Practical AI / GUIDE + WORKSHEET
Where a human should approve an AI workflow
Place review before consequential actions and give approvers the evidence, alternatives, and current state needed to decide.
THE STARTING POINT
Place approval where an action commits the business, changes an important record, or is difficult to reverse. Show the reviewer the source, proposed action, and uncertainty. An approval step is useful only when the person has the context, authority, and time to make a meaningful decision.
Identify the action boundary
Separate drafting, recommending, and executing. Summarizing a message may be a low-impact aid, while sending a promise to a customer or changing payment instructions has different consequences. Map what the workflow can do and what permissions each step holds. Keep permissions limited to the intended task rather than relying on a prompt to prevent every unwanted action.
Give reviewers a decision, not a rubber stamp
Show the original input, relevant record state, proposed change, and reason for escalation. Make edit, reject, and request-more-information paths available alongside approval. Avoid overwhelming reviewers with trivial checks that obscure important ones. Define who can approve which action and how the workflow behaves when nobody responds.
Recheck conditions before execution
The underlying record may change between recommendation and approval. Verify that the approved action still applies to the current state and that it has not already been executed. Preserve the approval reference with the resulting action. Monitor rejected suggestions and stale approvals to improve the process instead of treating every interruption as reviewer resistance.
WORKED EXAMPLE / ILLUSTRATIVE
A fictional customer-response workflow
AI drafts an explanation of a project delay. The project owner reviews the actual milestone and edits a date before approving. If the milestone changes again before sending, the draft returns for review instead of sending an outdated commitment.
| Stage | Permission | Evidence |
|---|---|---|
| Draft | Suggest text only | Source message and project state |
| Review | Approve/edit/reject | Current milestone and proposed commitment |
| Execute | Send approved version once | Approval and current-state check |
MAKE IT USEFUL
AI approval boundary
Design the decision and the state checks together.
Your notes stay in this page and are not sent to Smithers. Download or copy them before leaving; refreshing clears them.
Before you put it to work
- Place review before the consequential action.
- Provide source context.
- Allow rejection and correction.
- Check for stale approvals.
Adding a human button does not make a workflow reliable by itself. Review quality depends on evidence, workload, authority, and enforcement of the approval boundary.
IF THIS LOOKS FAMILIAR
Your tools. Your particular mess.
The worksheet is yours to use. If the difficult part is making it work with the systems you already have, that's the kind of thing we help with.