Risk tiers
- Low risk — internal drafts, tagging, research digests → AI can run continuously; humans sample-review.
- Medium risk — CRM updates, customer-visible drafts, scheduling → AI proposes; human approves high-impact items.
- High risk — money movement, legal commitments, public posts, irreversible deletes → human required; AI only prepares.
Gate design patterns
- Approve to send — drafts queue until a person releases them.
- Approve exceptions only — AI handles happy path; edge cases escalate.
- Dual control — two humans for high-risk actions (finance).
- Time-boxed autonomy — AI acts within limits (hours, spend thresholds).
Anti-patterns
- No logging of what the agent did
- Approvals on everything (rubber stamp, zero speed)
- Approvals on nothing “because the model is smart”
- Gates without SLAs (queue dies unread)
SMB default
Start stricter than you think. Loosen gates only after measured reliability. Pair with evaluation samples weekly during Pilot and monthly during Run. See How we work for production philosophy.
Want this applied to your stack?
Fixed-scope pilots for AI agents and automations. Map first. Ship one real workflow. Then run it.