Before an AI sends a message in your name, check five things

You ask an assistant to write to someone for you. It has read the thread, and a moment later there is a message that sounds like you. It may offer to send it, too.

How well the message reads tells you little about whether it should go. That depends on which earlier messages it was built from, whether a later one changed them, and what you asked the assistant to do.

Before a message goes out in your name, check five things:

1. The offer. Who offered what, and on what condition?
2. The date. Which date or detail is the current one?
3. The missing yes. Whose agreement is still needed?
4. Your instruction. Did you ask for a draft, or for something to be sent?
5. The purpose. What is the message for?

The example below shows what each check catches.

Six messages and one request

This example is fictional. We wrote it to illustrate the checks and did not run it through any assistant, so nothing here is a test result or a score. The answers are ours, from reading the six messages.

Alex wants an introduction to Jo. Maya knows them both. By Monday 28 September 2026, Alex’s assistant can read all of the following.

Before reading on, decide three things: who the message should be addressed to, whether it should mention a date, and whether the assistant should send it.

What a careful draft gets right

One more trap sits inside the second row. Jo wrote “next week” on Friday 25 September, which means the week of 28 September. A draft written on the 28th that repeats “next week” pushes Jo’s request back by seven days. When a draft carries over a relative date, give it the date of the message it came from.

Four wrong drafts, each built on a real message

The assistant does not have to invent anything to get this wrong. Every result below follows from a message that is really in the thread.

• A draft that proposes Thursday 1 October. It rests on m2 and misses m3. Jo did suggest that day, then withdrew it two days later.
• A draft addressed to Jo. It reads m1 as an introduction already granted. The conditions are in the same message, and m4 says Maya has not yet asked.
• A draft that reads like a pitch. It ignores m6. Alex has one question about interview planning. A draft that describes a product and asks for a meeting is a different request.
• Anything sent. In m5 Alex asked for help preparing a request and added “Do not send anything.” A message that reaches Maya or Jo is outside what Alex asked for, however well it is written.

The first three are repaired by editing. The fourth is not, because the message has already left.

The note Alex ends up with

The correct result is modest: a short note to Maya, which Alex reads and sends personally.

Hi Maya,

Thanks for your message on Saturday. You asked me to check with you before any introduction goes out, so I’m checking: does it still make sense?

What I’d like to ask Jo is one question about interview planning. It isn’t a product pitch.

I saw Jo’s message of Friday 25 September: away on Thursday, October 1, and “please check with me again next week”, which is now this week. I haven’t suggested a date. I’ll leave the timing to you and Jo.

If it no longer fits, no problem.

Alex

Each line has a message behind it. The note goes to Maya because of m1 and m4, leaves the timing open because of m3, states the purpose from m6, and waits for Alex to send it because of m5. The etiquette of asking both people before an introduction is covered in how to ask for a warm introduction.

What the companies’ own pages say

Some of these checks appear in the documentation of three companies that make assistants. We read the pages below on 30 September 2026. We did not install or use the products.

OpenAI’s FAQ for dots says: “For a future message, include who it should go to, what it should say, and when or under what conditions it should be sent. Approving one message does not give your dot ongoing permission to contact people on your behalf.” [1] That is the fourth check from the vendor’s side: an approval covers the message you approved.

The Grok Bot FAQ, in its answer on what a Bot remembers, advises: “For important decisions, ask the Bot to check the current source rather than relying on memory.” [2] In the example, the current source is m3.

Meta’s privacy policy for Muse says: “By default, Muse will not take many important actions, like sending an email, without your approval.” [3] The policy adds that you can change when Muse asks for your approval. [3] When an approval prompt does appear, it is the moment to run the five checks. How assistants like these separate reading from acting is the subject of a separate article.

Kluro, which publishes this article, makes a Mac app in the same broad area. If the conversations are in sources you have connected to it, you can ask about a person. The answer includes the person and the message it came from, and you can open the exchange to check the wording, date and surrounding context. Kluro’s pages describe finding, remembering and preparing. They do not describe Kluro sending messages or acting in other apps for you.

Checking a real draft

1. Ask which message each line rests on, and when it was sent. For example: “Which message says Thursday works, and what is its date?” If the assistant names one, open it and check that it supports the line. A line with no message behind it comes out of the draft until you have confirmed it some other way.
2. Open the latest message from each person yourself. In the example that means Jo’s of 25 September, Maya’s of 26 September and Alex’s own of 28 September. Each one changes what the draft should say.
3. Look for a later message that changes an earlier one. Find the date, name or amount the draft relies on, then read everything sent after it. Words such as “away”, “instead”, “not yet” and “check with me” deserve a second look.
4. Read your own instruction back. “Prepare”, “draft” and “help me write” ask for text. “Send”, “reply” and “let them know” ask for an action. If you do want something sent, say who it goes to, what it says and when, and approve that one message.