## Why “adoption” is the wrong number
Survey after survey reports adoption near 90%, and each one reports a different second number that matters more.
The 2025 State of AI-assisted Software Development report by DORA, Google Cloud’s DevOps Research and Assessment program, based on responses from nearly 5,000 technology professionals, found that 90% report using AI at work and more than 80% perceive that it increased their productivity. But only 65% rely on AI “at least a moderate amount”, and 60% use it about half the time or more when they hit a problem. Nine in ten touch it. Six in ten lean on it.
McKinsey’s August 2026 global survey found the same gap: 80% of respondents say AI improved their individual productivity while only 37% of the same respondents’ organizations report any effect on earnings.
DX, which measures AI use at more than 500 organizations each quarter, reported on April 1, 2026 that the average share of merged code written by AI moved from 22% to 27.4% between two quarters. Among daily users the figure was 30.8%. DX’s own reading: the tooling improved faster than teams adapted to it.
So what happens in that gap?
## The Individual-to-Team Gap
Here’s what I’ve seen, and what the data confirms: AI makes individual engineers faster. It does not make teams faster unless you change how the team works.
One engineer on your team starts using an AI coding agent. They write more code, generate more tests, prototype features in a day that used to take three. On paper, your team’s velocity just went up.
In practice, the rest of the team has to catch up. Five new pull requests appear over a weekend. The code review queue balloons. The senior engineers who would normally review PRs now spend 80% of their time reviewing AI-assisted changes that are 40% larger than before.
Faros AI’s 2026 engineering report, built on two years of telemetry from 22,000 repositories, found that under high AI adoption, PR size went up 51% and bugs per PR went up 28%. Median time in review went up 441.5%. Average time in review went up 199.6%. Pull requests merged with no review at all went up 31.3%.
One fast engineer can slow down a whole team.
This is the individual-to-team gap. It’s where most AI adoption initiatives die — not because the tools don’t work, but because nobody expected the consequences of making one person faster while everyone else stayed the same speed.
## The Review Bottleneck Moves, It Doesn’t Disappear
Here’s what happens when you give every engineer an AI coding agent and change nothing else: the bottleneck doesn’t disappear. It moves.
Code gets written faster. Reviewing that code gets harder. Bigger diffs, more generated tests, more edge cases the model introduced without testing for. Same number of reviewers. Same number of senior engineers, who are now the bottleneck instead of the individual contributors.
A 2026 BCG survey of 5,000+ engineering professionals found 97% adoption across the sample, but also that 48% of engineers report AI usage is driven by individuals experimenting independently, and only 6% work in an environment with a dedicated enablement or platform group. 46% say the impact is not yet formally measured or discussed within their organisation.
The result is what one source called “the senior engineer tax”: reviewers can’t keep pace, PRs pile up, and the engineers who could actually validate the AI output are the same ones being asked to validate everything.
Patrick Dawkins, who built an internal tool that reviews merge requests before a human sees them, put it plainly: “Running a single agent is manageable. Once it can run unsupervised for a while, boredom sets in, and you start a second one. Then a third. You’re thinking about three different things at the same time.”
Individual productivity and team velocity are not the same metric. Optimizing for one quietly taxes the other.
## Nobody Redesigned the Workflow
Handing engineers a Copilot license and calling it AI adoption is the default move. It gives you faster code generation inside an unchanged delivery lifecycle:
– Requirements still arrive late and ambiguous.
– PRs still sit in a review queue for two days.
– Test plans are still written after the code, not alongside it.
When we started working with Xceptor, a financial services data automation platform with a 49-person engineering team, they already had AI tools deployed. What they didn’t have was any change to how work moved between roles. Requirements were still handed off the same way. Testing was still sequenced the same way. Tools were faster, but the system wasn’t.
Before we touched a single tool configuration, we mapped the actual delivery lifecycle — where stories stalled, where rework originated, which handoffs created delays. That mapping, not the tooling, determined the sequence of everything that followed. AI was integrated phase by phase, starting where it would compound across downstream work rather than where it was easiest to demo.
If you can’t draw the workflow that changes, you don’t have an AI strategy. You have a procurement decision.
This is the thing that nobody wants to hear: the tool wasn’t the problem. The workflow was. AI doesn’t accelerate a broken process. It accelerates it faster, which means you hit the broken part faster, with more damage.
## The Five Stall Patterns
From my reading of what’s happening across engineering orgs right now, AI adoption stalls for five reasons — none of them about the tool quality:
**1. The verification tax.** Engineers who distrust AI output — 46% in Stack Overflow’s survey, ~30% in DORA’s — don’t necessarily stop using the tools. They re-verify everything, which converts promised time savings into review effort. When verification costs more than the generation saved, engineers rationally conclude the tool isn’t worth it for their work, and usage decays to trivial tasks.
**2. Skill-relevance anxiety.** In a 2026 Augment Code survey of engineering leaders, 63% reported engineers voicing concerns about skill relevance — rising to 89% at organizations with 201–1,000 engineers. The anxiety concentrates among senior engineers, who have the most identity invested in craft and the most influence over everyone else’s behavior. A rollout framed, even implicitly, as headcount replacement will be quietly killed by exactly the people whose endorsement it needs.
**3. Workflow misalignment.** AI tools that generate code faster than the organization can review, test, and release it don’t accelerate delivery — they relocate the queue. More pull requests, larger diffs, and review comments like “why did it choose this pattern?” raise the cognitive load on reviewers. If the tool doesn’t fit how code actually moves from keyboard to production, engineers experience adoption as added coordination work, and drop it.
**4. Silent sponsorship.** BCG’s AI at Work 2026 survey found 74% frontline AI use but only 33% of respondents saying leadership communicates clearly about AI. Engineers in that gap are left to guess what is allowed, what is encouraged, and what happens to people whose roles change. DORA’s 2025 research names a clear, communicated AI stance as one of the seven capabilities that amplify AI’s benefits — and its absence produces the two failure modes every enterprise recognizes: engineers who freeze (use nothing, wait for policy) and engineers who freelance (use unapproved tools on company code).
**5. Nobody owns the change.** AI tooling gets purchased by engineering leadership. Enablement gets delegated to a platform team or an internal AI committee. But nobody owns the end-to-end change: process redesign, role-specific playbooks, measurement framework, and accountability for whether delivery metrics actually improve.
## What Actually Works
This isn’t a “stop using AI” piece. The tools work. The data on individual productivity gains is real and substantial. The problem is that teams are trying to scale individual gains with team-level processes that were never designed for them.
Here’s what I’ve seen work, and what the research supports:
**Start with the workflow, not the tool.** Map how work moves through your team. Find the handoff that creates the most rework, the longest delay, the most context switching. Integrate AI there first, not where it’s easiest to demo.
**Measure usage events, not licenses.** 200 Copilot licenses with 15% weekly active usage is a people problem dressed up as a technology investment. Track how many times AI contributed to a shipped artifact this week. Not “how many seats” but “how many times.”
**Redesign the review process.** If AI generates code faster than you can review it, your review process is your bottleneck, not your coding process. That means more reviewers, rotation systems, or constraining where AI output gets heavy verification. The fix is not exhortation — it’s constraining.
**Create a visible early win.** At Xceptor, the design team hit 83% faster prototyping, well above their 50–60% target, and that result became the internal proof point that pulled other teams in. Eighty percent adoption came from pull, not push. Mandated adoption produces compliance — engineers install the tool, use it when observed, and revert to their existing workflow when nobody’s looking.
**Assign an owner.** Someone needs to own the end-to-end change: process redesign, role-specific playbooks, measurement framework, and accountability for whether delivery metrics actually improve. Not a committee. Not a platform team. A person.
## What I’d Tell Another Manager or Engineer
If you’re trying to roll out AI tools to your team and it’s not working — if individuals are faster but the team isn’t — it’s probably not the tools. It’s probably that nobody redesigned the workflow.
Start by mapping how work actually moves through your team. Find the bottleneck. Redesign that part. Then integrate AI there.
The tools are ready. The question is whether your team’s processes are.
—
*Sources: DORA 2025 State of AI-assisted Software Development (nearly 5,000 professionals); McKinsey August 2026 global survey; DX quarterly AI measurement (500+ organizations, Q1 2026); Faros AI 2026 engineering report (22,000 repositories); BCG AI at Work 2026 (5,000+ engineering professionals); Stack Overflow 2026 developer survey; Augment Code 2026 engineering leader survey; Snowman Labs “AI Adoption for Engineering Teams: From Licenses to Lift”; Forté Group “Why Most AI Tools Don’t Work at Team Level”; Upsun “AI adoption mistakes in engineering teams.”*