Control & safety
5 min read· Updated 2026-09

Designing approvals so AI is useful but never unsupervised

The point of approvals is not to slow the AI down. It is to make the boundary explicit so you can expand it deliberately.

Gate by consequence, not by task

Reversible work — a draft, an analysis, an internal task — rarely needs a gate. Irreversible work does: anything a customer receives, anything that spends money, anything that changes a policy or permission.

A sensible starting configuration

This is where most teams should begin in week one:

  • Outbound customer email: approval required, always.
  • Spend of any amount: approval required.
  • Internal tasks, notes, analysis, drafts: no approval.
  • Policy or permission changes: owner only.

Loosen one thing at a time

After a few weeks, pick the single category with the cleanest track record and raise its limit. Check the activity log a week later. If the quality held, do the next one.

Use the audit trail as a management tool

Every decision, approval and action is recorded with who or what did it. That record is what lets you delegate more without losing the ability to answer 'why did this happen?' months later.

Related guides