“Keep a human in the loop” is not a complete workflow design. The person needs to see what will happen, which account or record it affects, and what evidence supports it. Otherwise approval becomes a button people press without a meaningful decision.
Approve a concrete action
For an illustrative customer follow-up, show the recipient, full message, attachments, and intended channel. If the assistant changes the recipient or price after approval, the old decision should not authorize the new action.
For a CRM update, show the current and proposed values rather than asking whether the summary looks good. Reading a draft and authorizing a system change are different decisions. The interface and operating rules should make that distinction clear.
Handle stale and repeated decisions
The record may change while approval is pending. A colleague might resolve the ticket or update the quotation. Before execution, check that the relevant preconditions still hold. If they do not, return the changed proposal for review.
A repeated approval click or a retried request should not create two refunds, two messages, or two tasks. The system needs a durable way to recognize the same intended operation. This belongs in the integration, not only in a prompt that says “be careful.”
Make rejection useful
Let the reviewer correct the proposed content or explain why it should stop. Preserve the reason where it can improve the next attempt without turning a one-time exception into a permanent rule. “Use this discount for this customer” should not quietly become the default for everyone.
Measure the review burden. If people routinely rewrite everything, the assistant may be producing the wrong artifact or missing context. If they approve instantly without reading, the approval design may be hiding the consequence.
Start with a narrow authority boundary
An initial pilot can prepare evidence and drafts while employees execute actions in their usual systems. Direct execution can be added after the checks and recovery behavior are proven. This sequencing keeps the value test separate from a broad permissions rollout.
Golub Superintelligence can help scope those decision points in a custom integration. For individual preparation work, see Golub Bot; for a conversation-led process, see Golub Voice.