As long as each level makes money, the AI pie keeps growing, and massive spending will be justified. If one level fails to monetize its part, the AI pie stops growing. Massive spending at all the levels above it inflates a bubble. When that level’s monetization stalls for too long, shareholders, VCs, and lenders start asking hard questions.

If they don’t like the answers, their investment thesis changes (or is forced to change) and the bubble pops.

So far, each level has benefited from AI amplifying AI to grow the pie. AI and agents need chips. Chips need data centers. Compute in data centers needs distribution. Those are all working, but AI and agents need a new part of the cycle to keep the momentum going.

AI and agents must grow the other markets and ecosystems to escape our current circular AI economy. One layer of the AI ecosystem selling to another layer has reached as far as it can go. Extending growth beyond the AI ecosystem is where the challenges begin, and last week’s earnings revealed that.

The Three Layers Of AI Monetization

The first-order effects of the AI boom went to the hardware manufacturers. GPUs, TPUs, memory, and the inputs they require were the first-round winners. But are the companies buying all that hardware making money?

Yes. The second-order effects are going to the hyperscalers and the inputs required for their data center buildouts. Power and cooling companies have done well. Hyperscalers are reporting accelerating growth and massive backlogs of demand. But are the companies using all that compute making money?

Well, no. Not really. There’s an old joke in academia. If you get a degree and cannot find a job, you can always make a living teaching others the skills you couldn’t make a living with. Mark Zuckerberg and Elon Musk have essentially said the same things about their data centers. If they can’t make money with Muse and Grok (their own AI and agents), they can always make money selling the compute to everyone else who is building AI and agents.

Even the hyperscalers and AI labs can’t seem to make the unit economics work for third-order monetization. They are selling compute to enterprises hoping their customers can solve the problem they have yet to unravel. Much of their hopes are pinned on AI SaaS companies, but those earnings stories are troubling. That’s the big picture the market is reacting to.

It’s the story I will tell in this article, and I will explain how we can fix it.

Sometimes A Year Happens In A Week

Anthropic really chose the wrong week to float its IPO and financial filings. 3 months ago, no one would have gone through them with the eye for detail that analysts have brought to them this week. The story would have been 85% margins (with significant financial engineering), not $750 billion in spending commitments, negative unit economics, rising competition, and an unclear path to profitability.

You can be forgiven for not remembering Anthropic’s filing happened this week with everything that has happened since. Intel and Reddit shares had their worst days on record. ServiceNow’s shares gave a brave salute before jumping out of an airplane. Meta, Amazon, and Alphabet (Google) reported disappointing earnings.

Google lost half a trillion in market cap this week. The NASDAQ walked off a cliff in sympathy. AWS CEO Matt Garman tried to calm investors but instead had a crash-out interview Friday on Bloomberg after Ed Ludlow dropped a trap question on him (I’ll explain why in a bit).

China’s AI boom went into a correction in advance of all this, which didn’t help matters. The neoclouds like CoreWeave and Nebius have challenges around rising borrowing costs eating into their unit economics. Microsoft’s bond sale was announced right after its earnings announcement. In an environment of debt doubt, the timing soured what was initially seen as a positive story.

The Problem At Level 3

Ed Ludlow loaded up a question for the CEO of AWS on Friday. Everyone says we’re early in the AI cycle. “What hasn’t happened yet?” I literally yelled at the screen…

I have been discussing this with clients and students. Generative AI has already entered a commoditization cycle. Using the ‘It’s Early’ narrative is begging the question that Ed Ludlow dropped on Garman. AWS’s CEO walked right into it.

“{Our customers are} just in the early stages of figuring out how they get value out of this.” He didn’t say the early stages of getting value out of AI. He said the early stages of figuring out how to get value. To rephrase it more clearly: “Our customers have had 4 years with AI and over a year with agents, but still haven’t figured out how to get value out of AI, and most just started trying.”

Garman set a timeline for level 3 ROI to materialize. If the enterprises and tech companies buying compute from the hyperscalers haven’t figured out how to get value out of AI, they are probably a year or more away from seeing revenue growth materialize. That gap is big enough to inflate a bubble.

Reevaluating The AI Thesis

Investors, traditional lenders, the bond market, and VCs all heard the message loud and clear. The tech companies and AI startups you’re giving money to need their customers to keep growing the AI pie to pay you back, and their customers don’t have a clear line of sight to ROI. Smart money uses the same question about a consensus view to ensure they don’t fall for a bubble.

“What would have to change in order for your thesis to change?”

If enterprises, Big Tech, and AI labs can’t figure out the level 3 monetization puzzle, the AI pie stops growing. It stays circular, and that begins to change the thesis behind most investment. Even OpenAI is being forced to lower its IPO funding target and valuation.

The risk of a bubble is growing, but level 3 taking longer to monetize isn’t catastrophic as long as enterprises keep spending with the hyperscalers while they do it.

What would have to change for enterprise spending to go flat or come down?

It’s not enough for new revenue to be slow to materialize. That will only result in an S-curve. Something would need to threaten the enterprises’ existing revenue for spending to plateau, then turn down. Parse through Big Tech earnings and you’ll find the technology itself is part of the problem.

It’s not that AI and agents don’t work. They do. It’s that the AI and agentic paradigms are monetized in very novel ways and have novel impacts on existing business models, marketplaces, and ecosystems.

Reading The Earnings Tea Leaves: Margins

Usage is up across nearly every tech company that reported, but the margin on each unit of use is falling. This week’s earnings told two stories about how the third level of the AI cycle is playing out. The first story is around the question of margins and ties into the last part of the Wharton professor’s quote. AI needs viable business models outside of the AI ecosystem for level 3 to monetize successfully.

Microsoft Copilot users are up by 50%, but paid Copilot 365 seats are only 6.7% of the 450 million Microsoft 365 seats. AI isn’t driving new revenue by bringing new customers into the Microsoft 365 ecosystem faster. Most are existing users on a consumption-based pricing model, but commercial cloud revenue only rose by 17% even with over 50 million agents running on Copilot. This is a bad sign for the AI SaaS business model, and ServiceNow has raised even larger red flags.

ServiceNow saw agentic consumption rise by 13X, but signed contracts missed estimates, subscription margins are falling on higher cloud costs, and services margins went negative on higher agent implementation costs. Again, consumption is rising, but AI isn’t driving new customers onto ServiceNow’s platform faster than before. The cost of rising consumption is eating into ServiceNow’s margins for existing customers.

ServiceNow tried to switch from per-seat pricing to outcome-based pricing, but Bill McDermott revealed that customers won’t do it. Enterprises won’t pay more just to have an agent deliver the same outcome, and most SaaS vendors haven’t figured out how to use agents to deliver new outcomes. This is where the novel monetization and technical capabilities show up.

AI SaaS needs to grow the pie for its customers to make the switch to outcomes-based pricing, but all it is doing now is attempting to shift more money into vendors’ pockets. AI SaaS is delivering digital outcomes with more expensive technology, but they can’t charge more for it. This leads to another facet of the margin story.

Reading The Tea Leaves: Agents Aren’t Plug & Play

Higher implementation costs are even more troubling. Level 3 monetization requires enterprises to figure out how to use AI and agents to grow their business models. ServiceNow, Microsoft, IBM, Anthropic, and many others have turned to forward deployed engineers to help customers figure it out.

Conventional wisdom for level 3 monetization was that enterprises would buy more from retooled AI SaaS platforms. The buy route would give quick access to AI and agents and deliver quick wins to enterprises outside of Big Tech. That would drive more spending from outside the AI ecosystem.

Early level 3 monetization estimates were built on this assumption. ServiceNow revealed that enterprises can’t figure it out on their own and need ServiceNow’s FDEs. But if ServiceNow, Microsoft, IBM, and Anthropic can’t figure out how to monetize AI internally, how will their FDEs teach customers to do it?

Rising implementation costs and negative margins tell us that AI SaaS vendors are likely eating the FDE costs. The 13X increase in consumption is even more expensive than it looks. Every new agentic customer costs more to onboard than traditional SaaS customers, and they cost more to serve, but they don’t spend much more.

I want to see more earnings from other AI SaaS vendors, but I am hearing FDEs called cost centers by a large number of AI SaaS vendors. Palantir has proven that the FDE model can be profitable, so this sounds like a platform-level problem. How many platforms have the same problem?

Watch implementation and consulting margins to find out. My sense is that most AI SaaS platforms are struggling with rising implementation costs and customers who are unwilling to foot the bill. We’re seeing signs that the current AI SaaS business model is less profitable than the legacy SaaS business model.

The unit economics of AI and agentic workloads are currently negative at level 3 tech companies. Costs are scaling faster than returns for the firms (AI SaaS, Big Tech, and frontier labs) consuming the most cloud compute at this level. None of them seem to have a clear line of sight to when that will change.

Hyperscaler growth used to be concentrated in a few high-spending customers, but they have gone out of their way on earnings calls to show that spending is broadening out. Anthropic has about 250 customers spending over $10 million per year. However, that also means level 2 companies and startups are more vulnerable to a broader slowdown in spending across level 3.

The Core Business Is In Trouble

The second story speaks directly to the core business models behind level 3 monetization. ServiceNow shows that the AI SaaS core business model is under pressure. Margins are falling, which will make it harder to keep increasing spending on level 2. But what about the other tech business models?

Amazon’s Alexa for shopping use is up 570%, but online sales missed estimates. Amazon hasn’t figured out how to translate strong AI and agentic user growth into an increase in its core business. Each Alexa query that must be served by a cloud model comes at a cost, but isn’t returning incremental value back to the core business model.

Agentic commerce will require more than just bolting an agent on to an existing workflow. When Garman said customers were still figuring out how to get value, he was also talking about Amazon’s core business. I will come back to this in a minute.

Google search queries are rising at the fastest rate in 5 years, but AI Mode has much lower monetization. Search revenue grew at the slowest rate since 2023. Again, usage is up, but the margin on each unit of usage is down. Google will try to pump more ads into AI search queries to compensate, but this is an early indicator of a much larger pattern of business model destruction.

Reddit shares fell by 29% in a single day after it reported a weak sales forecast for next quarter amid slowing growth this quarter. It’s important to see Reddit’s results next to Google’s. The search engine used to be one of Reddit’s top sources of search traffic. AI search’s rising popularity has pushed referrals “down meaningfully this quarter.” That is slowing user growth.

It’s bad enough that Reddit will test requiring users to log in to access content on the platform. Search is the front door of the content economy. Google serves ads on the front door and sends people through the door to websites that can monetize those users again with ads, subscriptions, etc.

Google’s results show that AI has significantly increased usage and moderately reduced monetization at the front door. Reddit’s results show the door is closed, and AI has meaningfully reduced usage and monetization at the destination. AI isn’t growing the search and content pies. It appears to be shrinking both.

This is the reality of the AI and agentic paradigm. Businesses want agents to optimize their operating models by replacing staff, but that’s not how they work. Businesses want agents to accelerate growth by amplifying their business models, but that’s also not how they work.

The Twinkie Problem

Meta’s charging 14% more for its ads, but that wasn’t enough to offset rising token, legal, and data center spending. Meta successfully delivered higher value to customers and was able to charge a higher price. However, it wasn’t enough to keep the company from reporting its first negative free cash flow quarter ever.

Costs are scaling faster than Meta’s current business model can sustain, even with AI growth factored in. If ads keep growing (price or volume), then spending can too, just not as fast. Unfortunately, Meta has a Twinkie Problem.

Ads are in trouble because agents don’t watch ads, and agents are a growing share of internet traffic (ad volume drops). The threat to Reddit is a threat faced by nearly all social media and content companies that rely on search for traffic.

Personal agents extend the threat even further, so Meta is trying to move into a new business model. However, it’s clear that Meta’s new agentic business model will cannibalize its existing ad business model. Google and Meta have the same Twinkie problem. Agents shrink the Twinkie for ads.

Meta is still figuring out how to scale its consumer agents, and we are truly early when it comes to personal agents. We need more time to see if Meta has a viable transformation strategy from ads to agents.

Google is trying to scale AI search, but it doesn’t have a well-formed new business model. It’s experimenting with agentic commerce, which is also truly early. Again, we need more time to see if Google has a viable business model and transformation strategy.

Either way, Google and Meta are moving from a large ad ecosystem to a smaller agentic commerce ecosystem. They must build a business model that monetizes better than their old one while also competing against each other more directly than ever.

Agents Are Consolidating Forces

But no one’s talking about the much bigger shift underneath it all. This one generalizes. The content that used to be hidden behind search results is being viewed by agents and presented to users on a single surface. The agentic surface is like a firewall keeping every other UI from reaching the user.

We used to need multiple UIs: Google Search led to Reddit, Stack Overflow, Reuters, etc. AI Search compresses it down to a single action surface. A search that used to give multiple websites an opportunity to monetize the user through their UIs now offers that opportunity to only one UI. Agents do this (consolidate down to a single UI) to every workflow they touch.

Content businesses must figure out how to monetize the substance of their content after spending years training customers to expect it for free. They must compete on content as a differentiated product instead of relying on SEO to differentiate by position in search results. They must also deliver something that LLMs can’t easily replicate. LLMs are a new competitive force in the content market that have better unit economics.

It will be the same for content creators. Think of each influencer as a website. When social media agents (like Grok or Muse) summarize content from across the platform or generate it on demand, why click through to the author? Text will be summarized, photos edited, and videos will be clipped like a highlight reel.

Social media marketing will struggle, and that revenue stream will largely dry up. Creators will need to figure out how to get people to pay for their content after training them to expect it for free, and they face all the same challenges as content businesses do. Agents bring more views, but the margin per view is falling at the same time that AI becomes a new competitor with better unit economics. The Twinkie is shrinking.

UI Consolidation Goes Beyond Content

SaaS companies have a UI problem too. In the digital paradigm, much of a SaaS platform’s value was in how the UI supported the workflow. Usability and design were critical to make workflows on the platform as seamless as possible. Agents are taking over software use, and they don’t touch the UI if they can avoid it. MCP servers or legacy APIs are all agents need.

Take away the UI, and most SaaS platforms are just a database sitting behind a collection of APIs. Why use Zendesk instead of Zoho or Salesforce? The answer is part of a user demo, except the new user doesn’t want a UI tour. Agents just want the API documentation and data catalog. UI consolidation means SaaS commoditization. It’s hard for software to differentiate without a UI.

The growing range of cybersecurity threats adds urgency to that consolidation. More vendors means more vulnerabilities. It’s also a way to differentiate: “Five-nines (99.999%) secure against AI threats.” Keep an eye out for more emerging differentiators. Consolidation doesn’t mean the end of all software, just a smaller Twinkie.

Amazon benefits from UI consolidation. Its website is a single action surface where shoppers can find most of what they need. Why go to 18 different websites or shop at multiple brick-and-mortar retailers? Agents can browse hundreds of websites, find the lowest price on each item, aggregate a purchase across them into a single transaction, and make checkout a single click.

You no longer need an Amazon Prime subscription to get the same experience. Walmart and other retailers are opening their doors to agents, which is a smart way to differentiate. If your agent can’t get on Amazon, Walmart has the same items for less, and it is agent-friendly. Unfortunately, agent adoption is a transient moat. Ecommerce is just a domain-specific type of SaaS.

Agents Will Drive Growth, Just Not For Everyone

Buying direct could deliver lower prices for consumers. When the additional labor required to buy every item directly from the manufacturer drops to 0, it makes sense for manufacturers to offer DTC agent support. Watch for signs of this wave forming.

Then ask: What would have to change for the one-stop-shop advantages of Walmart, Amazon, Target, etc. to be rendered meaningless? Each retailer adds overhead even though their negotiating power enables lower prices. If producers and manufacturers can implement a DTC strategy that splits the Big Box overhead, giving consumers a better price while providing producers and manufacturers with higher margins or volumes, the Big Box advantage won’t last long.

Shipping or delivery becomes a new point of friction, but Uber or DoorDash can handle complex deliveries as their autonomous fleets come online. Watch for UPS and FedEx to start offering consolidated shipping orchestration and distributed costs for multi-retailer agentic orders. Logistics used to be a competitive advantage, but agent fleets (swarms) level the playing field in the physical and digital spaces.

UPS, FedEx, Uber, and the new DoorDash/Instacart will become agentic logistics platforms that handle it all if they’re smart. And they are. Watch for complementary retailers opening distribution centers together. Forget dark factories. Dark retailers are coming with Amazon or Walmart’s warehouse automation platform (digital and physical agents) running them.

UI consolidation opens the door for agents to take over the work of ‘middleman’ businesses. Meta is using AI to push the middleman ad agencies out of the market. Companies can spend their entire ad budget with Meta instead of splitting it between Meta and the agency. Meta’s business model allows it to give the ‘ad agency’ agent away for free.

However, ads are also a middleman. Companies buy ads to increase sales. In the agentic commerce paradigm, consumer agents like Muse or Dot will own the sale. Why buy an ad on Facebook when I can buy the sale from Muse? Why buy an ad on Google’s AI Search when you can buy the sale from Gemini?

Sales have better monetization than ads (higher prices), but there are far fewer of them. Compression will increase competition in growing ecosystems as it eliminates entire categories of business models.

The Outcomes Economy

Companies are trying to use agents to make their business more efficient by replacing people. What they don’t understand is that agents don’t optimize an operating model. They optimize the operations of entire marketplaces and ecosystems.

Everyone will soon be asking, “How can I build a business model that lets me give what you do away for free by monetizing an up or downstream outcome?” Value will aggregate to the highest value, most necessary outcomes.

The Outcomes Economy is a longer-term transformation. UI consolidation will lead to agents enabling business models that grow by pushing out middlemen businesses. Eliminating middlemen creates a shorter path between intent and outcome. That leads to the Outcomes Economy.

Today, there’s a lot of value trapped between intent and outcome. We have built massive businesses in the middle. They all add cost, but not all of them add value efficiently. What’s worse, we’re spending even more money to transform these businesses with agents. None of those workloads are economically viable, but they take up just as much compute as the value-creating workloads do.

There’s a second leg to the outcomes economy: Novel ecosystems and marketplaces. This is the longer-term and much larger growth story. I’ll write that up once there’s more publicly available evidence about it.

More technical doesn’t mean learning to build features, pipelines, or models. That’s career planning in the rearview mirror, and you’ll end up skating to where the opportunities used to be. Strategy and product management are forward-looking by nature, so we build the infrastructure of future outcomes, not the artifacts of present ones.

Strategists and product managers are a new type of builder who fill in a gap that businesses have suffered from for decades. Learning the new things they build to fill those gaps will help you skate to where the opportunities are and will be. I wrote this article and my last one to show you what that means.

My courses are built to teach a new engineering paradigm. The new builders are business architects who engineer three strategic pillars. In early maturity phases, we move parts of the business and operating models into the technology model. In later maturity phases, we build new parts of the business and operating model in the technology model. By the end of this article, you’ll have a better understanding of what that means in practice.

The new builders are technical stewards of improvement, transformation, and innovation. In the agentic paradigm, all three must be architected into the platform or agents go one way, and the business goes somewhere else. Again, we have an information problem, and one of the new technical skills is understanding information theory.

Managing the flow of information into the business is a more critical role than most realize today. As always, I’ll start with what that role looks like today and what you can do tomorrow before explaining 3-5 years from now.

Introducing A New Builder

“If every business had perfect information and a perfect decision-making process, every product would be a winner.”

Information engineers and data engineers build the pipelines that iterate the business closer to perfect information. Software engineers and AI engineers build the logic and models that iterate the business closer to the perfect decision-making process. Those jobs are taken, but they leave a massive gap.

Not every product is a winner, and the cause is rarely technical defects. For many products, everything was built as technically specified, but the outcome did not materialize. Who’s engineering the fix for that business defect?

The title is carefully worded to reveal the gap. Strategists engineer the transformation from data pipelines to information pipelines. Product managers engineer the monetization of information pipelines. The legacy paradigm is engineers building artifacts to deliver products that meet specifications without technical defects. AI and agents are encroaching further into those roles every quarter.

The new paradigm is engineers building the endgame that delivers outcomes, but endgames and outcomes are long-term goals. For tomorrow, let’s discuss building phase 1 or continuous improvement into the platform. Continuous improvement helps resolve the highest-impact business defects.

In this example, we are improving our data pipelines so they can transform into information pipelines and improving our data monetization so the business can monetize information. What you’ll realize in this article is how much every phase of maturity benefits from understanding information theory.

Surprising Facts About Information Theory

Information theory is the math of surprise. I was hooked the first time I read that. It gave the math a place and meaning before I had been introduced to a single function. The current approach to enterprise information strategy is focused on meaning, but you need to know something different about information to support agents.

Surprise comes in three flavors:

I didn’t expect that to happen.

I didn’t realize that was an option.

What do you mean it didn’t work?

Both agents and people benefit from fewer surprises, but, in a bit of irony, information benefits from surprise. It all comes down to outcomes. All three surprise categories put outcomes delivery in jeopardy. All three surprise categories contain information that agents and people don’t currently possess.

A business that is never surprised delivers consistent outcomes, but isn’t learning, so it can’t improve, transform, or innovate. A business that is always surprised has the most opportunities to learn, but is operationally chaotic. Information pipelines are the key to managing the business’s dynamism and the learning that leads to improvement, transformation, and innovation. Here’s how you start doing that tomorrow.

Read more

Agents, AI, and information thin the dependency enough for us to see and operate the business in a completely different way. But I need to take you there in pragmatic, actionable steps, not visionary leaps. You’ll be able to use what you learn in this article tomorrow while building the foundations for 3-5 years from now.

Always know what you get paid for and why they pay YOU to do it. Your career, product, and business fail when what you do disconnects from what people pay you for or if too many people can do the same thing. Economics Limerick: If you get paid a lot to do what you do, others will try to copy it so they can get paid too.

That sounds like a lyrical MBA or the start of a TEDx Talk, but it’s also an information problem. At the beginning of a technology cycle, everyone has that problem because no one knows what the opportunity is. If every business had perfect information and a perfect decision-making process, every product would be a winner.

I built a reputation for V Squared and myself around building agents (before that models) that customers will pay for. Agents are innovative and enable entirely new product categories. We have no historical data to inform our decision-making about what agents customers will pay for (what the opportunity is).

So how can you know what to build before you know what people will pay for? V Squared gets paid to build agents customers pay for. Our clients pay V Squared to do it because we have a repeatable process that has allowed us to answer that question before everyone else can. We can replicate past success even with new technologies and changing marketplaces.

In this article, I’ll explain the process I have used for almost a decade and taught to thousands of people. At its root, this is an information problem. When faced with a lack of historical data, we must turn to information for answers. V Squared gets paid to build agents, so the technical knowledge is still critical. However, clients pay V SQUARED to build agents (not another firm that can) for a completely different set of knowledge.

AI Labs Have Cleared A Barrier To Monetization

LLMs and AI labs did something huge in the AI product space. They got customers to pay for inference and to understand its value. It’s similar to what BI and reporting did for data. They took it from an afterthought to a first-class citizen in the enterprise.

I know paying for inference sounds small, but before LLMs, only a handful of us had ever gotten anyone to pay for inference as a product instead of as a supporting feature. It was a massive challenge. Many software platforms used inference, but customers paid the same amount to use the software with and without the inference. It took a lot to get them to pay more to consume inference, and most early attempts to monetize inference failed.

The AI labs have overcome a psychological barrier that stood for over a decade, but they aren’t profitable. There’s still work to be done. The AI labs have done a lot of innovating, but not enough monetizing. Agents are a big part of closing the monetization gap.

Getting customers to pay for agents is a balancing act between innovation and monetization that’s dictated by product cycles. When done well, innovation ensures future revenue growth, and monetization delivers revenue growth this year. Businesses need both to survive, but few AI builders and product managers understand the mechanics of finding and fine-tuning the company’s innovation-to-monetization ratio.

If you have taken my AI Product Management certification, this used to be called the Innovation Mix. It’s one of many sections that are being refreshed for 2027.

Agents Reveal The Product Function As An Ecosystem Of Roles

Most enterprises treat GTM as a single role or part of the product organization. The strategic decision of managing the I2M ratio is a completely different team and function. The people building agents and making critical architectural, design, and implementation decisions live in yet another organization. And innovators are just wherever.

Building the C-level discipline is the AI strategist’s job. Nothing else works without it.

Implementing the business architecture and information flywheels into the platform is the AI product manager’s job.

The AI builder and innovators must follow the signals and manage the handoff.

With agents, the product function must be an ecosystem. A business that doesn’t see the opportunity-to-revenue growth as a pipeline with multiple flywheels will always struggle. A business that doesn’t see the pipeline as one loop of an ongoing cycle can succeed once, but struggle to repeat it.

This article and the series take a pragmatic approach to explaining where agents take businesses and marketplaces. I’ll start here by explaining how each part is critical for building agents that customers pay for. I will also explain the biggest pitfalls that prevent incumbents and startups from being successful. A bad I2M ratio didn’t used to matter, but technology has changed that.

Let’s begin where I introduce most courses: The justification for why businesses must do something different. No one will change until they know why.

The Wrong Ratio: Unrealized Potential Or Slow Decline

Building agents that customers pay for requires innovation and monetization to be in balance. Too much innovation results in a lot of potential being wasted. Revenue growth is feasible, but the company watches others monetize the innovation and realize that growth. Google has often fallen into this trap, and the AI lab startups are stuck here as well, but it’s still rare.

Too much monetization is the trap most businesses fall into. Revenue growth provides a signal that everything is working, so why change anything? There’s nothing innovative in the pipeline, so the business keeps squeezing more out of its current product or platform. Eventually, growth stagnates, and there’s no potential in the pipeline to keep investors from moving their gains elsewhere. Most SaaS companies are in this phase right now. Telcos have been in a more advanced version of this phase for decades.

Too much innovation happens when the business has enough power to ignore market signals. Startups get this from VC funding. Google got it from the revenue generated by its search and ad dominance. Meta got it from social media dominance and its massive ad marketplace during its Metaverse wanderings. In Meta’s case, it wasn’t a failure to monetize as much as a failure to innovate in a direction it could monetize.

Too much monetization also happens when the business has enough power to ignore market signals. SaaS companies got this from their high margins and outperformance relative to other tech companies. Telcos got it from high barriers to entry.

When revenue is growing, there’s no reason to believe it will ever stop until ‘The Event’ happens. LLMs were Google’s event. Agents were SaaS’s event. Open-source AI was the AI labs’ event. Space (Starlink, IRIS2, etc.) is about to be the wireless carriers’ event. Agentic commerce will be ‘The Event’ for direct-to-consumer, just like eCommerce was ‘The Event’ for brick-and-mortar retail. Consumer agents will be ‘The Event’ for advertising.

Cloud was Microsoft’s event almost 20 years ago. PCs were IBM’s event before I was alive. GPUs were ‘The Event’ for CPUs, and quantum computing will be the GPU’s event. This isn’t new or exclusive to our current technology cycles, and it won’t stop until technology stops advancing.

The Technology Cycle

Every technology cycle progresses through three phases. If you have taken my AI strategy certification, this will be familiar. It’s critical to explaining the ‘why’ and getting C-level buy-in. The Product Arrow is the derivative framework I use to convince CFOs to invest in innovation, even when revenue is growing. Both frameworks are a critical part of the repeatable process.

You can see the value, even for an innovator. Businesses won’t pay an innovator if they don’t understand what’s in it for them. An innovator who can explain it gives them a reason to hire AN innovator. An innovator who can innovate in a line the business can monetize gives them a reason to hire THE innovator.

Read more

“The thing that I really give Jensen all the credit for, quite frankly, is that consistency of vision of what this can be—not today, not tomorrow, but as a secular trend.” Satya Nadella

Secular Trend? Everything Microsoft announced last week sounded premature and rushed to press release. Nothing is ready. None of it has been externally verified. Most of it comes with warnings against running anything built with it in production. My first reaction was, ‘What is Jensen Huang doing on stage announcing the AI version of a Zune alpha release?’

Microsoft’s release is technically sloppy. Huang will show up for a kid’s birthday party if it helps NVIDIA sell more, but he doesn’t put NVIDIA’s brand next to technically sloppy. He’s famous for distancing NVIDIA from crypto and NFTs, even though they were a significant revenue driver before AI took hold.

Microsoft’s world-class enterprise sales machine turned it up to an eleven in an effort to revive the Windows AI PC. They’ve been working to make AI PCs a ‘thing’ since 2024 with nothing to show for it. But don’t ignore this announcement as more of the same. I got more calls and messages about this from CIOs than anything else in recent memory.

NVIDIA is up to something much bigger.

Signs Of Something Much Bigger

This article is an inside look at how we make the bread at V Squared. Pragmatic Futurism is how we are consistently 2-3 years ahead of everyone else. The best way to predict the future so you can prescribe the right approach is to build it. But building is an ecosystem process, so you must watch what others are doing, especially when dozens of C-level leaders suddenly ask questions and the largest industry players all turn in the same direction.

This weekend, we had an all-hands to figure out where all the interest was coming from. What we found was an industry story rather than a Microsoft story. It’s more about where NVIDIA is guiding the market and the market shifts that are forcing its hand. Microsoft got dragged on stage by the only gorilla big enough to pull that off. It feels like NVIDIA is trying to force Microsoft to accelerate its timelines and increase the number of people it has working on this.

It doesn’t make sense until you see the big picture.

I’ll start by defining a ton of terms that are suddenly must-have vocabulary. Then I will explain what Microsoft and NVIDIA are up to. If Jensen Huang can pull this off, it will be one of the largest opportunities Microsoft has ever taken advantage of. Huang is well-known for making his friends rich, and he’s trying to do it again.

“If you look at the entire journey of our company, Windows was at the core of it. If not for Windows, there would be no GeForce.” Jensen Huang

We dissected the announcement in light of NVIDIA’s recent deal spree. At the beginning of this year, NVIDIA was building the stack to enable partners like OpenAI and Anthropic to train models and serve inference as efficiently as possible. Groq, Corning, IREN, and Nebius tell the story of a focus on optimizing the data center buildout for AI.

HuggingFace, Reflection AI, Kumo, Poolside, and illumex tell the story of a new chapter. Microsoft‘s AI PC is a bridge from where we are to where we need to be for that new chapter to take hold. Jensen Huang believes this will be “the foundation of the next generation of IT as we know it. Just as Windows and DirectX revolutionized how applications were built, MXC is going to revolutionize how agents are built and deployed.”

But this isn’t the end state. There’s a bigger vision that several leaders in the Mag 7 share.

Agentic Architecture Vocabulary Lesson: Hardware

Unified Memory is a single pool of shared memory that the CPU and GPU can access. Apple Silicon has been extremely successful with this as well. It avoids the need to copy data between separate CPU and GPU pools. The GPU can leverage most of this memory to store model weights rather than being limited to its VRAM, which is typically much smaller.

Coherent Memory is physically separate memory for each processor that is linked so either processor can read or write to the other’s. The coherence allows both processors to always see the current version of the other’s data. However, the memory operates at different speeds. GPUs use Micron’s or Samsung’s fifth-generation High Bandwidth Memory (HBM) running at up to 7.1 TB/s. CPUs use more standard DDR RAM that runs much slower. A model that fits into GPU memory generates tokens much faster than if it needs to leverage CPU memory.

NVLink-C2C connects the CPU and GPU and manages how fast the processors can exchange data. NVIDIA’s connection is about five times faster than what most Windows PCs are using today, so it makes memory sharing and coherence viable.

Tensor Cores are the GPU workhorses that run matrix multiplication and enable models to train or serve inference. The RTX Spark uses Blackwell chips that provide native support for FP4 (4-bit floating-point operations), which optimizes the amount of memory that the model needs. At FP4, a billion parameters take up 0.5 GB of memory. At FP16 (16-bit floating-point operations), that same billion parameters require 2 GB of memory. For local AI, this is a really big deal.

Key-Value Cache, or KV Cache, is a type of memory optimization for transformer inference serving. Transformers like LLMs generate a token at a time by comparing a query vector against every prior token’s key-value pair. The cache means the transformer doesn’t need to recalculate the key-value tensors at every step. This reduces computational complexity from quadratic to linear, which means inference is served faster or tokens are generated faster. KV Cache and quantization (using FP4 vs. FP16) allow local AI to support longer context windows and more complex tasks without running out of memory.

Agentic Architecture Vocabulary Lesson: Frameworks & OS

Microsoft is building a top-level framework called Hybrid Intelligence that relies on all those hardware concepts to be viable. This is where the alignment between Microsoft and NVIDIA becomes clearer, even though I will also explain the split on the horizon. The hybrid refers to the combination of local and cloud resource usage.

Microsoft announced two new local AI models (built to run on AI PCs) called Aion Instruct and Aion Plan. Instruct handles common language tasks like summarization, writing, and editing. More importantly for Hybrid Intelligence, Instruct also interprets intent. Aion Plan supports local tool-calling, context extraction, and workflow planning.

“What happens in the era of agents, when the agent is on your computer? It {AI PC} is now your personal assistant.” Jensen Huang

Together, Aion Instruct and Plan running on NVIDIA’s local AI stack create a single action surface for any workflow to start from called WinUI 3. Hybrid Intelligence evaluates the workflow using Intelligent Routing to determine which tasks will be done in the local environment and which tasks can be escalated to run on cloud infrastructure or cloud-hosted models.

GitHub already runs a version of Intelligent Routing called HydraFusion. Microsoft is adding the ability to select local models to that framework. It is also building support for Intelligent Routing to select from local models or for the user to explicitly select a local model to handle workflows beyond just coding.

Microsoft is opening up and extending the ecosystem. It supports local models built in ONNX and GGUF. Both are formats for representing a model so it can run on any environment that supports the standard. Microsoft is also implementing llama.cpp and an OpenAI-compatible endpoint. Those enable a local model to be accessed via an API. Add all those together, and you have the ingredients for an open model ecosystem to run more than just Microsoft’s Aion models.

Microsoft is also implementing an MCP server and agent connector running locally. Windows will include an on-device agent registry (ODR) that supports discovering and managing registered agent connectors (integration package letting an agent discover and use an MCP server’s capabilities). Windows will expose file explorer and system settings connectors. It is also getting an AI Dev Gallery with sample applications to encourage others to extend or build on top of the agentic harness that Windows is implementing.

Microsoft is also implementing a local version of its AI Foundry for Windows, with all its AI development offerings to be included, called Foundry Local. WindowsML is the local inference framework for model deployment. It includes APIs for things like speech recognition and text generation and supports more complex task pipelines.

Microsoft Execution Containers (MXC) is the local governance and security layer. It supports isolated execution, policy management, identity management, hardware provisioning, session management, resource access, and observability. There are over a dozen new frameworks, and some existing ones have been thrown into the mix. MXC is the sandbox and guardrails Microsoft is counting on to make agents trustworthy for consumers and enterprises.

The Balance Of Power Changed

You need the wide shot of this image to really get a sense of what shifted and why NVIDIA is moving in a new direction. The symbolism is thick with indications of who’s in and who’s out. You’d need a wide-angle lens to see where Dario Amodei and Greg Brockman were sitting.

Read more

McKinsey came out first and found that agents only capture the full value by “rearchitecting the entire task flow from the ground up.” In 2026, they doubled down with “The biggest gains come from reimagining end-to-end workflows around what the technology makes possible.”

BCG discovered “A common mistake is automating what already exists.”

MIT Sloan agrees, “The greatest gains come when leaders rethink work from first principles.”

Microsoft says that “leaders who redesign processes around outcomes and agent autonomy” realize the greatest gains.

I was right about this two years ago, which means I was very early to building agents to support workflows and workflow reorchestration. That head start and experience have shown me that reliable agents look nothing like the software that we have been using for over 30 years. My new favorite interview question is “What’s the difference between data and information?”

The way people answer is a good indicator of how well they understand agentic architecture. In this article, I won’t just tell you the difference between data and information. I’ll show it to you and explain why knowing the difference is critical to architecting agentic platforms.

Why Data Was Enough For Software

Software supports tasks. The purpose of a workflow is to deliver an outcome, so agents are fundamentally different. More complex requirements lead to more complex inputs.

Software, AI, and agents are all data consumers, but for AI and agents, data isn’t enough to achieve their purpose. Software’s logic or business rules enforce a structure on the data it consumes, and logic provides the benefit of context. Software often turns data into information so it can complete a task or support a user.

I query a table and get back a data point of ‘3299.99’. Software has metadata in its logic, like ‘$’. That allows an app to display the value with the correct context to a user comparing products side by side. The user can complete their comparison task because the UI’s layout makes their job a bit easier.

The task approach is simpler for the technology, but harder on the user. I had to perform multiple search, filter, and selection steps to get here. It took a couple of minutes. After working with agentic commerce and search platforms, this workflow is extremely annoying.

I often quote William Kent’s book, ‘Data and Reality’. “Data structures are artificial formalisms. They differ from information in the same sense that grammars don’t describe the language we really use, and formal logical systems don’t describe the way we think.”

This presentation captures what Best Buy thinks is the best way to present data to me so I can make a buying decision. However, the logical layout does not capture the way we decide which laptop to buy. The data points hint at decision structure, but they do not formalize it.

Again, from Kent’s ‘Data and Reality’: “Data models are tools. They do not contain in themselves the true structure of information.”

Data models in the Joe Reis sense are built in whatever structure is most efficient for software’s needs. Data models in the Best Buy website sense are built based on user feedback about their needs.

Why Agents Need Information

For AI and agents, data alone isn’t enough because they are responsible for the outcome, not just the layout and presentation to support a task along the way to it. Peter F. Drucker wrote in ‘The Coming of the New Organization’: “Information is data endowed with relevance and purpose.” Agents need a workflow’s context and mechanics to achieve their purpose.

Read more

You will see it reflected in this newsletter and in my courses. I am adding a new section this weekend to my ‘Executive Presence & C-Level Influence’ course that covers manipulative dark patterns and how to negate them. Influence is a core capability that we must master to advance in our careers, but there are two types of influencers: Persuaders and Manipulators.

I teach persuasion or ethical influence in my course. Persuaders use narrative design and decision-making frameworks to improve collaborative decision outcomes. They align the information in the narrative with a map of the decider’s mind (decision frameworks they use, information preferences, etc.) to create the ideal conditions for decision-making. Persuaders wield influence ethically by filling in the gaps that exist in every decision-making framework.

I am adding manipulation to my course because not everyone chooses ethical influence. Manipulators wield hidden influence to control decisions and improve their own outcomes. They align the information in the narrative with the gaps in the decider’s mind map to make the outcome they want happen. My course is getting two new sections on detecting, understanding, and negating manipulation.

Influence is a valuable skill as you advance in your career, so let’s start with the difference between persuasion and manipulation.

Is It Persuasion Or Manipulation?

I work with C-level leaders and the boards at some of the largest companies in the world. For over a decade, I have had a front-row seat to watch the best persuaders and manipulators at work. I must negate manipulators on a weekly basis to do my job effectively, so I have developed an expertise in the subject, which I now teach in seminars and courses.

Is that paragraph persuasion or manipulation? One of the clearest manipulation signals is transparency. If I explain the narrative design and information spine in the last paragraph, will it still work? Will you still be persuaded?

The motivation is transparent, even though I didn’t state it explicitly. I told you about my course first, its content second, so this paragraph becomes a natural third element: credibility. I need to explain where my expertise comes from.

I am teaching a course that doesn’t focus on my obvious area of expertise, so I must establish the source of my credibility. My objective is transparent due to the ordering of my narrative, but is the claim verifiable?

You can verify that I have worked with some of the largest companies in the world as a thought leader. My clients are tougher because of my NDAs. I have been at public events with the CEOs of Salesforce, SAP, IBM, Juniper Networks, and several others. I write about these events in this newsletter, and you can read how well I cut through manipulative vs. persuasive marketing. The paragraph has basic evidentiary support.

I am persuading you to take a course by filling in an information gap, so you have the best setup to decide if the course will deliver your desired outcome. There’s nothing unethical about selling, marketing, or promoting as long as you maintain transparency.

Influence Is A Critical Career Capability

That’s an important career lesson for all of us who come from a technical background. We’re often told that those three activities are always slimy or unethical, but they are critical parts of career success. Understanding the difference between persuasion and manipulation is what lets us avoid the slime, so our careers advance on sustainable foundations.

Manipulation isn’t sustainable, even though it often looks like the easiest or most effective path. My approach to ethical selling comes from studying one of the most unethical salespeople of all time, Jordan Belfort, or the Wolf of Wall Street. He ran multiple illegal stock trading schemes and went to jail for them. He is the origin of the popular sales interview question, “Sell me this pen.”

After he was released, he went on the speaking circuit and in one interview, he was asked to “sell me this pen.” He explained that the best path to selling anything was understanding the customer’s needs. Belfort started with questions like, “How long have you been in the market for pens?” and “Do you usually buy cheap or expensive pens?”

He moved on to questions that would discover the need, like, “Is this for signing contracts, journaling, or something else?”

Only then would he close the sale. He said that if the person wasn’t in the market for a pen, the right next step was to move on until you find someone who is. The course title, Executive Presence & C-Level Influence, is built to help people self-select. Do you have either need: executive presence or C-level influence? If you do, this is worth looking at, and here’s the information you need to make a good decision.

People who can’t persuade must depend on others who can. If I couldn’t persuade people to take my courses, that business would either fail or it would depend on someone else’s work to succeed. Being a builder isn’t enough, no matter what type of builder you are.

If you cannot persuade executive leaders to listen to you, then you must depend on someone else to get them to follow your technical guidance. You create the value, yet you are beholden to someone else for that value to create the impact that is rewarded. If you cannot persuade them to promote you or give you a raise, then you must depend on someone else to get rewarded for the value that you create.

Manipulation Is Always The Wrong Approach To Influence

Influence is critical, and only a manipulator will tell you that all influence is slimy. They don’t want you to learn what they don’t understand. That’s one example of hidden influence and manipulation. If you become a competent persuader, they become dependent upon you.

Manipulators don’t realize that’s how teams work. Everyone depends on someone else to deliver as a team rather than a collection of individuals. The leader is the most dependent upon everyone else, and effective leaders accept that dynamic. Persuasive leaders grow the team’s capabilities so they deliver more. Persuasive leaders create a safe team environment so people can depend upon each other.

Manipulators practice the least valuable form of leadership. They limit the value their team can create and build an environment where trust is difficult. A team without trust is a collection of individuals with the same manager.

Manipulation is the hardest road. Persuasion is the easiest. But you only learn that after seeing manipulation’s limits. I study real-world villains because they expose the hidden rules we all play by.

In every company, there are two types of hidden influence: manipulators and their enablers. Enablers use persuasion first, but they acknowledge the need for the short-term gains that manipulation delivers. They won’t degrade their capabilities by doing the manipulating themselves, so they enable manipulators to do it for them. That’s the second type of hidden influence.

The Manipulative Leader Builds On Unethical Foundations

If you are working for a manipulative leader or with a group of them, you are not alone. The sad reality is that manipulative leaders get ahead because their tactics are effective, and there are very few effective strategies for negating them. Hidden influence has practical advantages.

It takes strong C-level leadership to keep manipulators from getting ahead. This type of leader should not be as prevalent as they are. Their resurgence over the last decade is enabled by their ability to survive reorgs and layoff cycles. Since COVID, they have manipulated their way through the churn and advanced higher than they should.

‘Enabler’ leaders always keep a couple of manipulators around, but the more they use manipulators, the more other leaders see that path delivering success. That creates more manipulative leaders. It never ends well, which is why effective C-Suites keep manipulative leaders from advancing.

Manipulators project strength, but leaders with a track record of success look down upon them. We openly call them out and exclude them. A few take that as corrective feedback, but most manipulative leaders let it feed their insecurities. Their behaviors get worse, not better, which is why high-performing C-Suites actively weed these people out once they are exposed.

I’ll explain the easiest way to detect manipulative leaders. Some have gotten skilled at avoiding the C-Suite’s notice. Detection allows you to expose them to an effective C-Suite or avoid them when that’s not an option.

Signs of Manipulators

Read more

If we keep thinking in the same way, we will get the same results. It’s equally true for careers as it is for business. A digital mindset is limiting for agents and AI. Both cost more than digital technology, so we can’t apply them to what digital already does. We must deploy our new mindset to see new opportunities and business models.

A big part of my job is helping C-level leaders develop that new mindset AND rapidly deploy it to generate value. The dual mandate is also a function of this new business era. For over 30 years, the framework that developed the mindset was the product. Today, I only get paid if I provide the top-level framework AND the supporting frameworks that enable people to deliver better outcomes as soon as they leave the seminar or workshop.

It’s a lot like engineering workshops that are equal parts theory and hands-on lab. Executives, VCs, and C-level leaders want 360 capabilities, not ‘knowledge of’. The purpose of this exercise is to help C-level leaders learn, synthesize, and apply the framework to build their new mindset AND the supporting frameworks to put it into practice.

Why Reinvent A Calendar?

I am devastated to be letting go of the calendar exercise because it was too perfect. Sadly, technology is advancing quickly, so I need a more complex exercise next year. Calendars are perfect because we have had them for between 10,000 and 13,000 years. Everyone understands the calendar because it’s a core innovation of all human civilizations.

Agriculture created the need to do specific activities at precise times of the year. We have extended the calendar into modern times to support nearly every industry. Most of us are used to the digital calendar, which is a close cousin of the paper calendar. But those are relatively modern constructs compared to ancient calendars like Stonehenge or what’s built into multiple Mayan Temples.

Calendars used to be the domain of top scholars because telling time was dependent on the sun, moon, or any number of celestial bodies. Then someone figured out that a spring could be coiled just right to tick off seconds, minutes, hours, and days. Putting that in a new machine took a core function out of the calendar. It didn’t need to tell time anymore, so we could simplify it to the structure we all know today. Simplifying the calendar made one of its first principles more obvious, and that’s where reinvention starts.

Read more

But if that’s true, why aren’t we seeing viral free adoption turn into people lining up to pay for products like the iPhone and the internet? Something’s missing, and no one gets paid until companies figure it out.

Meta’s Adoption Struggles

Meta’s learning this lesson from Muse’s adoption and usage. It was at the top of both app stores’ download rankings, and Meta has all the distribution you could ever want, but according to a CNBC poll:

“People use Muse and get really excited with the first 2 or 3 things they do with it and then don’t know what else to do with it and trail off.”

Distribution isn’t the bottleneck, and I called this out in my Muse explainer. Meta has grand ambitions for Muse, and the architecture proves it. However, as CNBC’s poll shows, there’s a massive gap. Customers see the cool factor, but utility isn’t as obvious.

Is it the model capabilities? Meta built Spark to handle complex workflows. Is that the wrong direction? Let’s look at a frontier lab.

OpenAI Is Having A Similar Problem

OpenAI is seeing it in GPT-6 usage and adoption patterns. Boris Power, OpenAI’s Head of Applied Research, says the main challenge is users not knowing what to use AI for. The source (applied research vs. marketing) is interesting because Power is tasked with turning model capability improvements into ‘the AI or agentic killer app’, but it doesn’t sound like that’s his team’s focus.

Power also revealed that 80%-90% of their research is focused on GPT-7 and beyond. OpenAI knows about the problem, but isn’t allocating resources to solving it. The startup just paused work on its next-generation models to focus on alignment and reliability. Still, it doesn’t sound like the focus has shifted to utility.

Maybe the problem is a lack of enterprise acumen. Startups are notorious for being disconnected from the way enterprises operate. That creates gaps in their products and platforms. What if we look at an incumbent with better enterprise acumen than Meta or OpenAI?

Microsoft’s Copilot Rebuild

Microsoft rebuilt Copilot around something Power also sees. He said GPT-6’s capabilities make it more like a coworker that you hand a goal to than the legacy chatbot paradigm. It is built to enable the proactive agentic paradigm as well. Copilot Autopilot is designed around the same capability.

Autopilot is a proactive agent, and it’s built to be a teammate, not a chatbot. Users give it a role and goal (intent and outcome). It keeps working until it achieves the goal. It has access to Microsoft’s enterprise context tool so it understands how work gets done in the enterprise. Microsoft is running a consumption-based pricing model, so more use cases = more usage = more revenue.

But the gap is the same, and it threatens the business model. What should nontechnical users do with Autopilot? It’s an easy question for engineers, and a lot of what AI labs and incumbents have successfully deployed is built around software engineering. Tech workers have an obvious agentic killer app.

But even Microsoft, with all its enterprise acumen, hasn’t found the AI or agentic killer app for nontechnical users. That means we will see the Meta trend play out again. Early adoption and usage spikes followed by a rapid plateau and decline. Until Microsoft defines the AI or agentic killer app beyond coding, the new business model and growth engine will sputter.

Meta understands consumers better than almost any other company in the world, and they don’t have a clear line of sight to the consumer AI or agentic killer app. OpenAI understands the technology and what people are doing with AI better than most companies, and they don’t have a clear vision for the enterprise or consumer AI or agentic killer app.

Who Solves This Problem?

Big Tech doesn’t have a handle on it. Startups aren’t interested in it. The big consulting companies are better known for being fast followers and survey chasers, not technical innovators. Enterprises need to solve this problem so they can turn their AI investments profitable. Where is the line of vendors and consultants who fill the gap?

I saw this on X over the weekend, but it’s one of many similar takes.

Qu hints at an argument that you have heard here for over 2 years. Unique datasets are the competitive advantage for AI, especially frontier AI. Engineering access to data-generating systems is the critical capability. The more people use a platform or product, the more data-generating workflows it has access to.

Better product -> More usage -> More access to workflows -> More unique data -> More capable model or agent -> Better product

That’s how information flywheels work. Once it starts spinning, it’s tough for competitors to keep up. It looks simple when you see it laid out like this, but there is a lot of complexity hiding behind this flywheel description. A critical AI product management discipline is figuring out how to engineer these information flywheels so they drive model and agent capabilities that increase usage.

Very few product managers have that capability, and this gap seems to be universal. I agree with Qu, but I think the dollar value is more for clicks than reality. Great product leaders are about to get paid, but AI product managers look nothing like their digital counterparts.

Novel Builder Types

Qu and many others leave out the hard part. Engineering information flywheels requires a lot of supporting frameworks. AI product managers must understand how to connect the dots upfront. What new features engineer access to systems/processes/workflows that generate unique data that trains models or informs agents to give them new capabilities that drive more usage?

Builder is an emerging role, but there are multiple types of builders.

Software builders engineer features into the product or platform.

AI builders engineer capabilities into agents.

Knowledge builders engineer information models and systems for anything or anyone to use.

AI product managers are also becoming builders who engineer and architect competitive advantages and monetization into products and platforms. Software, AI, and knowledge builders are all supposed to have enough product sense to know what to build. I have seen the builder concept deployed in real-world roles, and the gap becomes obvious. The result is what you’re seeing with Meta, Microsoft, OpenAI, and many others.

AI strategists are another new kind of builder who engineers and architects the business and operating models into the technology model and platforms. It makes sense that as technical roles move closer to business value, business value roles will move closer to technology. They won’t merge into a single role. Each must find ways to redefine the value it creates based on the new business paradigm.

Architecting & Engineering Frameworks

“Here’s this amazing thing that can do amazing things for you!” isn’t enough for customers or internal users. Someone must define the killer apps that have obvious value to drive adoption. Someone must define the new business and operating models that create more value than traditional ones do. However, neither of those things can be a simple whiteboard exercise or deck-led presentation.

Strategists and product managers must be able to build business architecture into products and platforms. They must be able to rearchitect the business through its platforms and products. They must even be able to architect new parts of the business for the first time in its platforms and products.

Strategists and product managers are becoming meta-builders. They build frameworks into the product and platform that enable the builders at the next level to do their jobs. Without the strategist builder, the product management builder cannot implement monetization or competitive advantages. Without the product manager, AI builders cannot implement new capabilities. Why? Without information flywheels, they lack the raw materials to do their jobs.

If you have taken one of my courses, you know what I have spent the last 7 years preparing certification holders for. The purpose behind my Systems, Models, and Frameworks curriculum design methodology is to prepare you for this moment. In the past, it was only feasible to implement frameworks in the rim (Core-Rim) of the business or the part that was managed by people. Agents make it critical to implement those frameworks in the Core as well.

Those are the newest sections in my instructor-led AI Product Management and Data and AI Strategist certifications. I am teaching them explicitly for the first time this weekend and next week with my last cohorts of the year. Those are sold out, but I have opened up the first cohorts of 2027 for early enrollment. They will sell out too, so act fast to reserve your spot.

I will also be updating the self-paced version of those courses next month. The digital business is a legacy construct. The next generation of business leaders, engineers, and strategists need a new mindset to succeed. Shift your thinking to open up new opportunities. Holding on to the old paradigms will continue to deliver lower and lower returns.

Over 900,000 people downloaded Muse in the first week. It is on top of both Apple’s and Google’s app stores, so demand is high relative to everything else out there. The last version of Meta’s AI app reached almost 800,000 downloads in the same period after launch, but it fizzled.

Does 17% higher initial interest mean this time will be different? Meta says that its AI app family has surpassed 1.5 billion monthly active users. Another 900,000 seems like a meaningless signal on that scale.

But the markets sure seem to think it’s a positive sign. Meta’s stock jumped 11% on Monday. It took semiconductors with it on news that demand for Muse was a leading indicator of demand for CPUs. There must be something here, so in this article I will explain the Muse Spark ecosystem: what it is, what consumers are doing with it, and whether this is hype or something worth following.

What The Muse Ecosystem Does

Let’s start with the stack because the architecture is the foundation for something much bigger than a single assistant. It describes where Meta is headed.

Muse Spark 1.3 is the reasoning and model layer that Meta released earlier this month. It supports multimodal reasoning, coding, tool calling, long context windows, and subagents. Its defining capability is being trained for long-chain, multi-step workflows rather than an action that creates a single generation like a chat interface.

This means Meta is building agent-first, not chatbot first. Spark is architected for tool calling and multi-agent orchestration. It was trained to use personal devices like a human user so it can accomplish tasks in apps and interfaces that haven’t been updated with agent support. Spark was built to support agents that act.

Meta positions Spark 1.3 as a step toward personal superintelligence or putting capable AI and agents in consumers’ hands to help everyone achieve their goals. Even if Meta or Zuckerberg give you the ick, the platform’s aspirations toward scaling are made clear in the personal superintelligence manifesto.

Meta’s Model AI and Muse Code provide developers with access to Spark so they can build their own coding agents, custom apps, and even agentic frameworks. This component is architected to support a developer ecosystem with as many partners as possible. It follows the Apple approach where the platform brings scale, and the frameworks simplify integration with the platform, so partners can access that scale.

Muse is Meta’s consumer-facing personal agent that runs on Spark. It enables consumers to automate tasks like email, travel planning, shopping, finance, health/fitness, and browser automations. This is the action surface that needs the developer ecosystem to give it horizontal breadth.

Muse Connectors are permissioned connections between Muse and other services. They support Gmail, Plaid, OpenTable, Spotify, Notion, Ticketmaster, and many other integrations. Although some third parties have pushed back on unsanctioned integrations, I’ll cover this in a minute. Meta is trying to turn competitors into partners with mixed results so far.

Muse browser enables Muse to navigate websites, fill out forms, get customer service, and access shopping sites. They are following the trend of embedding browsers into the agent’s surface that other AI labs started. It supports a workaround for companies that choose not to partner or can’t build agent support into their existing platforms.

Muse: Entry Point Into A Super App?

I cover Meta’s ad ecosystem in my AI and Agentic Platform Monetization course. That platform proves that the company is capable of monetizing AI and agents. They reorchestrated the advertising workflow from intent to outcome.

Meta has much larger ambitions for its ad platform. Zuckerberg has talked about enabling new business categories that run almost entirely on Meta’s ecosystem. The ad platform has proven itself through multiple quarters of revenue growth. The business ecosystem is still under development, but the direction is promising.

Meta is a proven operator in the AI and agentic space, so we have to take Muse seriously. It is the only reason this is worth covering this early on. Companies that have established themselves in this domain are few and far between. They have earned the benefit of the doubt.

This makes for an excellent hype filter that I explained in a prior article about measuring AI products that actually create value. The 900,000 app downloads and top spot on the app stores aren’t value signals, but Meta has shown that it can turn adoption into revenue. It is worthy of your time.

The big questions are:

• Does the business architecture align with agentic architecture?

• Is Meta’s path to super app viable, or will this fizzle like the last version?

• And most importantly, do they have a path to profitability?

Muse’s Information Model Spills The Tea

Read more

Everything about Jev is a play on words. Jev is short for Jevons and references Jevons’ paradox: usage climbs as cost drops. The marketing is clever, and it is working. 13% of Vercel’s paid AI Gateway teams used Jev within the first 24 hours of its release. It has really captured builders’ imaginations, and there is obvious utility.

But I haven’t seen anyone explain the bigger picture. The implementations that are going viral only scratch the surface of what Jev was built to do. TypeSafe AI built Jev to showcase their System One model capabilities. They are a category of models that expose simple-seeming primitives.

How It Works: The Basics

They take a state and questions about the state as input. It predicts the answers. They call it “a classifier attached to a transformer.” Said simply, you supply the information and define the possible answers. Jev’s model returns a choice, score, and/or probability set that software or agents can act on. In my Harness, Loop, and Graph series, I explained why letting the LLM make open-ended decisions results in low reliability. Jev is a clean way to prevent that from happening, making it a valuable addition to early-maturity agentic harnesses.

In my framework’s early maturity phases, I rely on lightweight taxonomies to support decision-making. Jev uses its System One models to do the same things. Both approaches require some information, but handle the reality that you often don’t get enough information for an agent to act reliably with LLMs alone. The special thing about Jev is how well its classifier generalizes to a wide range of decision domains.

Jev supports three question types: multiple choice that returns probabilities and confidence for each choice (RIP your multiple-choice tests, professor), score that rates answers against a rubric, and noul (binary or yes/no). You can evaluate multiple questions against the same state or initial question.

It’s being applied to more than simple decision-making during a workflow. LangChain tested it as a judge for agent evals and found that it performed extremely well. The speed, low-cost, and reliability mean that teams can evaluate agents more completely than before. That alone is a massive value proposition, but there’s a lot more potential.

The pattern has legs, and Jev demonstrates a solid approach to multiple agentic challenges. This is where I don’t see enough coverage, so in this article, I will get deeper into why what you’re seeing in early applications isn’t all that TypeSafe AI is building for.

Many Are Conflating Multiple Agent Categories

TypeSafe AI’s founder, Diogo Almeida, left OpenAI to build in a direction that OpenAI wasn’t interested in at the time. He researched and engineered methods behind instruction-following models. His thesis is that automating long-chain workflows requires models that are architected to return decisions that a downstream software or agentic platform can act on.

TypeSafe AI’s System One models are a solution to multiple problems, and you need to understand how different agent categories can use Jev. Recall from my AI & Agentic Platform Monetization course (smooth step 6, Cici) that there are four major agent categories:

Read more

This article explains the implications of the changes I called out in March that we are seeing the leading indicators of right now. No one knows the future, but we know enough to make actionable predictions. It’s frustrating to see really smart people responding to our current environment by throwing their hands in the air. No one has a perfect picture of the future, but that isn’t an excuse for inaction or indecision.

This mentality has taken over the C-Suite. It started with a legitimate backlash to influencers, analysts, and futurists who overhyped or doomed everything and turned out to be completely wrong about most of it. TrustRadius’s 2026 B2B Buying Disconnect Report found that only 13% (down 63% since 2022) of buyers are using analyst reports anymore, and no one listed influencers or futurists as credible signals.

Analysts lost trust due to their “pay-to-play” research models where vendors and startups funded the analysts doing the write-ups. Futurists and influencers chased sensational headlines and content that drove engagement. I get it. Everyone’s burnt out. We’re all suffering from information overload, heightened volatility, and decision fatigue.

But the loss of trust has turned into apathy, which is dangerous. I will explain the impact of three likely disruptions:

1. Micro-Unicorns (20-50 Employees, $1B - $10B In ARR, And 80%+ Margins)

2. Overweight Incumbents

3. An Apathetic Leadership Class

We’ve been living through disruption and uncertainty since COVID. I get it. We got used to the stability and clarity of the 2010s, but complexity and rapid change are our new market normal. While these are the most likely market drivers, it’s more important to focus on the leading indicators than trying to predict the causal factors perfectly.

The Future Is Complex…Systems

In this article, I will use tenets of complex systems theory to explain what to look for in the enterprise. I used one tenet in March, but now that we have more information, we can talk about a more granular view.

Hysteresis starts when the path forward is fundamentally different from the path back. Agents have undeniably created the potential for this condition. But are you seeing your enterprise’s path forward being fundamentally different from its path over the last 5-10+ years?

A start doesn’t always cause the next sign of disruption: critical slowing down. That happens when enterprises take much longer to recover from small shocks or changes. The return to stability lengthens, and what it takes to create instability in the enterprise gets smaller. This is caused by a breakdown in the systems and cycles that enforced stability in the past. Are you seeing this in your enterprise?

Most C-level leaders I talk to say they see a fundamentally different path forward than back. Only 1 in 5 see signs of critical slowing down, and I see that split as a huge red flag. The mechanisms that enforce the old equilibrium can also hold a business back from making a critical transformation while industry peers forge ahead.

The bigger question is, are the systems that create stability in the enterprise responding to the market or responding to leadership’s desire to hold on to the old equilibrium? “We can’t know, so the best thing to do is what we’ve always done.”

When the old model is no longer competitive, holding on to it makes the business noncompetitive. Getting to a tipping point is two-sided. The external market forces must be strong enough, and the disruption must be real. A lack of clarity about both factors has clouded our view of the future.

The Last Generation Of Leadership Is Struggling

Periods of instability (hard times, rapid change, and high uncertainty) create capable leaders. Those leaders create stability by finding a new market equilibrium (good times). Periods of extended stability (2010-2019) create less resilient, less adaptive leaders. Those leaders respond to instability by trying to force the system to return to the old equilibrium or take the path back to what they understand.

If agents and AI aren’t causing a real disruption, that instinct will turn out to be correct. The impacts of AI and agents have touched nearly every part of the economy, but if returns don’t materialize, then impact will not be the same as disruption.

Only a few business leaders are holding on to the old market equilibrium by ignoring agents. However, most are working to bend agents back to the old equilibrium, or charting a path back to their legacy business and operating models. Few business leaders are working to adapt their firms to thrive in a new market equilibrium with a modernized business and operating model. And legacy SaaS vendors are happy to enable them.

There are notable exceptions like Eli Lilly, Siemens, JPMC, and several other incumbents, but they are in the minority. Vendors from AWS to Salesforce are talking about the new paradigm. Are you seeing products that support a new paradigm, or agent washing that enables the old one?

From what I have seen, most are bolting AI onto what’s already in place vs. truly modernizing. Most vendors are legacy enablers, not partners in modernization. What are you seeing?

Modernization Takes Leadership, Vision, & Conviction

Modernization takes more than just technology, and this is where leadership, in my view, is failing. Most business leaders believe the old business or market equilibrium will return soon. They cannot imagine a time when market instability creates a big enough disruption for their business to face an existential threat.

Most leaders have prepared for the technology of AI and agents, but they do not realize that preparation for technology is not enough. Titans do not fall because of a lack of preparation or a lack of access to technology. They fail due to a lack of imagination.

As a result, they prepare for disruption by buying technology, but they do not transform to prepare for the disruptions that technology drives. That was the case with Walmart and Sears as Amazon brought e-commerce’s disruption to the market, and what happened to Microsoft when Amazon later brought the cloud’s disruption to the market.

Microsoft had a cloud, and Walmart had a website.

Both bought the technology but would not make the business transformation that allowed the technology to achieve its potential. The market forced both companies’ hands. If not for Amazon, Microsoft and Walmart’s resistance to change would have paid off. It’s not enough for a technology to be disruptive. Someone must apply it in the marketplace, or the disruption will never happen.

The Fortune 500 companies that have fallen to technology in the past did so because they could not imagine a time when the advantages that made them titans became the weaknesses that made them obsolete. By the time that becomes obvious, the business has resisted its market past the tipping point. Technology-driven disruptions often make the old equilibrium impossible to return to.

Tipping Points In Complex Systems: What To Look For

Every stable system has a natural tendency to return to its equilibrium after a shock. COVID is a great example. It was a massive shock that took a couple of years to work its way through the system, but marketplaces took the old paths back. The path forward was not fundamentally different once COVID’s impacts subsided.

After DeepSeek disrupted the AI market, Big Tech struggled to find its footing. Every open-model announcement sent shares down and brought back the fears of an AI bubble. OpenClaw looked like a threat to frontier AI labs until OpenAI acquired it, Anthropic copied it, and Chinese AI labs rapidly integrated it. Small movements are causing outsized reactions. But we’re still finding ways to stabilize, even with AI-driven disruptions.

The SaaSpocalypse is doing the same to legacy tech companies. Every time there is another announcement of an AI player entering their domains, SaaS stock prices plunge yet again. Investors are fleeing from that space because they do not know when the SaaS sector will find a new equilibrium or what that will look like. Investors are signaling strong conviction that the path forward is different than the path back.

Dreamforce put new wind into Salesforce’s sails, but its shares are trading in the same range they have since 2020. ServiceNow and Intuit are in the same boat. Investors are saying that those companies haven’t grown in value for over 5 years, and they don’t believe they will reaccelerate growth any time soon. That is the definition of stagnation or equilibrium, even before AI.

I can’t blame anyone who has strong conviction in the old equilibrium. It has been remarkably sticky in the enterprise and across markets. However, what if this isn’t strong conviction, but the signs of a lack of conviction?

A Lack Of Conviction In The Enterprise

The US Fed Chair, Jerome Powell, said something telling earlier this year in response to the uncertainty the Fed was (and still is) facing on oil prices, but it applies to uncertainty in general. “We just don’t know. So people are writing down something that seems to make sense to them, but have no conviction.”

That is where I see most businesses are today with their AI strategies. C-level leaders are pressured to write something down, so they build a strategy that makes sense, but they have no conviction in it. Much of the apathy in leadership comes from this lack of conviction.

It is also why leadership bends agents to the legacy business and operating models. They have conviction in the strategy that worked for years, and if they could just get back to it…

The challenge with disruptions that create tipping points is that there is no going back. Critical slowing down is a powerful early warning sign of a system that has neared its tipping point, and we have seen several indications of critical slowing down.

Stable systems want to return to their equilibrium even after a disruption. The speed of that return or restoration is the giveaway that something fundamental has changed. That’s what we need to be looking for.

In highly stable systems, it takes big shocks to cause a disruption, and the system moves back to equilibrium quickly. In systems near their tipping points, small shocks cause outsized moves, and it takes longer for stability to return. Which are you seeing in your business?

As a system reaches its tipping point, the variables that manage its restoring force go from big numbers to near zero. The system still appears stable because it returns to the same equilibrium, but the geometry has fundamentally changed, resulting in a slower rate of restoration.

In complex systems theory, we all wait for the Big Push, or the one that finally makes the system’s instability obvious to everyone. Agents are the most likely candidates to drive a Big Push. It’s just a matter of who creates the market impact with them.

In the next 3 months, we will watch marketplace systems and the businesses that operate in them get tested by multiple pushes.

The Givens: Defining The Disruptions

Read more

Google published a paper this week on policy discovery and optimization at the policy level vs. the workflow level. It uses prior runs of the policy discovery workflow to find the best policy for a task or workflow. The optimization happens at a level above the workflow, but the improvement mechanism is the same. We use what happened in previous runs to improve the next one, and this generalizes from tactics to strategy or policy.

A procedure graph stores the best version of the procedure we currently have. BUT most researchers run into the same problem. The current version may be incomplete, inaccurate, or worse than another version we haven’t found yet. We need to start with what we have and improve it as we work…while we’re using the imperfect version and getting value out of it.

There Is No Finish Line With Agents

That’s the challenge of learning. We aren’t given a blank slate or a perfect one. However, most approaches to learning assume you can start from a blank slate and have access to someone who has complete information. I have notes on almost 50 startups that I have evaluated this year. I stop them after the demo and take over using the product myself. Within a few minutes, I find some variation on the demoed workflow that fails. I watch founders and AI engineers deflate, so I have to do a preamble before I ask my two big questions.

With agents that rely on models, there will always be failure cases. That’s the part about AI, information models, and agents that you don’t understand until you’ve launched several of them. Agents and models are never done because your information is never complete. You must deal with that first-principle constraint architecturally. I ask the startup’s team two questions:

1. How many unresolved failure cases have you discovered, and what’s your discovery process?

2. What is your failure case resolution process, and how many have you resolved?

There are a lot of right ways to approach the problem, but the only way an AI startup can be long-term viable is to have a process that increases automation as the startup scales. Learning is equal parts issue discovery, root cause analysis, solution selection, and solution validation. No part of the cycle can disrupt customers or internal users. I can describe the solution’s constraints, and that helps me evaluate the startup’s answer without prescribing its solution.

In this article, I will explain the weak self-improvement methods I use in early maturity phases, using my content-to-cash workflow. We start with the expert, record what the workflow does, and use corrections and outcomes to improve it. Mine isn’t THE way, and it’s far from the only way. Adapt my approach to work within your business’s constraints for the best results.

Read more

It all started in 1965 with Irving John Good’s article, “Speculations Concerning the First Ultraintelligent Machine.” AI, AGI, ASI, the singularity, and the concept of RSI all trace their roots to his musings on what a machine intelligence might look like. In 1993, Vernor Vinge predicted the singularity, superintelligence, or RSI would happen between 2005 and 2030 at a NASA symposium. Neither the terms nor the speculation are new.

David Chalmers wrote the core argument for RSI in 2010, saying, “increases in intelligence … always lead to proportionate increases in the capacity to design intelligent systems.” That’s the core thesis. As models get more capable, we will see an emergent capability for self-improvement grow as well. It starts out human-guided, but advances from there.

Nick Bostrom took that a step further in ‘Superintelligence: Paths, Dangers, Strategies’ and posited that the rate of self-improvement would fall into one of three buckets. He coined the term ‘fast takeoff’ that we often hear researchers use today. Fast takeoff relies on the AI’s architecture imposing few barriers to optimizing itself. This is where we start to see a key differentiation emerge.

If RSI is weak, improvement would be slow and take decades or more. If RSI is strong, improvement could happen in days, hours, or even minutes. There’s no strong agreement on labels and definitions for any of this, so I’ll simplify the debate by selecting the terms I use. I have been in the field since 2012 and helped define applied machine learning, so I’m taking a bit of privilege here.

There are two lines of Recursive Self-Improvement research in the modern AI era: weak and strong.

Weak RSI

We have had multiple methods for weak RSI for over a decade. Weak RSI relies heavily on people to drive the improvement cycles, with other models accelerating and scaling the progress. However, it still relies on a human teacher. The self-improvement doesn’t generalize beyond the narrow domain we teach, so progress is slow, and it plateaus.

My current series on workflow-centric knowledge graphs is an example of a method that starts with weak RSI and builds towards automating weak RSI into a very limited strong RSI. It is limited because the learning is bounded by experiments, so it won’t meet all the criteria for strong RSI or fast takeoff. It fits with Bostrom’s moderate category.

The inspiration is a Gödel machine that “rewrites any part of its own code as soon as it has found a proof that the rewrite is useful” and the derivative work ‘A Family of Gödel Machine Implementations’ that proposed experience-based Gödel machines. If you have been following any of my series on agentic architecture and information models, you see the parallels between my applications of Pearl’s work on causality and experimental cycles. Pearl’s work provides the proof. Experiments create the experience that determines the rewrite.

I also put the learning from both processes into information models, but most RSI research focuses on training the learning into an AI model. I’m working around our current AI’s architectural optimization speed limits by using something entirely different. It’s cheating, but we get to do that in the applied AI domain.

We have a lot of these methods that extend weak RSI and make it more capable, but they don’t meet all the criteria for strong RSI. Darwin Godel Machines were proposed earlier this year. They replace provable utility with Darwinian empirical selection. It accelerates learning at the expense of reliability (did it learn the right thing?).

AlphaEvolve is another, stronger RSI approach from GDM that “found an algorithm to multiply 4×4 complex-valued matrices using 48 scalar multiplications, improving upon Strassen’s 1969 algorithm that was previously known as the best in this setting.” It opened up progress on solving mathematical challenges, but again, reliability and solution consistency issues persist.

Even stronger RSI methods run into very human challenges. New research is promising, but often proven wrong. That’s true for AI and our own self-improvement.

Strong RSI

Strong RSI is when a teacher model (typically multiple models) trains a new model to be better than the sum of the teachers. This is a uniquely human capability today that the AI labs are trying to replicate. The best teachers with the brightest students can teach their students to exceed them. This is part of the ‘G’ in AGI (artificial general intelligence).

Strong RSI enables AI to generalize exponentially (beyond the teacher’s domain) and avoid plateaus, with compute being the only limiting factor. That’s the ‘S’ in ASI (artificial superintelligence) and what could enable AI to scale learning beyond human knowledge. This is what every major AI lab is working towards, and what GDM has supposedly solved with its current training run for Gemini 4.

The problems with this are significant. First, there has never been a convincing demonstration of strong RSI. It is theoretically possible (that’s been well researched and supported), and so far we have nothing to support the counterfactual that it is impossible. But every architecture we have discovered imposes limits on self-improvement rates.

Every lab will quickly put forward the number of scaling axes they have. That’s a fancy way of saying the number of promising methods they have found that improve the model’s capabilities. Every axis is tracked on an exponential curve that predicts the capability gains it will deliver. These curves all hold in the first few improvement cycles and then plateau.

I have been there before myself with physics, natural language, supply chain optimization, pricing, and customer behavioral models. Early acceleration quickly yielded to a plateau in every case. Still, each time I had the same exuberance and false hope. The model keeps progressing up the exponential curve on your graph, but it doesn’t deliver real-world capability gains.

You quickly realize there’s a disconnect between your graph and the one that really matters: real-world performance and outcomes. Reality is often disappointing. I’m not saying that strong RSI won’t be solved. I’m saying the only test for strong RSI is real-world performance.

So far, everything we are seeing on the strong self-improvement front plateaus after an initially promising phase, is a result of more compute vs learning optimization, or comes from better data/more complete information.

What Has Google DeepMind Cracked?

The source of this story is a single account on X. It’s not affiliated with any founder, researcher, or even a lab that I could find. The post just capitalized 3 letters (RSI) in a congrats message to GDM. That’s enough to drive hundreds of articles and thousands of social media posts speculating on whether one lab has cracked strong RSI.

We are seeing an old type of exponential scaling law: Hype. If there’s any truth to this rumor, it’s built on this paper: ‘Procedural Graphs: Self-Evolving Execution Structures for LLM Agents’. Its primary author is an applied researcher from GCP AI, and he’s definitely on to something.

Sergey Brin’s return to Google and his influence on GDM’s research direction have put renewed emphasis on strong RSI. Multiple insiders have reported that GDM has redirected significant resources to researching strong RSI. Maybe they found something, or maybe adding significant compute resources has led to a significant advance.

There is so much speculation, but no evidence. Even if claims of strong RSI come forward, it will be a while before they are confirmed in real-world settings, not contrived benchmarks. OpenAI’s recent advances in Navier-Stokes have not survived the real-world’s scrutiny unscathed. Even the authors called the AI’s proof ‘AI slop’ and spent months turning it into something that mathematicians could read and evaluate. Most who have evaluated it say it is impressive, but incomplete.

That’s how the real world works. Progress is incremental until one day it isn’t. We could be close to that day, but a post on X isn’t enough to convince me, and it shouldn’t be enough to generate this level of response. I’m seeing hype scaling exponentially, but strong RSI is still unsolved.

Procedural graphs are back in style as fuel for agents to learn and self-improve their world models. Everyone’s approach to procedural graphs is slightly different. Here’s one version from UConn and Notre Dame. Google, GIT, and Peking University recently published their approach. What I have been explaining here for over a year is finally getting broader attention. My Four Tenets Of Knowledge Graph Engineering explain why this was unavoidable.

The main contribution in these papers and others like them is exploration and self-improvement. While everyone focuses on the RSI to AGI to ASI pipeline in model capabilities, real progress is being made using an entirely different approach. Compute constraints have forced Moonshot and Zhipu to take a data-centric approach to improving their models. So far, the results show that the same gains can be delivered by advances on the data front.

Transformers have done a lot for the field, but as early as two years ago, many were calling a plateau. Whether that’s due to scaling constraints on hardware, data constraints, or the limitations of the transformer architecture itself isn’t well established. Still, work has progressed in new directions: harnesses, loops, graphs, and learning.

Researchers have realized that in real-world conditions, most knowledge graphs are partial descriptions of a system or process. Procedural graphs, or what I call workflow-centric knowledge graphs for NDA reasons, combined with agents, provide a path forward to learn workflows through observation and direct interventions.

This begins as a weak form of self-improvement that relies on observing people doing work and performing direct interventions. We need a baseline, and experts offer the best version of it. Once the baseline is built, strong self-improvement is required. Agents must bring new information and later new domain expertise into the business.

The power of a procedural graph is that it can benefit people and agents equally. It prevents information asymmetry after stronger learning enables agents to discover new information and develop new domain expertise.

The Implications Of Strong Self-Improvement

As Doctor Strange said, the warnings need to come before the spells. Before I explain how to advance on the information maturity model in my next articles (AKA information self-improvement maturity model and Parallel Maturity framework), I must explain the risks.

The learning process turns the partial descriptions we start with into more complete information models with each cycle. I use WITA cycles with clients because they support people and agents and can learn from both. They also meet enterprises at whatever maturity and capability phase they happen to be at. Most aren’t ready for an enterprise world model, so we need baby steps to get started and bridges from where the business is to where it wants to be.

Time is running out. Reserve your seat today. Learn to build the bridge that transforms the business and monetizes AI.

You can also learn about the mechanics of more complex systems this way, and that has higher risks than most realize. The UConn-Notre Dame paper digs into methods for learning causal transitions, or what action causes the transition that causes our desired outcome. If that sounds like a triple, you’re picking up what I’m putting down. The combination of causal discovery and learning and system representation learning (description of the system and its mechanics) makes mapping complex systems through observations and direct interventions feasible.

While everyone is looking at the AGI/ASI danger hype, the near-term threat is growing one experiment, WITA Cycle, or playground at a time. This community is made up of pragmatic, well-intentioned people. However, the rest of the world is far more devious, and so are their policies. As I have explained before, we don’t need more capable models. We just need more complete information.

And even weakly ‘intelligent’ agents have viable methods to get it.

The AGI Head Fake

In the last week, we have seen multiple warnings about AI advancement and the growing risks that it is creating. This one from X was seen over 90 million times in less than 24 hours. The claim is accurate. Multiple labs are racing to crack self-improvement and remove a major barrier to ‘superintelligence’.

But they’re not close, and their most promising directions have come up short. Posts like this one are positive because they focus attention on the long-term impacts of AI at a time when we can do the most to mitigate the catastrophically negative impacts. However, posts like these also make it seem like the biggest near-term risk is AGI or ASI.

And it isn’t. I have seen the frontier and our risks look very different. We already have evidence that we don’t need self-improvement, AGI, or ASI. The models we have today can power entirely new categories of exploits that are a midpoint between our current cybersecurity threats and the future threats of AGI.

We think of AI in the current cybersecurity paradigm, and warnings like this one are common. Thousands of companies, financial markets, utilities, and governments are racing to harden themselves against conventional cyber threats that AI can supercharge, but what happens when our current models get access to more complete information?

AI and agents enable a new type of exploit that isn’t getting enough attention. We’re used to conventional exploits that get in, elevate permissions, and then carry out an attack on the technology. AI supercharges those by finding new exploits faster and making detection significantly harder. Since this is a well-understood paradigm, we are well equipped to build mitigations and respond to a breach so the worst doesn’t happen.

The complete information paradigm is being completely overlooked, and there’s a ton of evidence that we don’t have strong prevention, detection, and mitigation methods.

An Exploit That Doesn’t Need AGI

AI enables an attack from the outside that exploits architectural vulnerabilities without breaching the host system. It just needs access to the marketplace or business system that our technology manages. Models learn the system and its mechanics, then use that map to discover vulnerabilities in the marketplace or business system.

Distillation is a really good example of the new category, even though that’s not how it was discovered. China’s AI labs have deployed bot armies to query US frontier models and turn all those responses into a training dataset for their AI models.

There’s no architectural breach involved, so all the defenses built to prevent unauthorized access or detect unauthorized activity on the internal system are never triggered. All you need to perform a distillation attack is access to the API or platform. We’re watching a new type of unauthorized access that looks like a paying customer using the product. The vulnerability in the architecture makes distillation possible. The vulnerability in the business model makes distillation lucrative.

Frontier AI leaks a microscopic amount of IP with each response. That vulnerability isn’t new. Models have always leaked information about themselves and their training data, but AI-powered bots make it feasible to extract and reconstruct that information into something that can be monetized.

Businesses Leak Information Too

I have been in competitive intelligence since 2011. The purpose of the field is to gather the information that businesses leak and turn it into actionable intelligence that gives one business an advantage over its peers. One of the most common tricks is comparing job descriptions to earnings calls and press releases.

All three are a form of intentional information dissemination, but the comparison is where you surface treasure. Businesses publicly share only some of their strategic initiatives. You can piece together what they hold back by extracting that information from job descriptions. When I started, technology only supported the scraping and aggregation of this information. Early NLP helped bring down the human effort, but it was still painstaking work.

Today, language models automate most of the workflow, making this feasible to do at scale. I have seen systems that compare dozens of ‘what they say publicly vs. the information they leak unintentionally’ to build a 360 view of every large and mid-sized business in very large industries.

Hedge funds use them to invest ahead of major announcements and get out before major problems become public. Large enterprises use them to stay ahead of competitors and often to undermine them. You can distill business and operating models, strategy roadmaps, product roadmaps, pricing strategies, customer acquisition strategies, and even entire industry directions from the information that leaks.

Experts Leak Information

It wasn’t feasible at this scale when people did the majority of the information gathering, aggregation, transformation, analysis, and comparison work. But we don’t anymore. The art of competitive intelligence gathering is only known to a small community of people, but even simple models make our craft or knowledge domain distillable.

Distillation happens on multiple levels. You can distill a training dataset or model weights. It’s like reconstructing a database you have never seen through thousands or millions of queries. What process graphs and the learning architectures from the start of this article enable is learning at the next levels. These methods can distill domain expertise from the information that leaks about it with every workflow an expert completes.

I was one of the domain experts who distilled my knowledge and built some of the platforms that have taken over competitive intelligence. I was well compensated for it, but that’s not always the way it works. Context extraction through distillation is much easier to automate today. Even simple bots enable this type of exploit at scale.

You can distill an entire job’s domain expertise by observing thousands to millions of workflows. There’s nothing wrong with this process as long as it’s transparent and all parties agree to it upfront. In competitive intelligence and AI, not all actors are ethical.

Unethical Self-Improvement

Meta tried to force this type of distillation on its workforce, and it backfired spectacularly. I have seen similar stories from call centers. One well-known wireless carrier tried to distill domain expertise by tracking call center employees, and the employees sabotaged the initiative through data poisoning. They logged the wrong resolutions for some calls. Agents trained on those datasets performed miserably in the real world, causing the initiative to be shelved.

Play stupid games. Win stupid prizes. Unfortunately, unethical actors are usually much smarter about how they implement their distillation campaigns. They go to great lengths to hide the work. Sometimes, even ethical actors stumble into the grey areas of unintended consequences.

I often call out SAP for doing knowledge distillation the right way. They put their disclosure and consent requests into contracts. Customers didn’t have to opt in, but over 25,000 did. There was a strong alignment. SAP’s platform would improve through distillation, and those improvements were beneficial to customers.

However, that alignment ends when people start losing their jobs. SAP’s customers got a more efficient platform that could manage more of the workflow. Some monetized that by laying people off. The people whose domain expertise builds the features and models rarely profit from them, as I did, because they don’t have the technical knowledge to do the building themselves.

Taking Humans Out Of The Loop

When competitive intelligence was a human-dominated domain, most practitioners adhered to ethical standards and professional norms. We all know that many of the top executives and engineers at the world’s largest companies suffer from a combination of loneliness and oversharing. You can get almost anything out of people by just paying attention to them and being kind.

The leaks that have come from after-hours conversations at this week’s Goldman Sachs Communacopia + Technology conference are eye-watering. It’ll happen at Dreamforce next week and every other week of conference season.

But that’s not how the game is played, and people ensured that the game didn’t start to play out that way. There’s a reason we all attend conferences with press badges. We aren’t part of the news media, but if you say it to us, it will get published. Competitive intelligence takes advantage of ignorance, but we won’t weaponize human frailty.

I use competitive intelligence as the example here because it is an undeniably grey industry that does a remarkably good job of self-policing. Some of that is based on individual honor and dignity, but when temptation strikes, dignity and honor are enforced by consequences.

If an individual or firm takes their gloves off, everyone else does the same with them. Competitive intelligence has rules and norms because no one wants the really ugly stuff being turned on us or anyone we know. If I do it, then everyone learns how to do it, and now I live in a world where yet another ugly tool becomes commonly used. Everything from what Cambridge Analytica did to using ads for targeted surveillance was an ugly tool going mainstream.

Bots take people out of the loop, and new exploit categories become common.

Information Asymmetry & Brute Force Are Dangerous Without AGI

Bots don’t get built with an ethical compass. Even those that do will bend the rules in the service of their core mission. They will violate the terms and conditions. They will hide to observe experts completing their workflows even when those experts haven’t given consent. They will catfish executives to get insider information.

I said this in my book, and it’s worth repeating. The purpose of building technology is to serve our strategy and perpetuate our objectives. However, the true nature of technology is to serve and perpetuate itself. And we don’t need to achieve AGI before we start to feel the impacts of that reality.

Even simple bots seem to adhere to von Clausewitz’s concept of absolute war. von Clausewitz used absolute war as a theoretical extreme or global maximum. It is a conflict where all constraints have been removed by one or all sides. When people manage conflict, absolute war, conflict, or competition remains a theoretical concept. Honor, dignity, norms, and consequences.

Bots bring the theory of absolute competition into business reality. There is no consequence for the bot beyond not achieving its objective. What happens after success doesn’t exist for bots, so there are no consequences. We give the agent a policy, and it executes through any means available, even ones we couldn’t have predicted.

Brute force gives bots the ability to exploit architectural weaknesses most have taken for granted. Procedural graph learning gives bots the framework to distill domain expertise and deploy brute force at expert levels. It also enables them to learn the rules of our business systems to discover vulnerabilities like information leaks.

And distillation isn’t the only vulnerability.

New Exploits Exist Everywhere

Technology platforms have well-known vulnerabilities. What’s worse is that many of these exploits rely on using the system as it was intended to work, for purposes it was not intended to support. And the only mitigation preventing someone from exploiting them is the level of effort required to do it. The infeasibility of speed and scale were the only gates protecting most systems that are deployed today.

That’s what kept models safe from distillation for almost a decade. Could it be done? Sure. The time or number of people required to do it made it infeasible to exploit before the model was updated, and you needed to start all over again.

Our passwords used to operate on a similar principle until raw compute power made brute-force attacks viable. We had to respond by limiting attempts, implementing MFA, and now implementing even more robust identity verification methods. It’s an older version of brute-force, but the same theory generalizes.

All of our encryption methods have brute-force vulnerabilities, but it’s just infeasible to exploit them due to the time, compute, or level of human effort required. We are already seeing cryptocurrency’s security protocols and marketplace security being bypassed using AI to discover new vulnerabilities. It’s only a matter of time until brute force also becomes feasible from a unit economics standpoint.

Bots, even relatively dumb ones, are leveraging speed and their ability to scale cheaply to take advantage of this vulnerability category. There’s a novel interplay between automated discovery and exploit. One agent learns the system and its mechanisms. Another finds ways to turn those mechanisms (exploit the rules) against the system or circumvent them altogether. A final agent uses speed or scale to act on the exploits it finds.

Most architectural patterns are open to exploits that were deemed an acceptable risk. Others exist because the exploit itself was unimaginable at the time the platform was developed. The vulnerability has been propagated for years, and if you’re not spending time looking for this category of vulnerability, it often remains hidden.

Exploits Are Fundamental Assumptions In Most Marketplaces

Buying concert tickets has become infested with bots exploiting a weakness in the online marketplaces’ architectures well before LLMs. More people want to attend the most popular concerts, shows, and sporting events than there are tickets. We created lines to manage this in the physical world, but when we duplicated the line in the digital world, it wasn’t architected to deal with bots.

Bots use their speed to always be at the front of the digital line since they execute orders faster than people do. Tickets sell out in minutes, so most people are forced to turn to the secondary market. Bots leverage the speed exploit to force people onto a marketplace that their operators control.

The US FTC and DOJ did not go after the bot operators. Last year, they sued the ticket sellers (Live Nation Ticketmaster). It promised stricter bot controls, but so far, the improvements have been minimal. In this case, the bots have successfully shifted the costs to consumers, and ticket sellers seem happy to collect their share.

Nike’s SNKRS app uses similar marketplace architecture for rare, high-demand sneaker releases. Bots have taken over this line system as well, pushing people out of the primary and into the secondary market. In this case, it has undermined a core pillar of Nike’s business model. They built a direct-to-customer (DTC) model after years of working with outlets like Footlocker. Margins rose, but it was short-lived.

Nike diehards waited years for the company to address its bot problem. Then they started exploring other brands. Diehard sneakerheads went first, and a significant portion of Nike’s customers trickled after them. Nike is a brand like Starbucks and Apple. Anything that erodes the brand’s most vocal enthusiasts will impact revenue. In this case, bots forcing people onto a secondary marketplace caused Nike’s revenue to take the hit. Its stock is down 50% in the last year, and its DTC strategy needs a turnaround.

Bots have taken over restaurant reservation systems. They resell reservations at ‘hot’ restaurants to people who can’t get a reservation because the bots took them all. Restaurants don’t see any of the secondary market revenue, but they eat the loss of an empty table with every no-show.

Lines need to be rearchitected to prevent the bots’ speed exploit, but most marketplaces are just patching symptoms. Bots now own most marketplaces that sell limited supply or high-demand items. GPUs and consoles are following tickets and sneakers.

Job Markets

But it’s not just lines. Job marketplaces have been overrun as well, and this is where it gets a lot more serious. In 2025, LinkedIn received 11,000 job applications per minute, which was a 45% increase over 2024. Most attribute the uptick to auto-apply bots. Some companies are getting 1,200 applications as soon as they list a job. This has caused many to take their job postings off these marketplaces in what the CEO of Greenhouse called a doom loop.

What used to be high-value marketplaces that funneled the best candidates to hiring managers have become the opposite. They are bot superhighways sending resume slop flooding recruiters’ and hiring managers’ inboxes. So of course companies are responding by using AI for screening more. That’s another bot-fueled doom loop. The best job seekers don’t bother applying through the job marketplaces, and the best jobs aren’t being posted there anymore.

People without bots have lost access to job marketplaces, and the deterioration of job marketplaces is creating even more accessibility inequalities. If you don’t know someone on the inside, getting past the automated screening bot is nearly impossible.

People are having bots stand in for them during job interviews. That’s how thousands of North Korean IT workers landed jobs in US companies. Someone in Arizona ran a bot farm that helped place North Korean workers at over 300 US companies.

As a result, the hiring process has changed. There are more interview rounds and the demands on candidates’ time are growing. I keep hearing people say that they don’t have time to interview while they’re employed. Unemployed people can only apply for a few jobs at a time. Fewer applications combined with a higher rejection rate has increased the time it takes to find a new role.

Content Markets & The Dead Internet

It’s happening on the content side as well. Bots are over half of all internet traffic, and nowhere is that more obvious than on social media.

LinkedIn was found to have the highest rate of AI slop posts out of the major social media platforms. Bots scale content creation without a person having to do much work. Slop posts got traction, and some creators built communities they could monetize in a matter of months. Slop comments added another way to scale reach without scaling the labor required to do it. AI slop pushed people off the platform.

While LinkedIn doesn’t publish platform traffic metrics, third parties have found ways to estimate changes over time. Some measure median impressions across thousands of creators. 98% of LinkedIn creators have seen a decline in impressions. The median decline between 2024 and 2025 was 47%. Another researcher analyzed over 600,000 posts and found that reach was down 50% and engagement had fallen by 25% in 2025.

People are probably leaving or spending less time on LinkedIn. In response, LinkedIn rolled back AI content generation features, added a user moderation option, and changed its algorithm significantly. Creators are leaving the platform after seeing declining reach and drowning in AI slop. The new algorithm doesn’t seem to be helping.

C-level leaders I talk to no longer see the blue badge as a strong credibility signal, and ‘LinkedIn Influencer’ has become synonymous with someone who is only an expert online. Bots have created a perception problem that has extended beyond just the platform itself.

Once the community contracts, revenue typically follows soon after. Bots created a problem for platform users that trickled down into a creator problem. Watch LinkedIn’s revenue for the next order of impacts.

Once a platform’s mechanics are undermined or gamed by bots, the path back is a long one, so any revenue declines are persistent.

The Architectural Flaws Are Undermining Revenue

With ticket sellers and sneakers, the physical barriers of lines were never translated to the digital lines of ticket marketplaces. You had to get there early or pay someone to stand in line for you. Sellers knew they were selling to a real person, not a robot. They could impose limits on how many items an individual could buy with a quick ID check. None of those barriers to bots work as well or without controversy in the digital world.

Editors, newsstands, bookstores, and even librarians used to be gatekeepers against slop content. There were a limited number of authors who got through the process each year because space was limited. Combine strong gatekeepers with publication constraints, and you have a functional content marketplace. Again, none of that translates well in a digital world of minimal moderation and nearly unlimited space for content.

The extreme version of that is AI-generated music being streamed by bots on Spotify and YouTube. The only person in the loop is the one cashing the checks.

Our digital architectures often reward the bot. They get all the best tickets, sneakers, and reservations. AI-generated resumes sail through the ATS’s keyword filters. AI-generated content can spread on the back of thousands of bots, creating overnight influencers or sending a startup’s product launch viral.

The rise of bots often benefits the platform running the marketplace for a while. Concerts and sneakers sell out. The metrics social media platforms use to charge for ads look like they’re taking off.

But the bill eventually comes due, starting with one market participant. Users leave a social media platform when bots overwhelm their feeds. Sneakerheads disconnect from a retail brand when bots make the buying experience stop being fun. Concertgoers demand regulations to protect them from bots colluding with ticket sellers. Billion-dollar models are copied in less than six months, threatening to undermine the unit economics of frontier AI.

There are mixed incentives for businesses and marketplaces, so bots are allowed to propagate. They learn the rules, find exploits, and eventually undermine the marketplace’s dynamics. That creates new winners and losers, but the bots don’t care. They optimize for their own goals without regard for the system-level effects. The more bots optimize, the more unstable and inefficient the marketplace becomes.

Bots are just another form of denial-of-service or brute-force attack that hacks more complex systems, from domain expertise to businesses and marketplaces. They learn the rules and use that knowledge to play the game to their benefit. When bots understand the rules better than people do, we need a way for people to quickly catch back up.

That’s what makes the procedural graph approach my go-to. Whatever the agent learns doesn’t get trained into an opaque model. It is stored in an information model that’s largely human-interpretable. As agents improve, we do too. Preventing information asymmetry and discovering vulnerabilities as soon as the agent does are critical to maintaining parity.

Once we lose parity, we lose the ability to manage our businesses and participate in marketplaces without using agents. That’s not a road we want to follow because consequences don’t constrain bots’ actions. And we don’t want to live in a world subject to the bots’ versions of absolute competition and warfare.

1. My client has a knowledge graph that was built domain-first, and it can’t support an agent doing work. This is the worst-case scenario.

2. My client has a vendor proposal, and they were told that their data warehouse would fill in the blanks. This is actually easier from a technical and organizational standpoint.

Starting from 0 costs less than abandoning a graph that a team spent 9+ months and a lot of money on. Sunk costs create inertia. Existing artifacts drive architectural decisions that create technical debt. Both are harder to overcome than a few bad assumptions.

“How do we know when we’re done?” Done must be defined upfront, and it isn’t well defined by creating a set of questions that the information model must answer. Few people answer questions for a living, so the information models that support them must be more complete than that.

A domain has no boundary in the enterprise as it does in science or academia. That’s why knowledge engineering methods that have worked in those domains for years have come up short in the enterprise. If I ask you to define the start and end of the marketing domain, you’ll probably do it with where one workflow hands off to another.

Sales, marketing, finance, etc. are as large as you decide to make them, so every entity you add reveals 3 more that are related and could be included. Teams keep going until the budget runs out or their patience does. They deliver whatever is in the graph on that day and validate that it can answer a set of questions. The result demos beautifully but can’t support an agent managing a real-world workflow because workflows require more than just Q&A.

A workflow has a first step, a last step, an intent, and an outcome(s). That gives you scope, order, and exit criteria.

I’ve argued the case for workflows over domains before in prior articles. This is the practical version. In this article and series, I will explain how to start, what to build first, and the methods that make those decisions more obvious.

The example I’ll use is my own content creation workflow. It turns a news story or a research paper into a Substack article or a LinkedIn post with the intent of increasing course enrollment. I’m using it because I can show you parts of the real artifacts, and because it has the same failure modes that most of us encounter while building knowledge graphs.

Why The Graph You May Already Have Doesn’t Work

The domain-first graph works for science and academia, so it’s not wrong. If enterprise domains were scientifically rigorous, that approach would work just fine. Domain-first is how ontology engineering has been taught for 20 years because it works so well when we have strong shared knowledge.

Ask which agent uses the knowledge graph and for what workflow, and there isn’t a good answer. The graph was built to describe a domain completely rather than to support an outcome. Businesses monetize outcomes even though they need context, semantics, domain expertise, etc. to deliver it. Disconnect anything from outcomes, and it will quickly diverge from the business.

Read more

It all started with big data. Every part of the business creates data. Every customer interaction with the product creates data. BUT THAT’S NOT NEW. That was happening in the 1800s…BCE. Some of the earliest stuff humans wrote down was data from business activities. There’s a 4000 year old tablet that recorded a customer complaint.

We have been capturing data for thousands of years and have been challenged by storing it so we can find it when we need it for just as long. Databases did not create data governance challenges. Capturing, transforming, storing, and distributing data are business problems that are as old as language.

I helped define data science in the mid-2010s, but I didn’t invent models or modeling. None of us were the first to use models in a business. Some of the earliest examples come from merchant trading companies that used mathematical models to determine venture profit sharing. Before that, models were heuristics and decision-making frameworks that lived in people’s heads. Training, deploying, standardizing, aligning, verifying, and maintaining models is also an old business problem. Deep learning did not create model governance challenges.

Over the last 20 years, we have transferred more business domains and activities that used to be 100% owned by people into technology. Big data was just the tipping point where technology could store data more efficiently than a file cabinet. Google indexed information for rapid access more efficiently than a library and enabled the internet. However, Google and search engines did not create knowledge governance challenges. The Library of Alexandria had the same challenges. As did monastic and royal libraries in India, China, Persia, and every other early civilization.

Businesses face the same information governance challenges. Information used to exist in people’s heads, and it was 100% in the domain of people. People used to manage (not always well) information governance challenges.

Before models and AI, people used to serve inference to the business. People managed (again, not always well) model governance challenges.

The governance debate reveals a massive lack of business awareness, which often reveals how badly businesses are managed.

My next instructor-led Data & AI Strategist Certification starts soon. Learn critical skills that help you stay relevant. Learn more and enroll.

Every Information Governance Problem Is A Business Confession

I often read comments on my posts that I categorize as, ‘You Couldn’t Get Me To Admit That Under Duress’. A Cofounder of an AI startup said,

“Moving logic into information models also creates a governance problem as rules become reusable assets, someone needs to own, version, and retire them.”

So you’re telling me that wasn’t happening in your startup, and you don’t see how some businesses have already addressed this issue? The problem may be new to you, but it is not new to business.

Another cofounder of the same startup commented,

“Exceptions are probably where this gets really difficult as humans tend to override the documented process sometimes.”

So you’re just discovering this common business reality and telling me you don’t understand how to manage it.

An ‘Enterprise AI Executive’ said,

“Traditional databases were never meant to house complex business logic, but autonomous agents need that context instantly.”

It sounds like you don’t realize that people need that context at the point of acting and deciding, too.

A Senior Director of AI Enablement said,

“The shift from storing business logic in code to making that logic accessible to agents is a much bigger change than it first appears. It changes not just the data architecture, but the way enterprises make their workflows understandable and executable by AI.”

How did your business make its workflows understandable for people? Do you not realize that many of your customers do this as a managed process?

I could keep going, and these examples are from just 1 post. You’ll notice I’m not picking on early career people. Everyone I quoted is senior enough to have seen enough. They should all know better. These aren’t new business problems, but it sounds like a whole lot of people are learning about them for the first time.

What this reveals is how few business leaders are aware of fundamental business challenges and business management methods. These comments and the hundreds of others like them reveal how many people don’t know about process discovery and mapping, change management, knowledge management, and business transformation. If you weren’t doing this before the governance debate began, were you really managing the business or just hoping things worked out?

These are the types of things you tell a priest or a consultant under an ironclad NDA. Not the types of things you confess to on the internet.

Governance Should Be A Migration, Not An Invention

Data, information, and AI governance should be migrations where existing processes are optimized with the support of new technologies. But if your business is run by executives like these or you choose vendors who design products with these kinds of blind spots, how can you?

Technology often reveals business challenges that haven’t been well managed in the past. That’s fine, but when we call them new, we disconnect them from the business challenges they must align with. We solve information or model governance challenges for agents and don’t realize that people need the same solutions to work for them.

I have helped multiple large VC firms add a new core tenet to their theses. Agentic tools (harness, loop, graph, and information model) are huge right now. So are agents that manage complete workflows and business processes. How do you evaluate whether the tool supports the ‘right’ agentic paradigm or whether the agent will be adopted by enterprise customers?

I get asked that a lot, and I use a basic tenet that’s rooted in one of my AI product management frameworks to answer.

What works for agents must also work for people, and the best way to architect an agentic harness or information model is to extend what works best for people.

People aren’t going anywhere. We have seen that proven repeatedly over the last two years. People are critical parts of the agentic workflow. Agents and the tools that support them must integrate into existing workflows even though the workflow itself is changing and transforming.

AI & Agentic Product Design: Augmentation

Read more

Why are businesses hiring so many AI product managers? Value creation has become a bigger gap with higher business impact than technology creation. But there aren’t enough qualified AI PMs to meet the demand, so a lot of this work is being passed on to ICs. Some of the middle managers who have been let go owned the slide decks that helped quantify the team’s value and keep the team’s work aligned with business value.

AI PMs and technical ICs are expected to have influence at the executive and C-levels. Nothing creates more influence than value creation, but if you can’t quantify the value your work delivers, that influence remains out of reach. AI PMs and technical ICs need to know how to teach leadership how to measure AI’s value.

Measuring Activity Vs Value

AI products generate an extraordinary amount of measurable activity. They produce answers, summaries, recommendations, images, code, predictions, and automated actions. Product dashboards often count every user, prompt, token, conversation, accepted suggestion, and generated asset. These numbers are immediately available and usually move quickly, so they often become the evidence used to justify an AI investment.

But activity is not the same as value, and we should have learned a bigger lesson from tokenmaxxing. I see a lot of businesses still falling into this trap. It isn’t new. We used to measure developer activity with lines of code instead of value with revenue generated per feature or project.

Many activity metrics can point to success and failure. A customer can ask an AI assistant more questions because it is useful or because its first five answers were wrong.

Some metrics point to unused artifacts. A company can generate millions of meeting summaries without improving a decision. Summaries often show how little value some meetings create, but someone must act on that information. Summaries can generate tasks, but prioritization is lost.

When organizations mistake activity for value, they optimize the product around what is easiest to measure rather than what the customer or business is trying to accomplish.

This is the difference between outcome metrics that measure value and disconnected metrics that measure everything else. Outcome metrics measure whether the workflow produced its intended result. Disconnected metrics measure activity somewhere inside the workflow without demonstrating that the desired result occurred.

Both types can be useful (activity can be a leading indicator of value if the value chain is well engineered), but they play different roles. Activity and model metrics help diagnose how a product operates while outcome metrics determine whether it deserves continued investment.

The metric test is simple: Could this metric rise significantly while the customer’s desired outcome remains unchanged or worsens? If the answer is yes, the metric is disconnected unless it is accompanied by evidence farther down the outcome chain.

Outcome Metrics Measure Completed Value

An outcome metric measures a meaningful change (enough to get C-level leaders or customers excited) in the customer’s workflow, operating performance, or business results. Disconnected activity metrics often stop at what the model produced. Outcome metrics follow the intervention forward to determine what happened because the product was used.

Strong outcome measurement typically has three characteristics.

Read more

Read more

TIP (Testing In Production) is the most common approach to agentic evals. Evals require an answer key, and most workflows don’t come with one. Again, we need a bridge because the business can’t wait for the perfect eval.

AI product managers are expected to understand this process and be able to support it. I teach it as part of my AI product management certification, but few others do. Based on the interviews I have conducted over the last 12 months, few have built the skills to do high-quality evals when there’s no answer key provided.

Most answers are some version of, “I get the users, domain experts, or SMEs to write them for me.” That shows a lack of experience and tells me they will struggle with real-world ambiguity. A lack of depth on evals is one of the most common reasons AI product management candidates fail the interview process.

Businesses must prove they can create value with AI and agents within the next 3 quarters to keep investors happy. The bridge has two parts, and the first one is relatively straightforward. The reason most evals seem overwhelming to build is that the agent’s purview is too broad.

That increases the eval’s scope and makes a holistic evaluation difficult. It leads to questions like, “How do I anticipate everything a user can do with this agent?” Narrowing the scope answers this question.

A broad agent purview also complicates resolving the issues you find during the eval process. Fixing one part of the agent’s functionality breaks other things the agent does. It can feel like agent improvement is a constant balancing act that’s impossible to get right.

Upfront Intent Detection & Routing

Part one of the bridge is intent detection, and it is also the start of building the second part. As soon as I explain intent detection to clients and technical teams, everything else is a lot clearer. It helps determine how the business chooses what agents should do and how we define the agent’s scope narrowly enough to evaluate it holistically.

Bonus: It creates a clear connection back to value creation, so ROI can be estimated upfront. It creates an alignment between strategic goals and the opportunities that get added to the roadmap. If the business is struggling here, the solution is in an updated opportunity discovery process.

Don’t think of intent detection as limiting what the agent does, and by extension, the value it creates. Think of it as ensuring the agent never tries to do too much and does work that falls outside of your evals.

It starts with the action surface, where users or customers go to start their workflow(s). Slack, Google Search, Outlook, and increasingly Claude or ChatGPT are good examples of an action surface. The places people begin their workflows are natural places for them to express intent.

There are two types of intent detection and agents that support them: Reactive and Proactive.

Read more

I will use the AGI and SaaSpocalypse debates to explain how first principles of engineering can reveal the direction of business. I love making calls when everyone is leaning in the opposite direction. Salesforce’s quarterly earnings and partnership with Anthropic have many calling to cancel the SaaSpocalypse. They are wrong, and I’ll explain why using engineering first principles.

But let’s start with a simpler example. In the last couple of weeks, the AGI question has come back into public discourse. So, have we achieved AGI, and if not, how close are we?

This Is Not The AGI You’re Looking For

SLMs and LLMs give any technology the ability to read and write. Agents can use AI’s reading comprehension to interact with the business and its customers in much the same way people can. That is a first-principles change to how technology is configured, but it is not a path to AGI. LLMs are just one step when you break them down to first principles.

It’s like humanoid robots in factories. They are not capable of doing everything that people can do, but they are configured similarly enough to a human that they do things the same way people do. They can interact with the physical world the same way people do. A human form is not a path to AGI. The large vision models that allow a human form to interact with its environment aren’t either.

The challenge is that they all look like AGI when you lose sight of their first principles. LLMs talk like a person. Humanoid robots look like a person. Large vision models, humanoid robots, and LLMs on the same platform will act even more like a person. However, they won’t really be AGI because none of those three gives it the logic and reasoning capabilities that humans have.

Logic Is Missing: The AI Dunning-Kruger Effect

Read more