The tools have changed. The bottleneck has moved.

There is no shortage of AI productivity statistics. What is harder to find is an honest account of what they actually mean for an engineering team trying to get work done.

Here is what we have found worth paying attention to — with the caveats included.

Adoption is no longer the question

In early 2026, The Pragmatic Engineer newsletter published findings from a survey of approximately 900 software engineers and engineering leaders (median experience 11–15 years, primarily Europe and the US). The headline numbers were striking: approximately 95% of respondents reported using AI tools at least weekly, and approximately 75% said they use AI for at least half of their engineering work.

These are not early adopters. This is the baseline.

The same survey found Claude Code — released in May 2025 — had become the most-used AI coding tool in approximately eight months, overtaking tools that had years of runway. Approximately 70% of respondents were using between two and four AI tools simultaneously.

“The question has shifted from ‘should our team use AI?’ to ‘how well is our team using it?’ Those are very different questions, and the second one is harder to answer.”

For context: the Stack Overflow 2025 Developer Survey, covering 49,000+ developers globally, reported approximately 84% were using or planning to use AI tools. Survey methodologies and populations differ, so the specific percentages across studies vary — but the directional picture is consistent.

The point is not the specific percentage. The point is that the question has shifted from “should our team use AI?” to “how well is our team using it?” Those are very different questions, and the second one is harder to answer.

Where the productivity gains actually live

The honest picture is more nuanced than the headlines suggest.

Addy Osmani, writing in August 2025, summarised the controlled study evidence: productivity improvements of roughly 20–30% on specific coding tasks appear consistently across several studies. What does not appear consistently is that translating into equivalent team-level output improvements. Writing code faster is one constraint. Review, testing, architecture decisions, coordination, and deployment are the others. Removing one bottleneck tends to surface the next one.

A 2025 study by Faros AI — tracking telemetry from over 1,200 engineering teams and more than 10,000 developers — found that teams with high AI adoption merged approximately 98% more pull requests, but saw PR review time increase by approximately 91%. More output hitting a fixed review ceiling creates a new bottleneck downstream.

“AI accelerates production; the bottleneck moves to review and validation. This is one of the most consistent findings across the field reports and research we have read.”

This pattern — AI accelerates production, bottleneck moves to review and validation — is one of the most consistent findings across the field reports and research we have read. It has a practical implication for how teams think about hiring.

What this means for how you staff

If your team’s bottleneck has moved from writing code to reviewing, validating, and integrating it, the value of a contractor who uses AI fluently changes.

A developer who uses AI well produces more output — but also tends to produce better-documented, more consistent output, because AI assistance rewards clear structure and penalises ambiguity. Their code tends to come with clearer variable names, better inline comments, and more explicit handling of edge cases — not because AI is perfect, but because prompting AI well requires being explicit about what you want, and that discipline tends to carry through.

The developers who create the most pressure on review queues with AI-assisted code are generally the ones who have not yet developed the verification habits to match. A developer who has been using these tools long enough to understand their failure modes spends more time reviewing their own output before it reaches your team — not less.

This is the gap that matters. Not “does this contractor use AI?” but “have they developed the judgement that makes AI output actually useful at team scale?”

The seniority factor

One finding from The Pragmatic Engineer’s 2026 survey is worth flagging for teams making staffing decisions: staff-level and senior engineers were the heaviest users of AI agents — approximately 63% used them regularly, compared to approximately 50% of regular engineers. Senior leaders were especially enthusiastic about Claude Code specifically.

This is counterintuitive if you assumed junior developers would be the fastest adopters (fewer established habits to change). The data suggests the opposite: experienced engineers, who have a clear model of what good output looks like, are better at getting useful output from AI tools — and better at catching when something is wrong.

It is also consistent with what we see in practice. The developers who use AI most effectively are not the ones using it to avoid thinking. They are the ones using it to do more of the thinking that matters.