Four quarters of AI in engineering, in four numbers.

The latest industry data on AI in engineering tells a story most leaders already feel. Code is being produced faster than ever. Review queues, quality, and the roadmap have barely moved. Here is what the Q3 2026 numbers show, what we think they mean, and what to measure next.

DX recently published its State of AI Impact in Engineering: Q3 Report, covering July to September 2026 across more than 500 engineering organizations. It combines system telemetry from source control, CI/CD and AI coding tools with developer survey data. It is one of the larger public datasets on AI in software delivery right now, so it is worth reading closely, including where the authors are careful to say what the data does not prove.

All figures below come from that report unless marked as a Waydev calculation. We have grouped them into five findings and added our own read on each.

Throughput is still climbing, and just started to ease

Median weekly throughput (DX uses a complexity-weighted measure it calls TrueThroughput) rose from 1.43 PRs per developer in Q4 2025 to 2.51 in Q3 2026, a 75.5% increase. Quarter-over-quarter growth held near 21% for three quarters, then dipped to 19.0% in Q3. That is the first sign of easing, and the report is clear it is too early to call a trend.

Review is where the time goes now

This is the most important chart in the report. Organizations were grouped by how much of their code is AI-authored. The more AI code, the longer a PR waits for its first human review: 40 minutes at the bottom tier, 82 minutes at the top, a 105% increase.

At the same time, median PR size more than doubled, from 46 lines of code in Q4 2025 to 102 in Q3 2026. That is up 122% year over year and 38% in a single quarter.

The report is careful here, and we think rightly so. The data supports a capacity explanation: more and bigger changesets chasing the same review bandwidth. Two other ideas, that AI changes take more effort per review, or that reviewers hesitate on AI-heavy PRs, are flagged as open hypotheses, not findings.

AI helps with tasks far more than with the system

Developer survey data shows big gains on bounded, diagnostic work with a short feedback loop, and almost nothing on work that requires holding the whole system in view. Overall perceived ease of delivery improved just 3% from Q2 to Q3, and the Developer Experience Index moved only about 1% at the median.

Ranked from largest to smallest gain, the factors sort almost perfectly from local to systemic. In plain terms, AI is helping developers get through existing friction faster. It is not yet removing the friction.

Quality is holding, with one signal to watch

Shipping far more without breaking more is a real achievement. Failed deployment recovery time stayed at roughly 30 minutes at the median every quarter. Change failure rate was also flat, with one exception: the median quarter-over-quarter change moved from 0% to about +0.5%. The report says this is not yet statistically distinguishable from normal variation, and names it as the metric to watch most closely next quarter. There is no evidence of high AI volume causing large outages. If there is an effect, it looks like a slow accumulation of small failures spread across many teams.

The saved time is not reaching the roadmap

Developers report saving up to 6.1 hours a week with AI. Yet the median innovation ratio, the share of time spent on new features rather than maintenance and fixes, moved only from 57% to 59% across four quarters. The report found innovation clustered between 50% and 65% regardless of how much code was AI-authored.

Session telemetry explains part of why. Looking at what AI sessions were actually used for, roughly a third went to maintenance and almost a quarter to fixes. Only about 11% went to building new features, and code review was one of the smallest categories of all.

Spend is growing much faster than delivery

Total AI spend across the sample grew from about $10M in Q4 2025 to $425M in Q3 2026, and tokens from 21 trillion to 602 trillion. Spend growth did slow sharply, from 342% in Q2 to 106% in Q3. Prices were not the driver: median cost per million tokens stayed between $0.58 and $0.78. The same seats are simply consuming far more tokens.

One detail deserves more attention. Cache-read tokens rose from 89.6% to 94.0% of all tokens. A growing share of every AI dollar is spent re-reading context the model has already seen, not producing new work.

Same models, same ceiling. The difference is context.

The report makes a striking observation: this is not a story of leaders pulling ahead and laggards falling behind. Throughput, developer experience and innovation are all converging across the sample. Since everyone has access to broadly comparable foundation models, model capability cannot explain the ceiling.

DX’s hypothesis is that the quality of context given to AI is a major differentiator, and it introduces a framework, CAFE(S), that scores context on Clarity, Actionability, Fidelity, Efficiency and Security. The authors are explicit that this mapping is diagnostic, not yet measured, and list other credible drivers too: codebase age and size, test coverage, architecture, review culture and team tenure. It connects to what researcher Margaret-Anne Storey calls intent debt: the rationale and constraints that live in people’s heads are invisible to an agent unless someone writes them down.

Six moves for engineering leaders

The report closes with recommendations we largely agree with. Here they are in our words, with how we would put each into practice.

How to read this data

Any benchmark is shaped by who is in it. The report’s sample comes from organizations that use DX’s own platform, which the authors note skews toward companies already investing in productivity measurement. Developer experience and time-saved figures are self-reported survey data, which the report triangulates against system telemetry. Results are aggregated at the company level, so variation between teams inside one organization is not visible. None of this undermines the direction of the findings. It does mean your own numbers matter more than any industry median.