Stripe Systems cut average pull request review cycle time by 60%, from 4.2 hours to 1.7 hours from submission to approval, after adding an AI-assisted code review pipeline to three production projects: a NestJS backend, a Flutter mobile app and a React frontend. Over six months, senior engineer review time fell from 45 to 18 minutes per PR, also 60%. Defects caught in review rose from about 8.2 to about 11.1 per 100 PRs (AI and human combined), and the false positive rate fell from 22% in month 1 to a stable 8% in months 4 to 6.
Stripe Systems Engineering wrote in a March 2026 post: "Code review is the most important quality gate in a software team, and it is also the most common bottleneck." Before the build, two months of measurement showed average review waits of 5.1 hours on NestJS, 3.8 on Flutter and 4.6 on React, and that 55-60% of review comments addressed issues that do not require human judgment.
The model has two tiers. Tier 1 is an automated AI review of every PR in under two minutes; the developer addresses valid AI comments before Tier 2, a human review focused on business logic, architecture fit, security, performance and overall design. A GitHub Actions workflow triggers on opened, synchronize and ready_for_review events and skips drafts. A gather-context.js script assembles the diff, related files, architecture-rules.md and coding-standards.md, capped at 80,000 characters. A review.js script sends this to claude-sonnet-4-20250514 using a NestJS, Flutter or React prompt template, and the model self-rates each finding so that only comments above a 0.7 confidence threshold are posted through Octokit.
Tuning cut false positives by raising the threshold from 0.5 to 0.7, adding project-specific rules, and a thumbs-up and thumbs-down feedback loop on each comment that the team reviews monthly. Quarterly survey scores rose from 3.2 to 4.3 for "AI review saves me time" and from 2.8 to 4.1 for "I trust AI review to handle mechanical checks". The token cost is approximately $80-120 per month across about 45 PRs per week. Setting up the pipeline, writing project-specific prompts and tuning false positives took about 3 weeks of one engineer's time, and prompt maintenance takes about 2-3 hours per month. In its Limitations and Honest Assessment section the post says: "AI review is not a replacement for human review. It is a filter that reduces the mechanical load on human reviewers."