Decide separately whether AI reply drafts should be sent

An AI assistant can provide a reply draft to the service team while the sending decision stays with the team. This function can be useful on its own. Whether to add automatic sending depends on the type of message, its evidence and the consequences of an incorrect statement.

Different messages need different workflows

A fixed acknowledgement can refer to a confirmed save operation. This allows a narrow workflow built around a checked template. A freely worded technical solution, by contrast, contains factual claims that need to match the current situation.

In this fictional example, a customer reports a failing data export. The internal case status is: investigation open, cause unconfirmed. An assumed reply draft nevertheless says, “The export error has been fixed.” The sentence is understandable, but the case status does not support it.

A review interface must make this discrepancy visible. Staff need the customer’s question, the current status and the draft together. A lone approval button without this evidence provides limited support for a substantive review.

Approval needs a specific content version

The Microsoft guidelines for human-AI interaction recommend making unwanted suggestions easy to dismiss and supporting efficient correction of incorrect results. In the example, the team should be able to edit the unsupported completion claim directly. Rejecting it must keep the draft in the internal work queue without accidentally triggering a message.

Approval applies to the reviewed content and the intended recipient. If the text or recipient changes afterwards, the workflow must resolve approval again. The sending service also needs its own permission checks. OWASP recommends human approval before high-impact actions and enforcing access rules in the system performing the operation. An approval button does not replace those rules.

Check the value of the chosen scope

This does not imply mandatory individual review of every automated message. A checked automated workflow may suit narrowly bounded, supported information. Where individual diagnosis or a consequential commitment is involved, a draft followed by staff handling may be the appropriate scope.

Assess whether suggestions are actually usable, how easily they can be corrected and whether the transition to sending remains clear. Automation can initially end at an internally editable draft. Extending it to autonomous sending needs its own suitable test cases and an accountable decision.

Sources

• Microsoft HAX Toolkit — HAX Design Library

• OWASP — LLM06:2025 Excessive Agency