Nearly every developer now codes with AI. The gap between teams is no longer the tools it's what they do around them. This report maps what the data actually says, the amplifier effect that decides who wins, and the specific things elite teams do that the rest skip.
The trap most teams walked into: roll out AI assistants, watch commits and closed tickets climb, and mistake more code for more progress - while stability slips and rework quietly piles up underneath the vanity metrics.
What the leaders do instead: treat AI as an amplifier of the system around it - small batches, strong tests, real review, and delivery metrics - so cheap code generation becomes real velocity instead of faster technical debt.
Reviewable Batches
Leaders resist letting AI generate huge changes at once. Small, testable changes were the foundation of high performance before AI - and they matter more now that generating a large change costs almost nothing
AI without a strong test and evaluation harness just produces unverified code faster. Elite teams point AI at their tests as much as their features, and trust the pipeline to catch regressions because they built it to.
Not Activity
The losing move is celebrating lines of code and accepted suggestions. The winning move is watching deployment frequency, lead time, change-failure rate, and restore time - and holding AI to them.
Stop treating "we use AI" as a strategy. Track the four delivery metrics that mattered before AI, and find out which teams actually got faster versus just busier.
AI makes big, risky batches cheap to produce. Keep changes small and reviewable so human review stays meaningful and stability holds.
Invest in the test and eval coverage that lets you trust the pipeline, and adapt code review for a world with far more code proposed per day.
AI accelerates the first 70%; human judgment owns the last 30% that decides whether it's safe to ship. Treat AI output as a draft under review, never as done.
It's a real GitHub number from 2023, measuring code suggested and accepted through Copilot, and it's often mis-cited as a current, universal figure. Treat it as evidence that AI writes a large share of first-draft code — not that half of all software is machine-authored.
On small, well-defined tasks, yes — a controlled study found a 55.8% speedup. On end-to-end delivery of production software the evidence is mixed, and early data even linked heavy adoption to reduced stability. Speed on a task is not speed to value.
Most likely because AI made it easy to ship larger, faster changes, and large batches hurt stability. The fix is process — smaller changes, stronger tests, tighter review — not less AI.
Strengthen the system around the AI: your tests, your review process, and your delivery metrics. That's what separates the teams compounding value from the teams accumulating debt.
CTOs, VPs of Engineering, and Heads of Platform deciding how to turn near-universal AI adoption into a real delivery advantage.
Drop your details and we'll send The State of AI-Assisted Engineering 2026 straight to your inbox - no spam, unsubscribe anytime.
Talk through how this applies to your roadmap with our engineering leads - a working session, not a sales pitch.
Download White Paper