Intercom lets an AI agent approve 19% of pull requests, with 93% of PRs now agent-driven
Intercom set a goal to double R&D productivity in 12 months and reports hitting it in nine, with downtime from breaking changes down 35% as deployments doubled. With more than 93% of PRs across its two main codebases now agent-driven, human review became the bottleneck, so it built a PR review agent that splits review into sub-agents (problem statement, intent vs diff, safety, logic, best practices), traces execution paths, uses Intercom-specific guidance refined by engineer feedback, and refuses to approve large PRs. A pilot of 100+ PRs had zero reverts and 6-16x faster time-to-approval at p75. In the first four weeks of wider rollout, 497 PRs went fully autonomous with Claude writing the code and the AI system approving and shipping. Over 19% of PRs are now auto-approved, and auditors Schellman were engaged early for SOC 2, HIPAA and ISO compliance.
The numbers (as reported)
| Agent-driven PRs | Over 93% (two main codebases) |
|---|---|
| Auto-approved PRs | Over 19%, no human reviewer |
| Pilot | 100+ PRs, zero reverts |
| Time-to-approval (p75) | 6-16x faster |
| Fully autonomous PRs | 497 in first four weeks of rollout |
| Revert rate, backend | 0.53% AI-authored vs 5.39% human |
| Revert rate, frontend | 0.22% AI-authored vs 2.00% human |
Facts
- Company / subject
- Intercom
- Tools
- Claude
- Codebase
- brownfield
- Work type
- Company rollout
- Stack
- —
- Published
- 2026-04-21
- Last checked
- 2026-11-06
What the source claims Practitioner (own blog/paper)
Claimed by the builder, writing about their own work. Not checked independently.
- Downtime from breaking code changes dropped 35% while deployments doubled -35% downtime
“downtime from breaking code changes dropped 35%, even as our deployments doubled”
- 6-16x improvement in time-to-approval at the 75th percentile in the pilot 6-16x
- Zero reverts of AI-approved PRs in the 100+ PR pilot 0 reverts
- AI-authored backend code reverted at 0.53% vs 5.39% for human-authored 0.53% vs 5.39%
- Doubled R&D productivity in nine months
“Nine months later, we did it.”
Caveats
Self-reported by Intercom. The post names Claude as writing the code but not the exact tool or the models behind the review agent. Revert-rate comparison is observational: AI-approved PRs are filtered to be small, which may favour them.
More stories by tool: Claude
Related topics: Company rollouts and costBrownfield codebases