AI coding assistants are boosting volume — but quality is suffering

The headline signal. Justin Reock from DX presents data from their platform covering over 200,000 engineers and 400+ organizations. The headline finding: AI coding assistants increase raw throughput metrics like deployment frequency and PR count, but those gains come with measurable degradation in quality, developer confidence, and delivery discipline. Change failure rates have become more volatile, with some organizations shipping 50% more defects relative to the industry benchmark of 4%.

The trust paradox. The most striking finding is a growing divergence between two traditionally correlated metrics. Perceived code maintainability rose nearly 4%, yet change confidence dropped 6%. Developers find the code easier to understand and modify but trust their own changes less. They are more afraid of breaking things than a year ago. This psychological tension is not being addressed by tooling alone.

PR size as a proxy. Average PR size has grown from 44 lines to 72 lines over a year. Reock attributes this to two forces: AI models generate average-quality, verbose code, and expensive CI pipelines incentivize developers to batch changes into fewer, larger PRs. Every extra line is a potential bug, vulnerability, or review burden. This directly undermines incremental delivery, which saw a 10% drop in developer sentiment.

Uneven distribution of gains. Junior developers use AI the most and spend more tokens per use case. Senior staff save about the same amount of time with fewer tokens, because they are better at spotting hallucinations and understanding architectural context. The median productivity gain across the study was a modest 7.7%. No organization came close to 2x, let alone 5x or 10x. The hype of order-of-magnitude improvements is not reflected in the data.

The real bottleneck is not code generation. Reock argues that even perfect code generation only addresses 14-16% of the overall value stream. The real constraints remain organizational: meeting-heavy days, context switching, slow CI pipelines, and friction in development environments. Time saved on non-bottleneck activities is worthless, as the Theory of Constraints suggests. The data supports this: small companies with less organizational friction see larger time savings.

What teams should do. The concrete recommendation is to measure AI impact through foundational developer experience metrics rather than vanity utilization numbers. Teams should compare cohorts of AI users vs. non-users across change failure rate, PR size, and cycle time. They should invest in platform readiness—clear documentation, modular code, non-flaky tests—because what is good for humans is also good for agents. Reock's framework focuses on three dimensions: utilization, impact, and cost.