AI-assisted development is no longer a minority sport: around 90% of professional developers use AI tools, spending a median of two hours a day working with them, and the DORA research program's latest report finds more than 80% saying AI has enhanced their productivity. At the individual level the effect is dramatic: studies measure roughly 21% more tasks completed and nearly twice as many pull requests merged.
Where the speed goes
Zoom out to the organization and the picture flattens: delivery metrics for whole teams often barely move. The code-quality data hints at why. Copy-pasted code has climbed from 8.3% to over 12% of changed lines, while refactoring has collapsed from roughly a quarter of all changes to under 10%. Meanwhile reliability is the top production concern for over half of enterprise decision-makers. Generation got faster; review, integration, testing and trust did not. Left unattended, AI throughput converts into review queues and technical debt rather than shipped features.
Discipline is the multiplier
- Hold the review bar and resource it: if PR volume doubles, review capacity and tooling have to double with it, or quality silently pays the bill.
- Let tests and CI gates be the arbiter: AI-written code merges on the same evidence as human-written code, ideally with the tests generated and hardened first.
- Schedule refactoring on purpose: the data says it won't happen by default anymore, so make it an explicit, recurring line item.
- Instrument delivery, not activity: measure lead time, change-failure rate and time-to-restore, and judge AI adoption by those, not by tasks completed.
This is the gap where we spend most of our engineering-advisory time: wiring AI acceleration into a delivery system (reviews, tests, observability) so the speed reaches production instead of piling up in front of it.