The Velocity Illusion
There's a stat making the rounds in 2026 that every engineering leader needs to sit with: AI tools now generate 41% of all code globally, yet code churn is expected to double this year. Delivery stability has decreased 7.2% according to Google's 2024 DORA report.
On the surface, everything looks better. PRs are moving faster. Cycle times are down. Industry median cycle time has dropped from 11 days in 2020 to under 7 days in 2026, driven largely by AI-assisted code review and better async practices.
But underneath those improving numbers, a different story is unfolding.
More Code, More Problems
About 66% of developers report that AI outputs are "almost correct" but still flawed — close enough to merge, broken enough to require rework. Research from GitClear analyzing over 211 million lines of code found that AI tools correlate with up to 9x higher code churn.
A recent MSR 2026 study examining 33,707 agent-authored PRs found a stark pattern: 28.3% of AI-generated PRs merge almost instantly (narrow, low-friction automation), but once a PR enters iterative review, many agents fail to converge. Reviewers spend real time on PRs that are ultimately abandoned.
This is the core tension: AI makes generating code nearly free, but reviewing and maintaining that code is still expensive.
The Metrics Gap
Traditional DORA metrics — deployment frequency, lead time, change failure rate, MTTR — remain valuable but increasingly insufficient on their own. They can tell you what is happening but not why. When AI inflates volume, your deployment frequency looks great while your rework rate quietly climbs.
The teams navigating this well are tracking a few additional signals:
- Code turnover rate — what percentage of recently merged code gets reverted or rewritten within 30 days
- AI vs. human rework ratio — if AI-generated code is being rewritten at 1.5x or higher the rate of human code, that's a red flag
- Innovation rate — the share of effort going to new features vs. bug fixes, maintenance, and rework
If innovation rate is declining despite rising velocity, AI is creating rework, not reducing it.
What Actually Helps
The answer isn't to stop using AI tools. It's to stop measuring only speed.
Enforce PR size limits. Track rework alongside throughput. Use tools that give you visibility into which PRs are high-risk before a reviewer spends time on them — Code Board's risk scoring does this automatically, but the principle matters more than the tool. Watch your change failure rate as closely as your deployment frequency.
Organizations that track quality alongside velocity consistently outperform those chasing speed alone. The teams that win in 2026 won't be the ones writing the most code. They'll be the ones whose code survives.
Top comments (0)