Put approval before consequential action
Classify proposed actions by consequence, reversibility and the system's demonstrated reliability. Drafting a response and sending it are different permissions. Reading a record and changing it are different permissions. Define which actions can run automatically, which require review and which remain outside scope. The reviewer needs the actual proposed effect, not merely a summary that the agent considers the action safe.
Worked example
Illustrative scenario; not a claim of measured client results.
A support assistant may draft a refund explanation, but issuing the refund can require an authorized person to review the amount, customer, reason and policy basis. Approval should apply to that specific action. If the amount or destination changes after approval, the system should not silently reuse the old decision. Keep a record of what was approved and what was executed.
Checks, limitations and next decisions
If approvals become routine rubber stamps, examine whether reviewers have enough context and time. Narrow the automated scope or improve the review screen. Do not treat a human's presence as proof of safety. Test rejected actions, changed inputs and interrupted sessions as well as approvals, and ensure the workflow stops cleanly when authorization is absent.
A practical starting point
Work through these checks in order. Use the results to decide what needs to change in your own workflow.
- Classify actions by reversibility and potential impact.
- Require review for sensitive or consequential changes.
- Show the proposed action, evidence and expected effect before approval.
A mistake to avoid
Do not make approval a meaningless button that hides the relevant information.
How to check progress
Track reviewer changes, rejected actions and incidents that passed approval.
Your action checklist
Use this checklist while reviewing your own setup. Checkmarks are temporary and stay in this open page.
For the wider decision process, read Choosing the first AI workflow to pilot.
Further reading
Official background resources for this topic. The workflow above is general guidance, not a guarantee of a particular result.
AI-assisted editorial content. Examples are illustrative, not client case studies. Read our editorial policy or report a correction.