Skip to main content

How to add human approvals before AI takes action

E
Written by Errin Veney

Before you begin

  • Begin in recommendation or shadow mode. Let a person approve the action while you measure performance.

  • Define which decisions are reversible and which always require explicit approval.

Design the review and approval process

Some workflows should begin with recommendations, not autonomous actions.

Use an approval step when the app can

  • Change access

  • Move money

  • Send an external message

  • Update an official record

  • Make a regulated determination

  • Reject a person or request

  • Trigger an irreversible process

  • Act on incomplete or conflicting information

Define

Who can approve?

What evidence must they see?

Which cases can follow the normal path?

Which cases always require review?

What gets recorded in the decision log?

How can the action be reversed?

A good first production mode is shadow mode. The app produces the recommendation beside the existing process, but does not execute it. Compare the results, correct the rules, and measure performance before increasing autonomy.

Common mistakes

  • Allowing the model to approve its own high-risk action.

  • Failing to show the reviewer the evidence and rule.

  • Creating an approval step with no response-time owner or escalation path.

Frequently asked questions

What is shadow mode?

The application produces a recommendation beside the existing process but does not execute the action. This lets the team compare results and correct rules safely.

Which actions should always require approval?

Irreversible, regulated, financial, external, access-changing, or high-impact actions should have explicit approval unless a documented policy says otherwise.

Did this answer your question?