AI app guides
Put human approval where an AI feature changes the business
Make review specific to the message, record or decision being acted on, with a clear result afterward.
Zocri · 2 min read · Sources checked
Choose the action that needs judgment
Anthropic's 2024 engineering guide describes agent checkpoints for human feedback and explicit stopping conditions. For a business app, a useful checkpoint is often where a draft becomes an external action. An assistant can prepare a customer reply, but the business may want a person to approve it before sending. Define that boundary around the actual effect, not around every sentence the assistant writes.
List which actions are informational, reversible or consequential in your workflow. A suggested category may need different review from changing a delivery address or sending a promise to a customer. The responsible business owner should choose the policy, with implementation permissions enforcing it.
Show the concrete proposal
An approval screen should identify the recipient or record, the proposed content or change and the relevant context. A button saying approve the assistant's plan does not tell the reviewer what will actually happen. Give them enough information to make the decision without opening several unrelated screens.
Keep editing and approval understandable. If the reviewer changes a draft, the approved version should be the version executed. If the assistant changes it afterward, apply the chosen review rule again. Preserve the connection between the reviewed proposal and the eventual action.
Handle refusal and interruption
Allow the reviewer to reject, revise or defer without accidentally triggering the action. Explain what remains pending and who owns it. If the review expires or the underlying record changes, show the mismatch rather than executing an old proposal against a new situation.
Choose stopping conditions for work that cannot resolve itself. Repeated attempts, missing information or an unavailable service should lead to a clear handoff. The human should receive the relevant evidence and unresolved decision, not a long transcript they must reconstruct before acting.
Verify the effect after approval
Approval authorizes an attempt; it does not prove completion. Show whether the message was accepted by the delivery service or the record update actually succeeded. If the operation fails, preserve the approved proposal and provide a safe recovery path that avoids unintended repeated actions.
Test approval, rejection, edited approval and a changed record while review is pending. Check permissions using separate roles. A meaningful checkpoint gives the person control over a specific business effect, then returns honest evidence about the result, while the app continues to enforce the boundaries that human attention alone cannot guarantee.