AI in engineering: Q2 2026 benchmarks & research readout

Engineering Enablement by DX

0:00

-38:51

AI in engineering: Q2 2026 benchmarks & research readout

Brian Houck and Justin Reock unpack DX’s latest AI Impact Report, exploring where AI is improving engineering velocity and where those gains are failing to translate into better outcomes.

Brian Houck

Aug 14, 2026

Listen and watch now on YouTube, Apple, and Spotify.

AI adoption among software developers is approaching 100%, AI-authored code now makes up more than half of merged code, and developers report saving more time with AI every quarter. But those gains aren’t translating evenly into better outcomes.

In this episode of Engineering Enablement, I sit down with Justin Reock, Deputy CTO at DX, to unpack findings from our latest AI Impact Report. We explore where AI is improving engineering velocity and developer experience, where concerns are emerging around PR size, change confidence, and failure rates, and why rising AI spend has yet to produce a comparable increase in innovation.

We also discuss how AI is changing the meaning of code maintainability and where developers’ AI-driven time savings may actually be going.

Some takeaways:

AI adoption is no longer the interesting question

AI adoption is approaching 100%, making adoption itself a less useful measure of success. DX sees 95% adoption through telemetry, and even developers outside that group are merging AI-generated code into production.
The more important question is how effectively developers are using AI. As adoption becomes nearly universal, organizations need to look beyond usage and understand how sophisticated their AI usage is and whether or not it is producing meaningful outcomes.

AI is increasing engineering velocity, but the gains need context

Developers report saving more time with AI each quarter, while PR throughput and deployment frequency are also increasing. Recent causal research gives additional evidence that AI is contributing to the increase in throughput.
More than half of merged code is now AI-authored. But metrics like time saved, code generated, and PR throughput don’t tell us on their own whether teams are creating more customer value.

Larger PRs create risks that extend beyond code review

Average PR size has increased from 42 to 72 lines of code, raising concerns about quality and code understanding. Larger PRs are harder to review and revert, more likely to be rubber-stamped, and can move more slowly through the development system.
Code bloat can also create a vicious cycle of rising AI costs. As agents generate larger codebases, future agents have more code to read into context, consuming additional tokens repeatedly rather than creating only a one-time cost.

Developers can change code more easily while trusting it less

Code maintainability is improving while change confidence is declining, even though those measures have historically been closely correlated. AI may make code easier to understand and modify while leaving developers less confident that their changes won’t break something in production.
AI may be changing what “maintainable” code actually means. If developers increasingly rely on agents to understand and modify code, traditional measures of maintainability and ownership may no longer capture the same things they once did.

AI is amplifying differences between engineering organizations

Change failure rates are becoming more volatile rather than moving uniformly in one direction. Some organizations continue to perform well while others are seeing higher failure rates, suggesting that AI can amplify existing strengths and weaknesses.
Developer experience is also showing concerning signals despite clear improvements in areas like documentation. DX’s Developer Experience Index has fallen about 2% over two quarters, with declines in drivers such as incremental delivery outweighing some of AI’s benefits.

AI spending is growing much faster than measurable innovation

AI spend has risen dramatically, reaching roughly 28 times year-ago levels among the largest companies in the dataset. At the same time, developer ramp-up has improved, but the innovation ratio has increased by only about one percentage point.
Time saved by AI can easily be absorbed by existing organizational friction. Meeting-heavy days, interruptions, build and test wait times, developer environment toil, and review delays remain major sources of lost time. In one real-world case study, reducing unnecessary meetings produced roughly twice the PR-throughput gains achieved through AI.
Ultimately, AI efficiency only matters if it translates into customer value. Removing friction and increasing throughput are useful, but the larger question is whether those gains allow organizations to ship more valuable software.