Sigao
A bright multi-lane highway funneling into a single open toll gate under a clear morning sky, cars queued in the sunlight.
All posts
Perspective·6 min read

Your developers are faster. Your releases aren't.

Every metric your developers can see says AI is working. Several metrics your board can see say it isn't. Both sets of numbers are telling the truth, and the gap between them is where AI budgets go to die.

By Chris Sims

There's a meeting happening this quarter in a thousand engineering organizations, and it runs the same way every time. Engineering presents the AI numbers: seats active, acceptance rates up, developers reporting real gains. Everyone in the room can feel the tools working. Then the CFO asks the only question a CFO ever needs to ask: did anything ship sooner? And the room goes quiet, because lead time hasn't moved, the release calendar hasn't moved, and the roadmap is hitting the same dates it hit last year.

Here's the part almost nobody says out loud: both halves of that meeting are right.

The individual-layer evidence is real. Adoption hit 84% in the 2025 Stack Overflow survey and 90% in DORA's 2025 data, with most developers reporting personal productivity gains, and in the most-cited controlled experiment, from 2023, developers using GitHub Copilot completed an isolated, well-specified task 55.8% faster than the control group. Nobody is imagining the speed. And yet the system-level ledger keeps running behind. DORA's 2024 report found a 25% increase in AI adoption associated with an estimated 1.5% drop in delivery throughput and a 7.2% drop in delivery stability; its 2025 report saw throughput finally edge modestly positive as teams adapted, while instability persisted. MIT's 2025 State of AI in Business study put roughly 95% of enterprise GenAI pilots at no measurable P&L impact.

Two honest dashboards, from independent public studies. Task-level gains are consistently real; system-level gains are consistently rare. Any explanation of your AI results has to account for both columns at once.

The strangest result of the period sits in the gap between those two dashboards. In mid-2025, the research group METR ran a randomized controlled trial with experienced open-source maintainers working in their own mature codebases: 16 developers, 246 real tasks, AI allowed on a random half. With AI, they were 19% slower. Before the study, they forecast AI would speed them up 24%. Afterward, having just lived the slowdown, they estimated it had sped them up 20%. That's a 39-point gap between felt speed and measured speed, in the same people, on the same tasks. And treat the slowdown itself as dated, because it is: METR's 2026 follow-up found the measured effect shrinking toward zero on newer tools, with confidence intervals that now include a real speedup, and the randomized design itself became hard to sustain because developers increasingly decline to work without AI, even when a study pays them to. What hasn't aged is the perception gap. Perception is not measurement, and if your AI business case rests on developer self-report, that gap is the figure that should worry you.

Where the saved time actually goes

If individuals are producing more and the system isn't shipping more, the surplus has to be piling up somewhere. It is, and DORA's researchers found the mechanism hiding in plain sight: AI makes it easy to write more code per change, batch sizes grow, and larger changes are riskier and harder to review. Their 2025 report compressed the whole pattern into a sentence that belongs on a wall in every engineering org: individual productivity boosts are frequently lost to downstream disorder. The same report calls AI an amplifier, magnifying an organization's existing strengths and weaknesses, which is a research finding and a warning label in one.

You can watch it happen in your own tooling. Time-to-first-commit falls. Open pull requests climb. Review queues age. The same four senior engineers who were the review path before AI are still the review path, except the queue feeding them now moves three times faster. And the code arriving in that queue is harder to review, not easier: in the 2025 Stack Overflow survey, 66% of developers said their biggest AI frustration is output that is almost right, and 45% reported that debugging AI-generated code takes longer than writing it themselves. Plausible-looking defects take longer to catch than obvious ones.

Notice what didn't happen in that story. Nobody made a mistake. The developers used the tools well, and the tools did what they promised. Every local decision was sound, and the system still got slower, because the effort wasn't saved. It was shifted downstream, onto the people and stages least equipped to absorb it.

Five signals it's happening to you

  • PR queue age rising while time-to-open-a-PR falls. The clearest single signature.
  • Diff sizes creeping up. AI makes 800-line changes cheap to produce and expensive to review.
  • Review depth thinning: approvals arriving faster than anyone could have read the change.
  • QA reopen rates and change-failure upticks two to six weeks after AI usage jumps.
  • Rework loops widening: tickets bouncing between done and not-done, "almost right" both times.

Every one of these is visible in tooling you already run. None of them appears on a vendor's dashboard.

The diagnosis that survives the data

The popular explanation is that the tools aren't good enough yet. That story doesn't survive contact with the evidence, because the same tools are producing real system-level gains in a minority of organizations. What separates them isn't the model. It's where the acceleration got aimed: at a step, or at the system. Software delivery is a flow of work through stations and queues, and its speed is set by its slowest point. Accelerate any other point and you don't get faster delivery. You get bigger piles. Code generation, in most organizations, was not the slowest point.

The stakes of leaving this unresolved are larger than a disappointing line item. Leaders who promised AI-driven acceleration are explaining flat delivery metrics to boards that read the same headlines they do, and trust erodes in both directions: executives stop believing engineering's productivity claims, and developers, 46% of whom told the 2025 Stack Overflow survey they distrust AI output accuracy, stop believing leadership's next initiative. The organizations that got burned once are the least likely to do the systemic work the second attempt requires.

The Sigao take

The AI productivity paradox is not evidence that AI doesn't work. It's evidence that most organizations pointed it somewhere it couldn't matter. The fix isn't a better model, and it isn't more seats; it's operational: map your delivery stream, find the constraint, aim the acceleration there, and manage the flow so the gains reach the calendar. We wrote the full playbook, including the 90-day plan and the metrics a board will actually bank, in our whitepaper The Bottleneck Isn't the Model. It's free, and it takes the argument this post opens to about thirty minutes of depth. If you want the dollar version of your own gap first, our AI value calculator takes three minutes.

Sources

Keep going

Where to go from here.

A straight read. We’ll tell you where your delivery stands and whether we can help.

Book a call

Newer post

Back to all posts

Older post

Your tech debt is a loan. Do you know the interest rate?