Where to put human approval in an AI workflow
An approval step should answer a concrete question: should this proposed action happen? It works best after the workflow has assembled enough evidence for a person to decide, and before the action changes an external system.
Make the review actionable
For a payment review, show the original invoice, extracted amount, currency, matched vendor, and proposed destination. Asking someone to approve a model's summary without the original evidence forces them to trust the very output they are checking. Define rejection behavior too: stopping the record, asking for missing information, and retrying with corrections are different operations.
- 1Identify the exact action the approval permitsIdentify the exact action the approval permits.
- 2Present the proposed record alongside its source…Present the proposed record alongside its source evidence.
- 3Decide what rejection, expiry, and duplicate responses…Decide what rejection, expiry, and duplicate responses should do.
Try it on a small example
- Identify the exact action the approval permits.
- Present the proposed record alongside its source evidence.
- Decide what rejection, expiry, and duplicate responses should do.
What to verify
Exercise the approval path with a real pause during a small test run. Confirm that later steps do not execute before approval and that a resolved request cannot accidentally be acted on twice. Baleybots exposes pending approvals and response operations through its public API. An event saying approval was requested can arrive after someone has resolved it, so read its current state before preparing another review.
Product details checked against the Public API reference on October 4, 2026. These guides describe documented behavior; availability depends on your account and the deployed service. Baleybots is in invite-only beta.