Four production AI workflows at Intercom that doubled R&D throughput by automating tech-debt clearing, platform plumbing, and bounded engineering work using Claude Code.
Intercom turned AI coding from individual experimentation into four production workflows that doubled R&D throughput across the engineering org.
The system is less a single agent than a repeatable operating model: identify high-volume engineering work, wrap it in a constrained workflow, and let Claude Code operate inside reviewable lanes. Human engineers stay responsible for intent, review, and merge decisions, while the AI handles repetitive implementation and cleanup work.
The interesting move is organizational. Instead of measuring AI by isolated prompt wins, Intercom treated each workflow as production infrastructure: scoped input, model-assisted execution, review checkpoints, and throughput reporting. That makes the workflow teachable to other teams rather than dependent on one unusually effective prompter.
Start with tasks that already have clear review standards: tech-debt cleanup, migration chores, tests, bug triage, or narrow implementation requests.
Workflow candidate checklist:
- Repeats every week
- Has clear acceptance criteria
- Can be reviewed in a pull request
- Fails safely when rejectedEach workflow gives the agent a bounded target, relevant files, success criteria, and the expected review artifact.
Task brief:
Goal: <one outcome>
Allowed files: <paths>
Done when: <tests/checks>
Return: PR summary + risks + verificationThe human loop is explicit: inspect the diff, require tests, reject weak changes, and feed the pattern back into the next workflow iteration.
| Tool | Version | Role | Why this tool |
|---|---|---|---|
| Claude Code | Current | Code execution | Works directly in a repo and can produce reviewable diffs instead of detached suggestions. |
| Workflow templates | Current | Orchestration | Turns individual AI use into repeatable team practice. |
| Pull requests | Current | Quality gate | Keeps AI output inside the same review surface as human code. |
| Throughput metrics | Current | Evaluation | Measures whether workflows actually change delivery, not just demo well. |
SystemYou are a repo-local engineering agent. Make the smallest change that satisfies the task and leave a reviewable trail.
User templateImplement <workflow task>. Stay inside <allowed paths>. Verify with <checks>. Return summary, risks, and follow-up work.
Privacy. Public discussion on the How I AI podcast; specifics drawn from public episode content.