Ask an engineering lead how the AI coding rollout is going and you'll usually get a confident answer. Ask the people reviewing the pull requests and the answer gets more careful.
That's now more than a hunch. Three surveys published in the first half of November 2026 asked leaders and staff about the same rollouts, and the leaders came out more positive every time.
This post maps those findings onto engineering teams and gives you a small, copyable check to run in your next retro.
The numbers that matter for dev teams
SlashData, "Faster code, heavier checks" (12 Nov 2026). 32nd Developer survey, n=11,099 professional developers in 122 countries.
- 67% of engineering leaders say their team has enough time to evaluate AI-assisted work.
- 60% of software and architecture professionals say the same. They're the ones doing the evaluating.
- 76% say AI tools are leading managers to expect more output from the same team size.
- 81% say AI reduces repetitive work but increases the need for review and validation.
PA Consulting (9 Nov 2026). 2,000 UK employees across sectors.
- 70% of senior leaders say AI helps them achieve things they couldn't before, against 38% of entry-level and junior staff.
- 60% say AI helps them work faster, and 37% say they then spend as long reviewing and refining the output as doing it themselves.
- 21% are unsure where accountability lies for AI-shaped work.
Brunswick Review, "The AI Disconnect" (5 Nov 2026). 2,000+ US employees, 65 senior leaders.
- 92% of leaders think AI will create new opportunities; 53% of employees agree.
- Frequent AI use peaks mid-career: 52% under five years' experience, 67% at five to 14 years, 48% at 15+.
Caveat: three different surveys, different countries, different questions. Compare direction, not the numbers. Brunswick's leader sample is small (65).
What this looks like in an engineering team
The following is my inference, not a finding from any of the surveys.
Leads see throughput. Reviewers see the diff. A lead's view of AI comes from dashboards: PRs merged, cycle time, licence usage. A reviewer's view comes from the PR that looked fine and wasn't. Both are real. They just measure different things, and only one of them usually reaches the planning meeting.
Capacity plans inherit the optimistic number. If the person estimating thinks review time is covered and the reviewers don't, the shortfall turns up as slower reviews, rubber-stamped reviews, or weekend work.
Juniors and veterans need different help. Brunswick found early-career staff have the fluency but not the context to know where to apply AI, and long-serving staff report less excitement and support. On a dev team that might mean juniors accepting suggestions they can't judge yet, and seniors quietly opting out. One "AI enablement" session won't fix both.
Accountability gets fuzzy at the merge button. If people are unsure who owns AI-shaped work, "the agent wrote it" starts to feel like a reason.
The artefact: a two-sided AI check
Run this anonymously. Ask leads (EMs, tech leads, anyone setting targets) and ICs (the people writing and reviewing code) the identical statements. Score % agree for each group. Track the gap, not the average.
The rows below are example values for illustration only. Replace them with your own team's answers.
## Two-sided AI check: <team>, <date>
Scale: % who agree or strongly agree. Same wording for both groups. Anonymous.
| # | Statement | Leads % | ICs % | Gap (pts) | Owner | Next action |
|---|---|---|---|---|---|---|
| 1 | I can explain what our AI plan means for my work | 90 | 55 | 35 | EM | Example: walk through the plan with real PRs from this repo |
| 2 | AI makes our code better, not only faster | 80 | 50 | 30 | Tech lead | Example: track rework and revert rate on AI-assisted PRs for 4 weeks |
| 3 | We have enough time to review AI-assisted changes | 70 | 45 | 25 | EM | Example: add review hours to sprint capacity before raising targets |
| 4 | I know who is accountable for AI-assisted code that ships | 85 | 60 | 25 | Tech lead | Example: write one line in CONTRIBUTING.md naming the PR author as owner |
| 5 | Juniors get enough support to judge AI suggestions | 75 | 40 | 35 | Senior devs | Example: pair each junior with someone a few years ahead for AI-heavy tickets |
Rules:
- Gap of 20+ points on any row: discuss in the next retro.
- Act on the biggest gap first, one action per cycle.
- Re-run the same statements next quarter. Do not reword them.
A few notes on using it:
- Keep the groups separate. If you blend the scores, the leads disappear into the total and the gap goes with them.
- Keep it small. Five statements, two minutes. People answer honestly when it's quick and anonymous.
- Don't argue with the IC number. The point is to find where the plan and the work disagree. The reviewer's answer is data about the work.
Limits
All three sources are surveys of self-reported views, not measured outcomes. Two come from consultancies and one from a firm that sells custom research. PA's figures are from its press release. None of them studies engineering teams specifically except SlashData. The check above is my suggestion for applying the pattern, not something the surveys tested.
Sources
- SlashData, "Faster code, heavier checks", 12 November 2026: https://www.slashdata.co/research/faster-code-heavier-checks
- PA Consulting, "AI risks widening the workplace divide, new report finds", 9 November 2026: https://www.paconsulting.com/newsroom/ai-risks-widening-the-workplace-divide-new-report-finds-9-november-2026
- Brunswick Review, "The AI Disconnect", 5 November 2026: https://review.brunswickgroup.com/article/ai-workplace-disconnect/
If you ran row 3 with your team this week, which way do you think the gap would go, and by how much?
Top comments (0)