Firm-level data shows AI review tools write about a quarter of review comments. Coverage, not finding count, is the number to ask a vendor for, and here is how to read it from config.
Every AI reviewer claims to read AGENTS.md and CLAUDE.md. What decides whether your standards land is which rule wins when two apply to the same file, and whether the tool can show you which one it used.
When an agent produces a change, the thing a team can actually review is the workflow that produced it. What that looks like in practice, with primary sources.
A developer tried to tell which of 102 F-Droid apps were LLM-generated and re-labeled 10% on a second pass. Detection is too weak to build review policy on.
Augment Code's 450K-file monorepo test puts a number on self-hosting review tools: $4,100 to $9,100 a month once infrastructure and maintenance enter. Method and sources.
Salesforce's code volume rose 30% but review time on the largest PRs flattened. That plateau is a disengagement signal, not a throughput win.
GitHub's own tutorial for reviewing AI-generated code is an 8-step checklist, and almost every step is assigned to a human reviewer. Read the primary source.
The review bottleneck is not that there is more code. The gains from AI shrink on mature codebases, so where you spend review effort shifts.
Field note on a peer-reviewed review reporting code-review time up +441% under AI-assisted development, and why the productivity number to watch is the one nobody markets.