How to review an AI agent’s reply before it reaches a contact

A helpful reply can become a problem if it promises a deadline you never agreed to or shares information with the wrong person. Drafting and sending are different actions. For a sensitive relationship, start by reviewing what the agent proposes before it reaches the contact.

Decide what the agent may say

Write a short boundary for the conversation: which topics are appropriate, what the agent must not promise, and when it should ask you. A reply that sounds confident is not evidence that an appointment, payment or delivery has been confirmed. Check the underlying calendar or record before approving the message.

Review a draft in APX

Supervised contacts let the agent prepare a reply while you decide whether to send it, answer yourself or ignore it. The contact’s conversation is sealed: the draft does not give that contact access to your private agent memory or tools. The guide explains which channels offer decision buttons and when you need the app instead.

If the person writes again before your decision, APX prepares a new draft for the accumulated messages and the earlier draft cannot be sent. If the model cannot produce a draft, do not mistake the absence of one for a reply sent to the person.

Check the result, not just the button

Read the incoming messages and the proposed answer together.
Verify names, dates, commitments and any private details.
Choose the intended action; if you answer yourself, check the recipient before sending.
Confirm the conversation shows the outcome. If delivery is uncertain, inspect it before retrying.

Supervision reduces the chance of an unreviewed answer; it cannot verify a factual claim for you. Start with a harmless conversation and adjust the rules based on what you actually observe.