Intercom: The Review Gate Is the Decision, and Auto-Approval Can Be Made Safe

Intercom's AI Is Approving Our Pull Requests. Here's How We Made It Safe. (Kesha Mykhailov and Niamh Young, April 2026) is the review-gate report, the third of the three reports. Across Intercom's two main codebases, over 93% of PRs are agent-driven and over 19% are auto-approved with no human reviewer, and the auto-approved code reverted less often than human-reviewed code (backend 0.53% vs 5.39%, frontend 0.22% vs 2.00%). The line that carries it: "Human review is not a guarantee of safety. It never was."

What it establishes:

• Review is a step you can automate. A decomposed reviewer (independent sub-agents for intent, safety, and correctness) traces execution paths through the codebase that humans "rarely had time to" follow.
• Scope refusal is the safety control. The agent will not approve large PRs, which forces small, reversible changes, and any engineer can request a human review at any time.
• Evidence by construction. Every decision is "labelled, logged, and queryable," and Intercom's auditors (Schellman) confirmed the process meets SOC 2, HIPAA, ISO 27001, ISO 42001, and AIUC-1. The engineer who ships stays accountable.
• What it leaves open. The criteria behind the 19%, the execution environment, and agent identity. Intercom notes most of its largest outages came from infrastructure, not changes a PR review would catch.

The guide builds on it in review as a gate, auto-approval by risk, and evidence as a byproduct.