Intercom's AI PR Review Agent — 93% Agent-Driven, 19% Auto-Approved, AI Code Reverts 10× Lower Than Human

Title

Intercom's AI PR Review Agent — 93% Agent-Driven, 19% Auto-Approved, AI Code Reverts 10× Lower Than Human

Content

Intercom — the San Francisco AI customer service platform — has built and deployed an in-house "PR review Agent" across its two main codebases. Today, over 93% of Intercom's pull requests are Agent-driven, and over 19% are auto-approved with no human reviewer in the loop. The agent has been audited by Schellman against SOC 2, HIPAA, ISO 27001, ISO 42001, and AIUC-1. Per Intercom's engineering team (authors Kesha and Niamh Young, Apr 2026): "Speed is not the enemy of safety. It's a prerequisite for it. Accumulating code creates risk. Shipping small batches minimizes it. The faster you ship, the smaller each change is, and the easier it is to catch problems, and roll back when something goes wrong as the context is still fresh in your head." The motivation: when an AI agent can produce a working implementation in minutes, waiting hours or days for human review creates an impedance mismatch — either the queue backs up, or — more dangerously — humans start rubber-stamping. Intercom's CTO Darragh Curran set an explicit goal to double R&D productivity within 12 months; they hit it in 9 months. Downtime from breaking code changes dropped 35% even as deployments doubled. PR review was the next bottleneck. The PR review Agent decomposes review into separate sub-jobs handled by independent sub-agents: one assesses problem-description quality, another checks whether the diff aligns with stated intent, another reviews for safety concerns, another for logical correctness, another against best practices and anti-patterns. The agent also traces execution paths through the codebase — something humans rarely had time for. Reviews are grounded in Intercom-specific guidance that engineers continuously refine, with engineers flagging helpful comments to feed a flywheel. Pilot (100+ PRs): zero reverts of AI-approved PRs; 6–16× improvement in time-to-approval at the 75th percentile. First 4 weeks of broader rollout: 497 PRs went fully autonomous — Claude writing the code, Intercom's AI approval system reviewing, approving, and shipping to production. The agent is strict: it won't approve large PRs, instead flagging and requiring them to be broken down, creating a direct incentive to ship smaller, more incremental changes. Operational metrics: 93% of PRs Agent-driven; 19% auto-approved. AI-authored backend revert rate 0.53% (vs 5.39% for human-authored, ~10× lower). AI-authored frontend revert rate 0.22% (vs 2.00%, ~9× lower). Pilot: zero reverts. 4-week rollout: 497 fully autonomous PRs. Downtime from breaking code changes: -35% with deployments doubled. Intercom is also building a parallel agent that "proactively diagnoses issues in production," with the team noting: "We'll share more on this soon." Automated approval is never forced — any engineer can request human review; the human who ships remains accountable. Every AI-approved PR is labeled, logged, and queryable.