AI adoption in software engineering has crossed the 90% threshold, rendering control groups without neural networks practically obsolete. Yet, while code is generated in a couple of keystrokes, overall delivery to production has not accelerated. Developer productivity analytics platform DX (part of Atlassian) published its Q2 2026 findings based on data from over 500 engineering organizations. With DX deployed across GitHub, Airbnb, Pinterest, and Morgan Stanley, the benchmark offers an unvarnished cross-section of the market.
Output surges run into architectural bottlenecks
The input metrics appear compelling at first glance: median TrueThroughput grew 37% over four quarters, climbing from 1.42 to 1.94 pull requests (PRs) per developer per week. The share of AI-generated code in merged changes surged from 34% in Q1 2026 to 52% in Q2. However, the throughput gains were captured almost entirely by pure tech companies and agile teams of under 100 engineers, while large enterprises stalled.
Code is written faster, but delivery pipelines choke at the verification stage. Over the course of aggressive AI adoption, the median PR size nearly doubled.
Artificial intelligence drastically lowers the cost of writing new code, but fails to reduce the cost of reviewing it.
Bloated PRs are difficult to parse, linger in code reviews, encourage rubber-stamping, and dramatically raise the odds of letting critical errors slip into integration.
Pipeline fatigue and hidden technical debt
Instead of the promised engineering bliss, teams face deteriorating sentiment: the Developer Experience Index (DXI) slipped from 67 to 65 points over four quarters. Codebase maintainability is worsening, PR sizes are ballooning, reviews are slowing down, and incremental delivery is eroding.
For senior leadership, the takeaway is clear: buying AI assistant seats without overhauling the end-to-end delivery pipeline only inflates operational costs. Text generation is cheap, but skilled engineering attention remains scarce and expensive. Without re-engineering validation workflows, deploying Copilot simply shifts the traffic jam from the code editor directly into code review and QA.