Everything You Can Count Is Up. Everything That Matters Is Flat.

Everything easy to count is soaring but is that the whole story?

DX just published its Q2 2026 State of AI Impact in Engineering report, measuring 500+ organizations. I lead analytics and insights. And I read this report the way I read everything: as a measurement story. On that level, it's one of the most quietly damning things I've seen this year.

Here's the shape of it.

Everything easy to count is soaring. AI adoption is above 90%, so ubiquitous the researchers say "are we using AI" is no longer even a useful question. More than half of all code, 52.7%, is now AI-authored, up from 19% three quarters ago. Throughput is up 37%. Deploys are up double digits. AI spend has gone from roughly $1,500 to $44,000 per quarter for a typical org, and from $9 to $191 per developer. Every number a dashboard loves is pointing up and to the right.

Now the numbers nobody puts on the slide.

Developers' perceived rate of delivery has been flat for a full year. The pipeline got measurably faster and the humans in it don't feel it. The Developer Experience Index actually fell, and the report estimates each single point of that drop costs about ten hours per engineer per year to friction. Change confidence, whether engineers trust their own changes won't break things, dropped 6.1% even as code became easier to read. And the one that stops me cold: the innovation ratio, the share of time spent building new things versus maintaining old ones, has been essentially flat at around 58% for the entire year. All that saved time, four to six hours a week per developer, and it is not showing up as more innovation. Only 6% of executives in a related study said they could point to clear organization-wide AI ROI.

Read those two lists together and the story is unmistakable. We got much better at producing output and no better at producing value. The activity metrics and the outcome metrics have completely decoupled, and almost everyone is watching the activity metrics.

This is the exact trap I keep writing about. Adoption was never the finish line. Output is not the same as impact. And when you can measure the easy thing, you stop asking about the hard one.

Two findings hit especially close to my world.

First, the regulated-industry twist. The sectors moving slowest on throughput, Financial Services and traditional industries, report the highest confidence in their software quality. Industries showing significant speed are also scoring lower on perceived quality. Speed and quality did not move together, and the report ties the difference to something I believe deeply: the slower sectors ran more structured, more governed rollouts. Governance looked like the drag. It turned out to be the moat.

Second, the human one. The report finds engineers pull back on AI-driven speed specifically when a change is high-risk or hard to undo. That's not fear of the tool. That's accountability. People feel responsible for what ships under their name, so they slow down exactly where the stakes are real. I find that genuinely reassuring. It's the same instinct behind a rule I hold: you own every word you send, no matter what helped you write it.

The report's own bottom line is one every leader should tape to the wall: pair every speed or spend metric with a quality and experience counterweight. Track change confidence next to throughput. Track innovation ratio next to time saved. Otherwise you are just shipping faster and calling motion progress.

None of this is anti-AI. The gains are real. The onboarding wins alone are remarkable. It's a caution about what we choose to measure, because in the AI era the easy metric and the true one have never been further apart.

Credit to the DX Research team for the report. For the leaders here: if your AI dashboard shows adoption and output but not confidence, quality, and innovation, you're watching the half of the picture designed to look good. What's your counterweight metric?