual state of the product.

The delivery workflow is structured around three steps: turning ambiguous product needs into structured PRDs, organizing specifications into delivery-ready scopes, and transforming them into executable tickets. The goal is not to remove the product manager from the process, but to remove part of the repetitive drafting and formatting work so the PM can focus on review, judgment, and arbitration. The agents also help surface missing decisions, technical tradeoffs, and unclear assumptions early in the process. Because the workflows are connected to the codebase and existing system logic, specifications stay closer to delivery realities from the start.

That same logic extends upstream into discovery work.

Here too, Marion uses a chain of three agents, fed by several sources: bugs reported through support tools such as Intercom, user and sales calls, and competitor benchmarking. These inputs help structure discovery faster and surface recurring themes more efficiently. Because the workflows are grounded in broader product context, discovery outputs stay closer to existing constraints.

Similar workflows are also applied to workshop preparation and alignment sessions. For example, when preparing a session with a client or internal team, AI can help generate a first structure, identify relevant epics, draft user stories, and prepare supporting material. The result is not final, but it gives her a strong working base much faster.

The impact is practical and immediate: faster preparation, clearer first drafts, more standardized outputs, and less time lost to repetitive synthesis.

The gains are visible. Specification work can be divided by three to five. Discovery preparation can reach a meaningful first version in about one hour. Workshop preparation can be divided by around three.

But the larger lesson is methodological. In practice, the system behaves less like a content generator and more like a contextual product copilot embedded in the company’s delivery workflows. Marion’s work suggests that AI is most valuable not when it replaces judgment, but when it strengthens execution in the invisible parts of the process.

📈 Trends

The trend that excites Marion most is the speed at which AI is evolving inside product management itself. Not only because models are improving, but because their practical usefulness in day-to-day work is becoming easier to see. Productivity gains may still be difficult to measure precisely, but they are increasingly felt by the people using the tools.

At the same time, adoption is uneven.

Large companies often move more slowly, partly because of confidentiality constraints, data risk, and organizational caution. Scale-ups, by contrast, often move much faster, sometimes quickly enough to make others feel left behind. Marion also points to a cultural factor: more digitally native teams often adopt these tools more naturally.

One of the most interesting ideas from the interview is that AI at work may be shifting from individual experimentation to organizational capability. In some companies, this is no longer just a few motivated early adopters testing tools in a corner. Internal agents are being rolled out more broadly, training is being built, organizational models are adapting, and usage metrics are being tracked. That is a very different stage of maturity.

Marion is now experiencing that shift directly. What started as AI usage inside the product team is becoming a broader internal structuring effort. Now, part of her role evolves toward AI operations, with around half of her time dedicated to helping teams across the company adopt better AI practices: guidelines, agents, workflows, and tracking.

🔗 Follow Them

You can follow Marion on LinkedIn, where she shares practical learnings on product management, AI workflows, and hands-on experimentation.

Thank you for tuning in to this edition of the AI Builders Interview series! 🎙️

🔗 Dive deeper into the minds behind AI innovation—explore more conversations with the builders shaping the future, and check out 150+ real-world AI case studies in our growing library:

Discover more interviews of AI builders

✨ We hope this dialogue offered fresh insights and sparked new ideas!

🛠️ Know an AI builder doing great work? Nominate them for a future interview!

Disclaimer: This content was (obviously 😉) built with the assistance of AI.

Across Atlassian and Meta, the real shift isn’t a smarter assistant for each developer—it’s routing AI through the organization’s control points (PR review, test pipelines) and adding ranking and verification layers so engineers curate pre-vetted changes rather than wade through noise. This reassigns work: machines generate and filter, platforms enforce quality, humans decide—compressing early passes while preserving decision rights. Copying the surface (an IDE copilot) without the gates and evaluation stack will increase churn; the operating model upgrade is the propose–verify–accept loop, backed by shared infrastructure.

AI Operating Model Case — Atlassian

Organization Type: Public

Industry: Software & Technology

Initiative starting date: Not disclosed
Evidence Quality Score: 3/5

Scope of AI in Development Lifecycle: AI in engineering delivery

Description: Atlassian uses internal AI systems to augment software engineering delivery, focusing on code review rather than the full product lifecycle. Rovo Dev performs AI-assisted pull request review, automatically generating inline review comments and suggestions that developers can act on. An ML-based comment ranker filters and orders LLM-generated review comments using Atlassian’s proprietary data to reduce noise and surface the most useful feedback. A self-hosted Inference Engine centralizes AI infrastructure for production LLM and search-model workloads, acting as a shared platform layer for multiple internal engineering tools. Evidence in the notes is limited to engineering delivery; other lifecycle stages are not documented.

Nature of Integration: AI is integrated into Atlassian’s engineering workflow primarily through Rovo Dev, which plugs into the existing pull request review process to automatically analyze code changes and generate inline comments. These AI-generated comments are then passed through an ML-based comment ranker trained on proprietary signals, which filters and orders the suggestions before developers see them, reducing noise in the review UI. Underneath these tools, a self-hosted Inference Engine provides a centralized platform for running LLM and search models in production, indicating that multiple internal systems share a common AI infrastructure. The architecture therefore combines agent-like review workflows, a custom classifier for comment ranking, and a managed inference layer rather than standalone SaaS assistants. Developers remain firmly in control, using AI output as advisory input to their reviews and deciding which suggestions to accept or ignore.

Workflow Impact: Before these systems, reviewers performed a fully manual first pass over pull requests, scanning for style, correctness, and quality issues and writing all comments themselves. With Rovo Dev, an initial layer of automated review runs on each PR, generating suggested comments that highlight potential issues, while the ML ranker filters and prioritizes these comments so only the most relevant ones appear to reviewers. This reduces the time spent on low-value scanning and repetitive feedback and lets developers focus on higher-value judgment calls and nuanced code quality decisions. The workflow still functions without AI, but code review is now AI-assisted by default for participating teams, compressing early review steps and shortening review cycles, even though no numeric cycle-time reduction is disclosed. As a result, engineers’ roles shift slightly from discovering all issues manually toward supervising and curating AI-generated feedback.

Claimed Impact: The notes state that Atlassian positions Rovo Dev as a way to accelerate code reviews and reduce pull request cycle time, implying gains in engineering throughput and responsiveness. However, no concrete metrics, such as percentage reductions in review time or increases in PR throughput, are publicly disclosed in the cited sources. The impact is therefore characterized qualitatively as faster iteration and less noisy review comments, driven by AI automation and ML filtering in the review process. These claims come from Atlassian’s own materials and coverage rather than independent validation, so they should be treated as company-reported benefits without quantified evidence.

Transformation Classification: Level 2

AI is embedded into a specific engineering workflow—pull request review—where it automates the first pass of review comments and filters them through an ML ranker. This meaningfully optimizes a recurring development step and is backed by a dedicated inference platform, going beyond ad hoc use of generic coding assistants. However, the change is constrained to engineering delivery rather than representing a redesign of the entire product development operating model across discovery, prioritization, testing, and release. The evidence supports a focused workflow optimization rather than a full operating model integration.

Strategic Signal: Atlassian’s approach signals a shift from generic coding copilots toward AI embedded directly into key governance points of engineering workflow, with Engineering as the primary beneficiary. By tying AI into pull request review and layering an ML-based ranker on top, Atlassian treats review quality and noise reduction as a systematic optimization problem rather than an individual developer choice. The self-hosted Inference Engine suggests a medium level of defensibility, since replicating a centralized AI platform and proprietary ranking logic requires internal infrastructure and data, but the underlying models and concepts remain broadly accessible to other firms. For other organizations, this case highlights that early high-leverage AI integrations come from upgrading existing control points (like code review) with AI orchestration rather than building entirely new workflows. It also indicates that AI operating models can evolve incrementally, starting with narrow but deeply integrated optimizations in the engineering lifecycle.

Retrieve the case and more (including their sources) on the AI in USE website

AI Operating Model Case — Meta

Organization Type: Tech giant

Industry: Software & Technology

Initiative starting date: 2024
Evidence Quality Score: 4/5

Scope of AI in Development Lifecycle: AI in engineering delivery, AI in testing, AI in release

Description: Meta uses internal AI systems (TestGen-LLM and reported Devmate/Metamate) to shift parts of software testing and coding from manual engineer effort to AI-generated outputs with verification and review gates.

TestGen-LLM generates candidate unit-test improvements that are filtered by build, pass, flakiness, and coverage checks before engineers decide what to accept.

Devmate/Metamate is described as an internal coding assistant that analyzes failed tests, identifies issues, and submits proposed fixes for human review.

This change primarily affects engineering delivery and testing, creating a more governed, AI-assisted software workflow rather than standalone autocomplete usage.

Nature of Integration: Meta integrates TestGen-LLM into an internal engineering workflow that proposes unit-test improvements and then runs automated verification gates (build correctness, reliable passing, flakiness screening, and coverage checks) before changes reach engineer review. The TestGen-LLM system is described as an LLM-based approach using an ensemble with prompts and hyperparameters plus programmatic filtering, rather than a simple developer-side chatbot. In parallel, Devmate/Metamate is reported as an internal coding assistant that can analyze failed tests, identify issues, and submit fixes for human review, indicating a more agent-like interaction pattern in day-to-day engineering tasks. Across both systems, humans retain control through review and acceptance decisions, and the notes do not disclose full internal deployment architecture beyond the described workflow and verification steps.

Workflow Impact: Before these systems, engineers manually identified weak unit tests, wrote additional cases, iterated on failures, and debugged test regressions through standard review and CI loops. After adoption, TestGen-LLM generates candidate test additions and improvements and narrows them via automated checks, reducing the amount of manual test authoring and exploratory iteration required before an engineer sees a viable change. Devmate/Metamate is reported to offload parts of failed-test analysis and fix preparation by producing reviewable code changes rather than only offering suggestions inline. The workflow becomes more dependent on AI for accelerating routine test-improvement and debugging loops, while engineers remain responsible for reviewing and accepting changes that reach production.

Claimed Impact: Meta reports quantified outcomes for TestGen-LLM from internal use, including 75% of generated test cases building correctly, 57% passing reliably, 25% increasing coverage, 11.5% of targeted classes improved during Instagram and Facebook test-a-thons, and 73% of recommendations accepted for production deployment. These results indicate operational improvements primarily in testing quality and throughput (more test improvements reaching production) rather than a purely qualitative productivity claim. For Devmate/Metamate, the notes describe anecdotal productivity improvements (e.g., a 30-minute task reduced to 15 minutes and workload reduced by half), but these are not presented as audited or systematically measured metrics. Overall, the TestGen-LLM impact is supported by a Meta-authored paper with specific measurements, while broader coding-assistant productivity claims are less verifiable in the provided sources.

Transformation Classification: Level 2

Meta shows evidence of AI embedded into specific engineering workflows, especially unit-test improvement with automated verification and human review gates.

The documented change materially optimizes testing and parts of debugging, but the notes do not establish an end-to-end AI-run delivery operating model.

Human approval remains central, and broader agentic coding claims are reported secondhand without full architectural disclosure.

Strategic Signal: Meta’s approach signals a shift from individual AI assistance to governed workflow integration where AI outputs are systematically filtered and then reviewed before production acceptance, primarily benefiting Engineering. The defensibility comes from tying LLM outputs to internal engineering processes (test-a-thons, CI-style verification gates, and acceptance into production) and from operating at Meta’s codebase scale. Replication is moderate: other large software organizations can adopt LLM-assisted test generation, but matching the same reliability filters, evaluation discipline, and production acceptance workflow requires mature infrastructure and process rigor. The broader signal is that meaningful AI leverage in software development is increasingly driven by embedding AI into verification-heavy pipelines, not just providing developers with chat-based help.

Retrieve the case and more (including their sources) on the AI in USE website

Thank you for being part of the AI in USE community! 🌟

🔗 Visit our website AI in USE to explore the full library of AI use cases and discover how AI is transforming industries across the globe:

Explore 100+ real-world use cases

✨ We hope these AI use cases spark inspiration!
💬 Let us know which one intrigued you the most—your feedback helps us grow!

Disclaimer: This content was (obviously 😉) built with the assistance of AI.

These three launches reveal a quieter product shift: AI value is increasingly won in the “permission layer” around the model — privacy guarantees, governed access, and workflows that make outputs reusable, not just impressive. The key mistake many teams still make is shipping a clever prompt box and calling it a product, while users are blocked by risk (can I trust it?), effort (can I get to the data?), or iteration drag (can I get to a usable version fast?). The opportunity is to design AI as a system that removes avoidance behaviors — so adoption becomes the default, not a leap of faith.

Proton – Lumo Privacy-First AI Assistant

Organization Type: Scale-up
AI Purpose: Augment
Type of AI Model: Generative AI
AI Application Type: User Experience
Targeted Industry: Software & Technology
Target Group (AI User): Consumers, Employees

Use Case Description:
Proton’s Lumo is a privacy-first AI assistant for users who want the benefits of everyday AI — writing, summarizing, translating, coding, research, shopping assistance, and document analysis — without giving sensitive prompts or files to mainstream AI platforms. The core user problem is not lack of AI capability, but lack of trust: many individuals and organizations avoid using AI for confidential work because chats may be stored, reviewed, used for training, or processed through third-party infrastructure. Lumo positions itself as a direct alternative to general-purpose assistants by making privacy the primary product promise rather than a secondary setting. It offers web and mobile access, optional encrypted chat history, Ghost mode for non-persistent conversations, file upload and analysis, web search, Proton Drive integration, and Projects for organizing AI work. This lets users apply AI to personal, professional, and document-heavy workflows while keeping activity inside Proton’s privacy-oriented ecosystem. Strategically, Lumo extends Proton’s encrypted productivity suite into AI. The move resembles how DuckDuckGo differentiated search through “no tracking,” or how Apple uses privacy as a trust and ecosystem argument: the AI assistant is not positioned as the most powerful model, but as the safer default for users and organizations unwilling to trade confidentiality for convenience.

Key Features:

• Provides a private AI chat experience for writing, summarizing, translation, coding help, and general questions.

• Lets users enable encrypted chat history so past conversations remain available without being readable by Proton.

• Offers Ghost mode so chats are not saved and disappear when the session ends.

• Accepts file uploads and answers questions or generates summaries based on the uploaded documents.

• Connects to Proton Drive so users can reference selected files and folders during chat.

• Supports Projects so users can organize ongoing work and link project-specific Drive folders for repeated use.

• Keeps web search off by default and only uses it when the user turns it on.

Results:

• Reduced privacy and confidentiality barriers to using an AI assistant for sensitive prompts and documents.

• Increased reuse of Proton Drive content by enabling question-answering and summarization directly over stored files.

• Expanded Proton’s product suite from encrypted productivity tools into a privacy-first AI assistant workflow.

Launch Date: July 2025

Retrieve the case and more (including their sources) on the AI in USE website

Supaboard – Plain-English Self-Serve Business Intelligence

Organization Type: Start-up
AI Purpose: Augment
Type of AI Model: Generative AI
AI Application Type: Operations
Targeted Industry: Software & Technology
Target Group (AI User): Operations teams, Finance teams, Marketers

Use Case Description:
Supaboard is a business intelligence product that lets business teams connect company data and get KPI answers, charts, and dashboards by asking questions in plain English instead of writing SQL or waiting on analysts. The problem it addresses is the friction that prevents non-technical teams from using data in daily decisions: routine questions often require specialist support, metrics are spread across tools, and definitions are not always consistent across teams. As a result, business users either wait too long for answers or fall back on spreadsheets, intuition, or partial data. Supaboard turns a user’s question into an analysis and returns an explanation and visuals that can be reused as dashboards and scheduled reports. Operators in functions like finance, marketing, and business operations use it to investigate performance, track variance, and produce recurring reporting in one place. The measurable value is faster reporting cycles and reduced manual reconciliation work across spreadsheets and SaaS systems. Strategically, it positions BI as a day-to-day operating layer for non-technical teams, not a separate analyst queue.

Key Features:

• Connects common business data sources and makes them available for analysis in one workspace.

• Answers plain-English questions with KPI outputs, charts, and written explanations.

• Creates dashboards from user prompts and keeps them updated as underlying data changes.

• Supports deeper follow-up questions to move from a headline metric to drivers and trends.

• Applies role-based access controls so users only see the data they are permitted to view.

• Delivers recurring outputs through scheduled reports, alerts, and workflow integrations.

Results:

• Reached 13 paying companies at MVP stage.

• Ranked #1 product of the day and #1 of the week on Product Hunt in Feb. 2026.

• Reported usage by 1000+ teams.

• Reduced time to bring data together from hours to minutes.

Launch Date: April 2025

Retrieve the case and more (including their sources) on the AI in USE website

Higgsfield AI – Vibe Motion Conversational Motion Design

Organization Type: Scale-up
AI Purpose: Augment
Type of AI Model: Generative AI
AI Application Type: Sales & Marketing
Targeted Industry: Software & Technology
Target Group (AI User): Marketers, Designers

Use Case Description:
Higgsfield Vibe Motion is a motion-design workflow that turns conversational creative direction into structured animated assets for marketing and brand content. The problem it addresses is that teams responsible for social-first campaigns need many fast, on-brand variations, but traditional motion design depends on specialist tools and time-intensive keyframing. Users describe the intended look and feel—such as motion style, tone, and brand treatment—and the product generates motion graphics outputs without requiring timeline-based animation work. Marketers and designers use it to produce items like kinetic typography, logo reveals, presentations, and animated brand visuals, often by incorporating existing assets such as logos and images. This reduces iteration cycles between creative intent and technical execution, while keeping the workflow structured enough to support repeatability across campaigns.

Key Features:

• Generates motion-design outputs from conversational prompts that describe creative intent.

• Provides structured motion generation to support consistent iterations and branded variations.

• Creates common marketing motion formats such as kinetic typography, logo reveals, and presentation-style animations.

• Incorporates user-provided assets like logos, SVGs, images, and footage into animations.

• Lets users work inside a broader video workspace that supports multiple video models and creative workflows.

Results:

• Reached around $10M ARR within weeks after launching the browser-based product.

• Reported that social media marketers account for about 85% of platform usage.

• Generated roughly 4 million videos per day across the platform.

• Reduced typical creation cycles to 2–5 minutes for generated videos in the broader workflow described.

• Improved first-pass usability in the broader Click-to-Ad workflow to one or two attempts instead of five or six.

Launch Date: January 2026

Retrieve the case and more (including their sources) on the AI in USE website

Thank you for being part of the AI in USE community! 🌟

🔗 Visit our website AI in USE to explore the full library of AI use cases and discover how AI is transforming industries across the globe:

Explore 100+ real-world use cases

✨ We hope these AI use cases spark inspiration!
💬 Let us know which one intrigued you the most—your feedback helps us grow!

Disclaimer: This content was (obviously 😉) built with the assistance of AI.

These three brilliant use cases show something encouraging: the strongest AI products don’t just add intelligence, they make good work easier to start, shape, and finish. The real product lift comes from designing around context—so users can guide the system, review what it learns, and keep momentum even as workflows get complex. In other words, AI works best when the product helps people get the most out of it.

Miro – Board-aware AI for workshop synthesis and diagram generation

Organization Type: Scale-up
AI Purpose: Augment
Type of AI Model: Generative AI
AI Application Type: R&D / Product development
Targeted Industry: Software & Technology
Target Group (AI User): Product managers, Software engineers, Designers

Use Case Description:
Miro embeds AI directly inside its collaborative boards so teams turn messy workshop outputs into structured, reusable product artifacts without leaving the canvas. The core problem is that research notes, stickies, diagrams, and screenshots already live in Miro, but converting them into briefs, themes, and diagrams normally takes hours of manual synthesis and formatting. Users select relevant board objects as context and ask Miro to generate summaries, themes, action cards, and first-draft diagrams that appear as editable content on the board. Product teams and facilitators use it to clean up after discovery sessions, while engineering teams use it to draft and iterate technical diagrams faster. The measurable value is reduced time spent on synthesis and diagram setup, with the practical constraint that output quality drops on cluttered boards and users report limited control for broad, custom, multi-element instructions. Strategically, this positions Miro’s board as the primary “context layer” for product work, reducing tool-switching and keeping downstream deliverables anchored to the team’s shared workspace.

Key Features:

• Generates summaries and structured themes from selected sticky notes and workshop content.

• Creates editable diagrams from a text prompt directly on the canvas.

• Refines generated diagrams and content through follow-up instructions and manual edits in the board.

• Generates new artifacts such as briefs, tables, and timelines from selected board objects.

• Converts uploaded hand-drawn diagram images into editable digital diagrams.

• Lets administrators enable or disable AI features for teams and apply enterprise security controls.

Results:

• Users report saving hours when turning customer conversation notes and workshop outputs into structured research synthesis.

• Teams reduce time spent on first-draft diagram layout and manual shape formatting by generating an editable starting point.

• Workshop cleanup time decreases by converting clustered stickies into themes and action-ready outputs.

• AI usage is constrained by plan-based AI credits, creating adoption friction for heavy diagramming workflows.

• Output consistency declines on cluttered or large boards, increasing rework time and reducing perceived reliability.

Launch Date: May 2023

Retrieve the case and more (including their sources) on the AI in USE website

Tinder – Chemistry Curated Match Recommendations

Organization Type: Tech giant
AI Purpose: Augment
Type of AI Model: Generative AI, Computer vision
AI Application Type: User Experience
Targeted Industry: Media & Entertainment
Target Group (AI User): Consumers

Use Case Description:
Tinder’s Chemistry product changes discovery from open-ended swiping to a smaller set of curated profile recommendations designed to reduce “swipe fatigue” and improve conversation quality. The problem it addresses is that users face too many low-relevance profiles and low confidence that time spent browsing leads to meaningful chats. Chemistry uses AI signals from interactive Q&A and, when a user opts in, visual analysis of selected camera-roll photos to infer interests and help tailor recommendations and profile presentation. Users interact with the system by answering prompts, reviewing and approving Photo Insights, and managing what the product has learned about them. The intended value is faster preference learning for better early-session relevance, less time spent sifting through profiles, and stronger conversion from matches to multi-message conversations. Strategically, the feature positions Tinder’s turnaround around recommendation quality and user trust, while making privacy controls and transparency central to adoption.

Key Features:

• Replaces endless swiping with “Drops” that present a limited set of curated profile recommendations.

• Asks users interactive questions to capture preferences and personality signals for matching.

• Generates Photo Insights from user-selected camera-roll photos that the user reviews before saving.

• Suggests profile photos and creates collages to help users build stronger profiles with less manual effort.

• Lets users view and delete Chemistry insights in an in-app control hub.

Results:

• Improved Tinder “Sparks Coverage” from -1% YoY in Dec. 2024 to +4% YoY in Dec. 2025.

• Improved total “Sparks” from -11% YoY in Dec. 2024 to -5% YoY in Dec. 2025.

• Recorded a $14M negative impact on Tinder direct revenue from product testing, because Tinder prioritized product experiments and user-outcome testing over short-term monetization during the rollout cycle.

Launch Date: November 2025

Retrieve the case and more (including their sources) on the AI in USE website

Qevlar AI – Autonomous SOC Alert Investigation Assistant

Organization Type: Start-up
AI Purpose: Automate
Type of AI Model: Generative AI
AI Application Type: Risk management
Targeted Industry: Software & Technology, Business Services
Target Group (AI User): Security teams

Use Case Description:
Qevlar AI provides an autonomous investigation layer for Security Operations Centers that turns incoming security alerts into completed investigations and documented decisions. SOC teams face high alert volumes, inconsistent manual triage, and analyst fatigue, which leaves limited time for proactive security improvement. The product ingests alerts from existing security tools, enriches them with internal and external context, determines whether they are malicious or benign, and generates a structured incident report with recommended next actions. SOC analysts use the outputs to validate conclusions, escalate true incidents, and close benign cases with audit-ready documentation, while SOC leaders use the standardized investigations to reduce backlog and improve consistency. The measurable impact is faster investigations and higher alert coverage, with a large share of benign tickets closed automatically. Strategically, the platform captures investigation work as reusable organizational intelligence rather than losing it in one-off ticket closures.

Key Features:

• Automatically starts an investigation when an alert arrives from SIEM and endpoint detection tools.

• Enriches each alert with relevant internal telemetry and external threat context.

• Classifies alerts as malicious or benign with a documented rationale for review.

• Generates a structured incident report that consolidates evidence, investigation steps, and recommended remediation actions.

• Closes benign tickets automatically based on configured rules and review thresholds.

• Supports MSSP multi-tenant operations with separate customer context and isolated investigations.

Results:

• Reduced average investigation time to 3 minutes.

• Delivered a 10x reduction in investigation time.

• Enabled automatic closure of up to 80% of tickets.

• Reported 99.8% classification accuracy.

• Increased investigation coverage to 100% of alerts with full context.

• Reported an average 300% ROI for MSSP deployments.

Launch Date: Invalid Date

Retrieve the case and more (including their sources) on the AI in USE website

Thank you for being part of the AI in USE community! 🌟

🔗 Visit our website AI in USE to explore the full library of AI use cases and discover how AI is transforming industries across the globe:

Explore 100+ real-world use cases

✨ We hope these AI use cases spark inspiration!
💬 Let us know which one intrigued you the most—your feedback helps us grow!

Disclaimer: This content was (obviously 😉) built with the assistance of AI.

Both cases point to a practical design rule: make AI changes local enough to own end-to-end, and make the control explicit. One overlays AI on mature rituals to collapse friction; the other rebuilds a single product stream around evaluation, retrieval quality, and tight release cadence. The hidden constraint isn’t model choice but scaffolding—structured workflows, shared context, or a proprietary knowledge base—without which automation amplifies noise instead of throughput; pick the smallest loop you can measure and govern, then scale.

AI Operating Model Case — Alan

Organization Type: Scale-up

Industry: Healthcare

Initiative starting date: Not disclosed
Evidence Quality Score: 4/5

Scope of AI in Development Lifecycle: AI in discovery, AI in user research, AI in engineering delivery, AI in post launch analysis

Description: Alan uses AI as an internal productivity layer across product discovery and engineering support, without changing its underlying product operating model. In discovery, AI automates finding relevant users, drafting SQL, sending personalized interview invitations, and scheduling interviews, turning a manual weekly task into an automated workflow. In engineering, Slack-embedded Dust assistants answer technical questions and centralize knowledge sharing between engineers. These systems sit on top of Alan’s structured product model (Areas, Crews, quarterly reviews, 6-week cycles), compressing high-friction tasks rather than replacing planning or prioritization.

Nature of Integration: Alan integrates AI as a set of embedded workflow tools rather than a single centralized platform. For product discovery, it chains LLM-based components with automation services (e.g., SQL generation, email personalization, and calendar booking) to create semi-automated interview workflows triggered by product teams. For engineering, Dust-powered assistants are integrated directly into Slack, where engineers ask questions and receive AI-generated answers enriched by shared technical context. The architecture relies on external large language models and agentic workflows orchestrated across SaaS tools, rather than fully custom in-house models. Humans remain in full control: they define targeting rules, approve outreach, conduct the interviews, and validate or refine AI answers in Slack threads.

Workflow Impact: Previously, PMs manually identified users, wrote SQL or asked analysts for help, crafted invitation emails, and iterated through scheduling back-and-forth for interviews; the new AI workflow automates user selection, SQL drafting, outreach, and calendar scheduling. This reduces the operational overhead of running recurring user interviews and lets PMs maintain a higher cadence of discovery without extra coordination cost. On the engineering side, the Slack-based assistant replaces part of the informal Q&A that required waiting for colleagues to respond in real time, turning many questions into self-service lookups with AI-synthesized answers. AI is encouraged and operationally useful but not mandatory: core processes like roadmap definition, prioritization, and release management still run on Alan’s existing product rituals. Roles shift slightly, with PMs spending more time on discovery conversations and less on logistics, and engineers relying more on shared, AI-accessible knowledge than on direct one-to-one support.

Claimed Impact: Alan reports that the automated discovery workflow reduced interview-scheduling time by 90%, from about one hour per week to around five minutes, while enabling 2–3 user interviews per week. For engineering, the Dust assistant is described as one of the most used internal agents, with 22,608 questions asked and an estimated 10–20% reduction in project completion time. These outcomes primarily represent cycle-time reduction and productivity gains in specific sub-workflows rather than a change in overall product throughput metrics. All impact figures come from Alan and Dust’s own case descriptions, so they are directionally informative but not independently verified.

Transformation Classification: Level 2

Alan applies AI to compress specific high-friction workflows—user recruitment and scheduling for interviews, and internal engineering Q&A—while keeping its broader product operating system (Areas, Crews, quarterly reviews, 6-week cycles) intact. There is no disclosed AI integration in core prioritization, PRD writing, design, testing, release, or incident workflows. The change is therefore best characterized as targeted workflow optimization within existing structures, not a re-architecture of the overall operating model. AI is layered onto established routines to make them faster and less manual rather than redefining how product and engineering work end-to-end.

Strategic Signal: Alan’s approach shows how engineering and product teams can gain meaningful efficiency by embedding AI into clearly defined, repeatable workflows instead of trying to overhaul their entire operating model. The strategy has moderate defensibility: the underlying tools (Dust, Slack, automation and scheduling platforms) are broadly available, but the effectiveness depends on Alan’s existing discipline in product rituals and its willingness to formalize discovery and Q&A flows. Engineering benefits most from turning knowledge access into a Slack-native, always-on capability that complements peer support rather than replacing it. Because the architecture is based on standard components, other companies can replicate the pattern, but they need sufficiently structured processes and data to see similar gains. The broader signal is that AI’s near-term leverage often comes from surgical automation of operational bottlenecks inside a mature product organization, not from wholesale reinvention of the development lifecycle.

Retrieve the case and more (including their sources) on the AI in USE website

AI Operating Model Case — Stack Overflow

Organization Type: Other

Industry: Software & Technology

Initiative starting date: 2026
Evidence Quality Score: 3/5

Scope of AI in Development Lifecycle: AI in discovery, AI in user research, AI in prioritization, AI in design, AI in engineering delivery, AI in testing, AI in release, AI in post launch analysis

Description: Stack Overflow’s main internal operating-model change is a structured, AI-focused product-development stream for AI Assist, spanning discovery, alpha/beta validation, model evaluation, architecture integration, and recurring release iteration. This stream relies on systematic user interviews, surveys, community feedback, model benchmarking, and post-launch monitoring rather than a one-off feature launch. The change is concentrated in the end-to-end lifecycle for AI Assist itself, not in company-wide engineering operations or generic internal AI tools. Cross-functional teams align product research, UX design, search/reranker tuning, LLM/RAG architecture, microservice integration, authentication, and release management around this AI-centric workflow.

Nature of Integration: Stack Overflow integrates AI primarily through a dedicated AI Assist product-development stream that embeds AI considerations at each lifecycle stage. Teams run user research and discovery specifically to inform LLM/RAG design, then feed benchmark results and search/reranker tuning into the engineering pipeline that powers AI Assist’s answer generation. The underlying architecture uses large language models coupled with retrieval over Stack Overflow / Stack Exchange content and internal search infrastructure, with additional microservices and authentication (e.g., JWT plumbing) to connect into the existing stack. AI is not an autonomous internal agent; instead, it is a core component in the product architecture whose behavior is shaped by model evaluation, prompt optimization, and relevance tuning overseen by humans. Human control remains high, with PMs, engineers, and designers defining requirements, running experiments, and governing releases while AI handles answer generation within the product.

Workflow Impact: The initiative does not replace internal PM or engineering tasks end-to-end but restructures how the AI Assist team works from discovery through release. Previously, a feature might have shipped after limited experimentation; now, user interviews, surveys, alpha/beta usage, and model benchmarks form formal inputs into model selection, prompt design, and search/reranker tuning before and after launch. Manual ad hoc evaluation of answer quality is partially replaced by more systematic model benchmarking and search relevance experimentation aligned to the RAG pipeline. AI becomes a dependency for delivering and iterating AI Assist, especially for answer quality, response speed, and UX behavior, so the product team’s workflow is built around evaluating and tuning these AI components. The cadence of work shifts toward recurring, AI-focused updates (e.g., faster responses, shareable conversations, markdown rendering, richer context links), requiring tighter coupling between product, design, and infrastructure roles for each iteration.

Claimed Impact: Stack Overflow reports measurable improvements from this AI-centric workflow, including at least a 35% reduction in AI Assist response times after refining search relevance, reranking, and prompts. The company also cites adoption and engagement metrics such as over 285,000 technologists using AI Assist, up to 6,400 daily messages among the most active users, and 75% of discussions focused on highly technical content. These outcomes indicate both cycle-time improvements in the AI experience and substantial usage, suggesting that the iterative model-evaluation-and-release loop delivers tangible product performance gains. All metrics originate from Stack Overflow’s own communications, so they are directionally informative but not independently verified.

Transformation Classification: Level 2

The case shows a structured, AI-specific optimization of the product-development workflow for AI Assist, tying together discovery, user research, benchmarking, RAG/search tuning, and release iteration. AI is central to this product stream, but there is no evidence of AI reshaping company-wide engineering norms such as PRD drafting, generalized prioritization automation, or broad test/incident workflows. The change is therefore best characterized as workflow optimization around a flagship AI initiative rather than a full operating-model shift for the entire organization.

Strategic Signal: Stack Overflow’s operating model around AI Assist signals that engineering teams can anchor development workflows on continuous AI evaluation and tuning when they control a high-quality, domain-specific knowledge corpus. The approach is moderately defensible because it leverages proprietary Q&A data, established search and relevance infrastructure, and tight integration with existing microservices, making it harder to replicate without similar assets. Replication is still feasible for other organizations with strong internal content and search capabilities, but it requires coordinated product, engineering, and infra investment rather than just plugging in an off-the-shelf LLM. Engineering is the primary beneficiary function, since its delivery rhythm is now organized around model benchmarking, RAG/search adjustments, and AI-driven performance metrics. More broadly, the case illustrates that meaningful AI operating models often emerge first around a single AI-intensive product stream before expanding into organization-wide engineering practices.

Retrieve the case and more (including their sources) on the AI in USE website

Thank you for being part of the AI in USE community! 🌟

🔗 Visit our website AI in USE to explore the full library of AI use cases and discover how AI is transforming industries across the globe:

Explore 100+ real-world use cases

✨ We hope these AI use cases spark inspiration!
💬 Let us know which one intrigued you the most—your feedback helps us grow!

Disclaimer: This content was (obviously 😉) built with the assistance of AI.

This week’s cases highlight an uncomfortable truth: most AI products fail not because the model is weak, but because the workflow still forces humans to do the slow, credibility-critical stitching at the end. The strongest implementations don’t chase maximal autonomy — they compress the feedback loop (review, validation, latency) so outputs are usable in the moment they matter. The product lesson: treat “time-to-trust” as a first-class metric, and design the system around the points where users hesitate, verify, or abandon.

🧠 Grammarly — shifts writing from post-editing to pre-send confidence, embedding credibility checks where decisions happen

🧱 Fujitsu — turns modernization from expert-dependent archaeology into reviewable artifacts teams can act on fast

⚡ Kog — makes real-time AI experiences viable by removing latency as the hidden tax on product iteration

Grammarly – Context-aware writing agents for faster, higher-quality communication

Organization Type: Scale-up
AI Purpose: Augment
Type of AI Model: Generative AI
AI Application Type: User Experience
Targeted Industry: Software & Technology, Education
Target Group (AI User): Consumers, Students & learners, Employees

Use Case Description:
Grammarly is no longer only the spelling-and-tone assistant people use inside Gmail, Google Docs, LinkedIn, or workplace apps. With Grammarly Docs and its AI agents, the company is expanding from correcting text after it is written to helping users draft, revise, validate, and review writing before it is shared or submitted. The Coda and Superhuman acquisitions signal a broader ambition to become an AI productivity platform for communication-heavy work, although the clearest current use case remains writing quality, writing feedback, and writing confidence. The problem Grammarly addresses is the friction of producing clear, polished, credible, and audience-appropriate writing across professional and academic contexts. Users do not only need typo correction; they need help improving clarity, tone, structure, and argument quality, while identifying where claims may need support or where the text may not meet reader or evaluator expectations. In Grammarly Docs, specialized agents such as Reader Reactions, AI Grader, Citation Finder, and, in the current agent directory, Fact Checker give users feedback on how their writing may land, whether it meets a rubric, and where credibility gaps may remain. Across browser, desktop, mobile, and workplace apps, Grammarly still keeps its original value proposition: inline proofreading, rewriting, paraphrasing, and tone adjustment at the point of writing. The strategic move is therefore twofold: defend Grammarly’s core “writing assistant everywhere” position while expanding into a more complete AI writing workflow. The differentiation is not that Grammarly alone offers rewriting or AI feedback — many tools now do. Its stronger claim is distribution, trust in writing workflows, and a tighter focus on communication quality, credibility signals, and user control rather than generic document generation.

Key Features:

• Provides inline proofreading suggestions for clarity, structure, and audience fit as users write.

• Generates paraphrases that adapt tone, style, and custom voice to match the user’s intent.

• Predicts likely reader reactions, including key takeaways, confusion points, and open questions.

• Estimates an assignment grade and gives rubric- and course-aligned feedback before submission.

• Flags claims that need support, suggests supporting and disputing evidence, and formats citations.

• Checks text against databases, academic papers, websites, and published works for plagiarism.

• Scores the likelihood that text is AI-generated to support review and policy workflows.

Results:

• Reach: 40M+ daily users, 50,000 organizations, and 3,000 education institutions.

• Distribution: planned rollout across 500,000+ websites and apps.

• Business scale: Reuters reported >$700M annual revenue and profitability.

• AI-agent impact: no disclosed metrics yet on adoption, accuracy, time saved, or learning outcomes.

Launch Date: August 2025

Retrieve the case and more (including their sources) on the AI in USE website

Fujitsu – AI Legacy Code Design-Document Generator

Organization Type: Corporate
AI Purpose: Augment
Type of AI Model: Generative AI
AI Application Type: Operations
Targeted Industry: Software & Technology, Financial Services
Target Group (AI User): IT administrators / System integrators, Software engineers

Use Case Description:
Fujitsu Application Transform is a SaaS capability that helps modernization teams understand legacy systems by turning source code into usable design documents. Many enterprises running COBOL and other older languages lack current documentation, and critical system knowledge sits with a shrinking pool of experts, slowing maintenance and migration programs. The product analyzes the existing codebase and generates design documentation that reflects current specifications, with linked relationships intended to reduce missing or incorrect details. Legacy-system engineers and systems integration teams use it during discovery and planning to replace manual reverse-engineering work. Fujitsu reports large reductions in documentation effort and improved completeness and readability versus conventional approaches, accelerating modernization timelines while keeping human review in the loop.

Key Features:

• Analyzes COBOL and Java source code to extract system structure and behavior.

• Generates design documents automatically from analyzed source code.

• Links related code elements in the generated output to support more complete documentation reviews.

• Exports generated documents in Markdown and Excel formats for reuse in enterprise workflows.

Results:

• Reduced design-document generation work time by approximately 97%.

• Improved COBOL design-information comprehensiveness by 95% versus general generative AI and conventional analysis comparisons.

• Improved design-document readability by 60% versus conventional methods.

Launch Date: March 2026

Retrieve the case and more (including their sources) on the AI in USE website

Kog – Real-time LLM Inference Engine for Low-Latency AI Apps

Organization Type: Start-up
AI Purpose: Augment
Type of AI Model: Generative AI
AI Application Type: R&D / Product development
Targeted Industry: Software & Technology
Target Group (AI User): Software engineers, Data & analytics, IT administrators / System integrators

Use Case Description:
Kog provides a drop-in inference engine that speeds up token generation for companies serving LLM-powered applications where every second of delay affects the user experience. The business problem is that standard inference stacks are optimized for throughput and batch serving, which leaves real-time products constrained by per-request latency and GPU cost. Kog’s product replaces the inference layer while keeping compatibility with common vLLM-style interfaces, so teams run existing LLM workloads faster without rewriting application code. Software engineers and platform teams use it to improve responsiveness for interactive experiences such as voice agents, coding agents, and other latency-sensitive AI features. The measurable value is faster responses on the same hardware and improved GPU utilization, which reduces serving cost per request and makes real-time AI features viable at scale. Strategically, Kog positions inference performance as a competitive lever for AI product teams by enabling more iterative agent loops within a user-acceptable wait time.

Key Features:

• Provides a vLLM-compatible inference endpoint to swap in faster serving without changing client integrations.

• Accelerates sequential token generation to reduce end-user waiting time in real-time LLM experiences.

• Optimizes multi-GPU communication to reduce cross-device latency during inference.

• Packages evaluation and deployment options via API access and container-based distribution.

Results:

• Up to 3.5× faster token generation versus vLLM and TensorRT-LLM in an AMD Instinct MI300X benchmark.

• Cross-GPU latency of 4 microseconds, up to 4× faster than existing communication libraries in the same benchmark.

• 1,368 tokens per second on a sequential-generation benchmark for Llama-3 8B.

• Published benchmark coverage across 1B to 32B active-parameter models including Llama, Mistral, and Qwen.

Launch Date: July 2025

Retrieve the case and more (including their sources) on the AI in USE website

Thank you for being part of the AI in USE community! 🌟

🔗 Visit our website AI in USE to explore the full library of AI use cases and discover how AI is transforming industries across the globe:

Explore 100+ real-world use cases

✨ We hope these AI use cases spark inspiration!
💬 Let us know which one intrigued you the most—your feedback helps us grow!

Disclaimer: This content was (obviously 😉) built with the assistance of AI.

The real change is that coordination work—turning signals into tickets, tickets into code, and incidents into actions—is being encoded as event-driven agents wired into systems of record, not left to human glue. That compresses cycle time but also moves governance upstream: you manage integrations, routing logic, metrics, and ownership (the “superb builder,” the DRI), rather than supervising every step in real time. The mistake is to deploy copilots without recoding handoffs and decision rights—doing so accelerates output while your old review and control bottlenecks harden.

AI Operating Model Case — Coinbase

Organization Type: Public

Industry: Financial Services

Initiative starting date: 2024
Evidence Quality Score: 3/5

Scope of AI in Development Lifecycle: AI in discovery, AI in feedback clustering, AI in prioritization, AI in prd spec, AI in design, AI in engineering delivery, AI in testing, AI in release, AI in post launch analysis

Description: Coinbase has shifted from ad hoc LLM usage to a structured, SDLC-spanning operating model where engineers use tools like Cursor, Copilot, Claude Code, internal MCP integrations, and custom agents to move from feedback and tickets to code, tests, reviews, and incident workflows faster. AI now assists with PR drafting, test generation, feedback-to-ticket conversion, and context-driven actions across systems such as Linear, Slack, Datadog, Sentry, and Snowflake. This sits primarily in engineering delivery and incident management but touches discovery, prioritization, PRD, and design in lighter ways. Organizationally, Coinbase tracks AI usage with leadership metrics, runs company-wide enablement, and has created a dedicated adoption role to drive consistent usage.

Nature of Integration: Coinbase integrates AI through a hybrid model that combines individual developer tools (Cursor, Copilot, Claude Code) with deeply embedded agents wired into core engineering and incident-management systems. Internal MCP-style integrations let agents operate over Linear, Slack, Datadog, Sentry, Snowflake, and other platforms so engineers can trigger context-aware actions directly from their existing workflows. The architecture is described as agentic workflows that route tasks and context across tools rather than a single monolithic AI platform. Large language models are orchestrated against Coinbase’s internal context and routing layer, but the specific underlying model vendors are not detailed. Humans remain in control of initiating tasks, reviewing AI-generated code, tests, and tickets, and approving changes, while agents handle much of the transformation and coordination work between systems.

Workflow Impact: Previously, engineers and teams manually converted feedback into tickets, wrote PRs and tests from scratch, triaged bugs and incidents by hand, and coordinated reviews and incident workflows across multiple tools. With the AI operating model, many of these steps are partially automated: agents help turn feedback into structured tickets, draft PRs and tests, expand test coverage, and drive incident triage and coordination using live context from monitoring and data tools. This makes AI an encouraged and increasingly embedded part of day-to-day engineering work, though it is not formally mandatory in every step. The net effect is faster movement from signal to code and from code to review, with shorter PR review cycles and compressed feedback-to-release paths. Roles also shift: engineering managers are expected to spend less time in meetings and more in code, and Coinbase has introduced a “superb builder” role focused on driving AI adoption and workflow redesign.

Claimed Impact: Coinbase reports several quantitative indicators of impact from this AI-enabled workflow. PR review cycle time has been reduced from roughly 150 hours to around 15 hours, indicating a substantial cycle-time improvement. In highly instrumented “speedrun” events, engineers achieved about 70 PRs in 15 minutes and 300–400 PRs in 30 minutes, suggesting bursts of dramatically higher throughput. Adoption is also high, with every engineer reported to have used Cursor by February 2025, pointing to broad behavioral change rather than a niche experiment. These figures are company-reported metrics from internal writeups and leadership interviews, so they are directionally informative but not independently verified.

Transformation Classification: Level 3

Coinbase has moved beyond isolated AI tools to build an internal, agentic layer that orchestrates work across ticketing, coding, testing, and incident systems. Leadership tracks AI usage and impact, and a dedicated “superb builder” role and enablement programs exist to standardize workflows across more than 1,000 engineers. The operating model from feedback to code, review, and incident handling is being structurally redesigned around AI assistance, rather than AI being a peripheral productivity add-on.

Strategic Signal: Coinbase’s approach signals that engineering functions can gain durable advantage by building a context-rich AI layer on top of their existing SDLC and incident tooling rather than relying solely on off-the-shelf copilots. The strategy is moderately defensible because it depends on custom integrations, internal routing logic, and organization-wide enablement practices that are specific to Coinbase’s stack and scale. At the same time, other large engineering organizations with mature tooling can replicate the basic pattern, so the main barrier is execution quality, not exclusive technology. Engineering is the primary beneficiary, as routine translation work between feedback, tickets, code, and incidents is offloaded to agents, freeing engineers and managers to focus on higher-leverage decisions. The broader signal is that AI’s next productivity gains in software development will come from integrated, agentic workflows that span multiple systems, backed by explicit leadership sponsorship and change management.

Retrieve the case and more (including their sources) on the AI in USE website

AI Operating Model Case — Dust

Organization Type: Start-up

Industry: Software & Technology

Initiative starting date: Not disclosed
Evidence Quality Score: 2/5

Scope of AI in Development Lifecycle: AI in engineering delivery

Description: Dust integrates an internal agent platform with a PM-light product model where engineers act as DRIs for initiatives. Agents are embedded into core tools like Slack, Notion, GitHub, and Confluence, with Temporal workflows triggering AI processing whenever messages, documents, or pull requests change. This shifts internal workflows from ad-hoc manual coordination and context-gathering toward event-driven automation and self-serve knowledge access. Organizationally, engineers take broader end-to-end ownership while AI agents handle much of the glue work previously done through manual search, ticket creation, and cross-tool routing.

Nature of Integration: Dust uses AI through an internal agent platform that is directly embedded into everyday tools such as Slack, Notion, GitHub, and Confluence. Temporal workflows orchestrate these agents, triggering AI runs when events occur, like a new Slack message, an updated Notion page, or an opened pull request. The architecture is explicitly agentic and event-driven, with agents retrieving context from multiple systems and then taking actions such as commenting on PRs or creating structured tickets. Humans design, configure, and test agents in a preview environment and decide when to publish them into production workflows, while the agents then operate autonomously in response to live events. This places AI as an integrated layer within Dust’s internal systems rather than as a separate, optional assistant.

Workflow Impact: The system reduces manual searching across fragmented tools and eliminates many low-level coordination tasks such as turning a Slack-reported bug into a properly formatted ticket or manually bringing design-doc context into code review. Previously, team members would have had to switch between tools, copy-paste information, and remember to follow up on issues; now, agents automatically route context and trigger next steps when predefined events occur. Because these workflows are driven by live events, work becomes more continuous, with fewer batching or periodic coordination rituals needed to move information between systems. While engineers already own initiatives as DRIs under Dust’s PM-light model, AI agents further support their broader scope by lowering the operational and knowledge-management overhead. In practice, this results in engineers spending relatively more time on substantive product and code decisions and less on mechanical orchestration across tools.

Claimed Impact: Dust discloses quantified impact primarily through a customer case (Assembled) rather than its own internal product organization. In that case, Dust reports 95% internal adoption across 120+ employees and “hundreds of hours saved,” framing the benefit as both time savings and widespread reliance on agentic workflows. No equivalent numerical metrics are provided for Dust’s own engineering or product teams, so the internal effect is described qualitatively only. The impact claims therefore rest mainly on company-provided narratives and a single detailed customer example, which suggests meaningful operational gains but lacks independent or internally quantified validation.

Transformation Classification: Level 3

Dust combines a non-traditional ownership model—engineers as DRIs with minimal PM layer—with an internally built agent platform embedded in core communication and knowledge tools. Event-driven agents now handle cross-tool coordination, context retrieval, and action triggering, which structurally alters how work is initiated and routed. This goes beyond individual AI assist: it reconfigures ownership, reduces handoffs, and makes automated workflows a standard part of how engineering delivery operates. However, there is limited direct evidence that AI is systematically embedded across every product-development stage.

Strategic Signal: Dust’s approach signals that engineering organizations can pair an engineer-owned product model with deeply integrated agents to reshape internal operations. The primary beneficiaries are engineers, who gain broader product responsibility while offloading a significant amount of search, coordination, and follow-up work to AI-driven workflows. Strategically, the model is moderately defensible because it depends on well-designed internal integrations, event orchestration, and organizational norms rather than only on generic LLM capabilities. Replication is feasible for companies with similar technical maturity and willingness to adjust roles, but it is more complex than simply rolling out a generic copilot. The broader signal is that meaningful AI operating-model change comes from combining organizational redesign with agentic infrastructure woven into daily systems, not from isolated tooling experiments.

Retrieve the case and more (including their sources) on the AI in USE website

Thank you for being part of the AI in USE community! 🌟

🔗 Visit our website AI in USE to explore the full library of AI use cases and discover how AI is transforming industries across the globe:

Explore 100+ real-world use cases

✨ We hope these AI use cases spark inspiration!
💬 Let us know which one intrigued you the most—your feedback helps us grow!

Disclaimer: This content was (obviously 😉) built with the assistance of AI.