Code Review

AI reviews code changes, flags defects or issues, and in some cases helps approve pull requests before engineers do a final check.

Business impact

• automated code reviews completed — Citigroup reported about 1 million automated code reviews completed in the first eight months of 2025.
• developer time capacity freed up — Citigroup said the automation freed up 100,000 hours each week.
• software bugs caught earlier in the development cycle — GM reported bugs were caught 10x more often earlier in the development cycle.
• grounded F1 score — GitHub said Copilot code review reached a 40.1% grounded F1 score on ReviewBench.
• 工数削減 — NEC reported up to 66% less effort in an internal validation of Metabob.

Data requirements

• code changes (text) — The systems review source code and pull request diffs.
• pull requests (text) — ReviewBench and production workflows both center on pull request changes.
• test files (text) — Shopify reported that the internal tool analyzed test files and automatically fixed common issues.
• repository context (text) — GitHub says the reviewer uses wider repository context.

AI methods and techniques

• automated code review — The system inspects pull requests or code changes and flags likely issues for review.
• AI-assisted review — Engineers use the AI output to support their own review decisions.
• benchmark-based evaluation — GitHub uses ReviewBench to measure and improve code review quality before production experiments.
• agentic workflow — GitHub says Copilot code review moved to an agentic architecture that pulls in wider repository context.

AI models and model families

Claude Sonnet 5

Industries

Real-world evidence

7 documented case studies on record.

Companies using this: Citigroup, DoorDash, GM, NEC, Shopify, Uber, GitHub.