An AI can send the email. The recipient still has to live with it.

Meta says Muse runs in a dedicated cloud computer, uses a separate Sentinel agent to review actions before they reach the internet, and shows an audit trail of completed and planned work. It also says people can disconnect services or change access later. Those details describe a real effort to put boundaries around a broad kind of access.

But a permission prompt can still hide the part that matters socially. “Send email” is not a useful description when the message accepts a meeting, commits a colleague to a deadline, turns down a request, or tells a seller you are ready to buy. A purchase confirmation can be equally vague if it leaves out the merchant, item, total, delivery date, return conditions, and whether a substitution is allowed.

The person using the agent should see the exact consequence in ordinary language. So should the recipient, through a message that does not pretend the commitment was personally typed if that distinction matters to the relationship.

Different actions need different last looks. Before an email goes out, show the recipient list, the final text, the account it is leaving from, any promise or date it contains, and whether it is a draft or a send. Before travel is booked, show each traveler, flight or room, total, cancellation terms, and what happens if a preferred option disappears. Before a purchase, show the merchant, item, quantity, delivery address, final charge, and return path.

That is not a request for a scary technical dashboard. It is a way to avoid making a person reconstruct a decision from a log after someone else has already acted on it. Meta says Muse will ask for approval on sensitive actions. The important product question is whether that approval screen carries enough of the actual situation to make approval meaningful.

If you are evaluating AI assistants, begin with a small task whose result you can inspect: ask it to prepare, but not send, a reply; compare two refundable travel options without booking; or assemble a cart without checkout. Then look for the details you would need if the result were wrong.

A practical trial should answer four things:

1. Can you see what information the assistant used? 2. Can you tell the difference between a suggestion, a prepared action, and a completed action? 3. Does the approval name the person, money, date, or record that will change? 4. Can you correct or reverse the result without a scavenger hunt?

The point is not to make every small task feel dangerous. It is to reserve confidence for the moments when an AI assistant turns a private instruction into someone else’s obligation.