Baseline metrics
These metrics become the baseline that later assessments compare against. They are recorded, not scored: they tell you whether the scores are matched by what actually happens.
| Metric | Why it matters |
|---|---|
| Lead time, request to production | The end-to-end result, not a stage |
| Review pickup time | How long a change waits for its first human look |
| Human approvals and wait per PR | Whether human attention is being spent or queued |
| Share of merges with zero review, by tier | Where the safety net is thin |
| Change failure and revert rate | Quality of what ships |
| Incidents per 100 merged changes | Whether speed is creating production risk |
| CI queue time and flake rate | Platform fit for agent volume |
| Time to roll back | Cost of being wrong |
| Changes linked to acceptance criteria | Specification discipline |
| Backlog trend | Whether extra output is absorbed or piling up |
Every metric here is better when lower, apart from links to acceptance criteria. Pick one for the prediction in your decision record.
Reported reference points
From vendor reports, so best treated as directional. Each vendor sells measurement tools, so each has a reason to find a problem worth measuring.
- Review waits. LinearB's 2026 benchmarks (8.1 million pull requests from 4,800 teams) report agentic pull requests waiting 5.3 times longer for pickup than unassisted ones, and AI-generated pull requests accepted 32.7% of the time against 84.4% for manual ones.
- Unreviewed merges. Faros AI's 2026 report (22,000 developers, comparing each organization's lowest and highest AI adoption periods) reports pull requests merged without any review up 31.3% and the incidents-to-PR ratio up 242.7%.
- More output, longer queues. DX's Q3 2026 report (500+ companies) shows weekly pull requests per developer up 75.5% over four quarters, and a 105% longer wait for first review at the organizations producing the most AI code.
How each figure was checked is on the evidence page.
Goodhart's warning
When a measure becomes a target it stops measuring well. Report these metrics together, never one alone, and never treat pull-request counts or the share of AI-written code as success.