From code completions to agents: what AI adoption is really worth
Across engaged Copilot Business and Enterprise users, 62.1% are still primarily engaging with code completions, rather than agentic capabilities.
If you measure adoption by engagement alone, that looks like success: they're using AI. But it's also a signal that most of your developers are stalled at the most basic form of AI use, and that the real value is still on the table. Reporting adoption and driving it forward are two different jobs. The second means looking past whether developers use AI to how.
Adoption is a curve
A common mistake we see is treating adoption as all or nothing: a developer either "uses AI" or doesn't. In practice, that binary erases the nuance that matters. Developers engage with AI across a wide spectrum, and the difference between a developer accepting code completions in their IDE and a developer orchestrating multiple agents is meaningful. Collapse them into one active user count and a team steadily moving from experimentation to AI-native engineering looks identical to a team that's stalled.
Instead of a single active-user figure, the GitHub Copilot impact dashboard groups developers into AI adoption cohorts based on how they've actually used Copilot over the past 28 days. These cohorts span four phases: no cohort, code first, agent first, and multi-agent. The point isn’t the classification logic itself; AI is evolving quickly towards more multi-threaded, agentic experiences, and the logic is built to evolve with it. What endures is the framework: a consistent way to understand where developers are across a spectrum of adoption and how that's changing over time.
Cohort phases:
Phase 0 — No cohort: User did not meet the engagement criteria for any phase.
Phase 1 — Code first: User engaged with code completion and/or IDE agent mode.
Phase 2 — Agent first: User engaged with a single GitHub-based agent surface (i.e., Copilot cloud agent, Copilot code review, or Copilot CLI).
Phase 3 — Multi-agent: User engaged with two or more GitHub-based agent surfaces, or with the new GitHub Copilot app.
That framing is what makes the data actionable. You can watch cohort mix shift month over month to see whether developers are graduating into more advanced phases. And if they’re not, the dashboard’s recommendations offer a path to enable more agentic work. You can also assess adoption at the team level to find exactly where momentum is building and where it’s stuck to target enablement investment where it’s most needed.
Copilot usage metrics transformed our rollout from guesswork into a data-driven initiative. By providing a single source of truth about who is using Copilot, how usage is evolving, and where adoption is increasing, we can confidently focus our enablement efforts and scale effectively.
Akihito Mizoe, Director, Application Services Division, Hitachi, Ltd.
Proxies for ROI beat a single number
While adoption matters, at some point every leader has to justify the investment, and that pressure naturally pulls toward a single, tidy ROI figure. Rolled up far enough, ROI hides where value is coming from and whether it will hold.
We take a different approach. Rather than manufacturing a precise-looking dollar amount, we focus on measurable improvements that act as leading signals for ROI. Two of the most telling are how much more work is getting done and how quickly. The dashboard tracks both: pull requests merged and time-to-merge.
For a quick read on Copilot’s impact, the dashboard compares these metrics between passive and engaged developers. For a more precise view, those same metrics are mapped to adoption cohorts, showing the lift between each stage and the additional impact that can be unlocked by advancing developers to the next.
When we evaluated Copilot usage at thousands of enterprises, we found that phase progressions generally yielded higher PR throughput.
Starting phase | Ending phase | Percent of accounts where average throughput was higher for users in ending phase vs. starting phase |
Phase 1 (Code first) | Phase 2 (Agent first) | 77.6% |
Phase 2 (Agent first) | Phase 3 (Multi-agent) | 77.2% |
Phase 1 (Code first) | Phase 3 (Multi-agent) | 93.3% |
And when we look at the magnitude, we can see just how much value could be realized when developers progress between stages.
Merged PR gains on average:
From phase 1 to phase 2: 78% more PRs
From phase 2 to phase 3: 41% more PRs
From phase 1 to phase 3: 151% more PRs
Measure what matters
A single adoption number tells you how many developers touched AI. It can't distinguish a completion-accepter from an agent-orchestrator—and that's a significant gap. Our dashboard is built to show you the shape of your adoption and how that maps back to impact, not just its size. Because you can't move a curve you can't see.
