The latest data from DX’s State of AI Impact in Engineering: Q2 2026 shows that PR (Pull Request) throughput increased significantly over the past year, while developers are also delegating a growing share of coding work to AI.
At first glance, this looks like the productivity breakthrough companies have been waiting for.
But there is a complication: PRs are also getting much larger, developer confidence in making changes has declined, and the share of engineering effort going toward innovation versus maintenance has remained relatively flat.
In other words, more code is moving through the system, but that does not automatically mean more value is reaching customers.
This may point to a new kind of engineering bottleneck. For years, writing code was one of the expensive parts of software development. AI is rapidly reducing that constraint. But reviewing, validating, integrating, testing, understanding, and ultimately taking responsibility for that code still require significant human attention.
The bottleneck may therefore be shifting from code generation to code verification.
That matters because many organizations still measure AI impact primarily through developer productivity: hours saved, code produced, PRs merged, or licenses adopted. These metrics are useful, but they are not the end goal.
The real question should be: Has AI increased the organization’s ability to safely deliver useful outcomes to customers?
If AI produces larger volumes of code while review queues grow, systems become harder to reason about, and teams spend more time managing the resulting complexity, some of the apparent productivity gain may simply be moving downstream.
The next phase of AI adoption in engineering will therefore require more than giving developers better coding assistants. Companies will need to redesign the entire software delivery system around the new economics of code: stronger automated testing, AI-assisted reviews, smaller and more autonomous changes, better observability, agentic QA, and much faster feedback from production.
The competitive advantage may ultimately belong not to the companies that generate the most code with AI, but to those that can turn AI-generated capacity into more customer value without creating more complexity.
Perhaps the most important metric for AI in engineering is therefore not:
How much faster are our developers coding?
But:
How much more value can our engineering system safely deliver?
