The Human Side of AI · 05

Read it before it goes out: using AI assistants at work

The assistant wrote the email in a few seconds. It reads well. Your name is at the bottom, and in a moment it will be in someone else's inbox.

AI assistants are good at first drafts: replies, summaries, follow ups, the note you have written a hundred times before. That is a real help.

The risk is not the draft. It is sending it without reading it. Once a message goes out, it is yours. The person who gets it does not know or care which parts a tool wrote. They know it came from you.

What assistants get wrong

The mistakes are rarely dramatic. They are small and easy to miss when you are moving fast:

• A detail that sounds right and is not. A date, a price, a name, a promise about when something will arrive.
• A summary that drops the one point that mattered, or quietly changes what someone said.
• A tone that does not fit. Too cheerful for a complaint, too stiff for a customer you have known for years, too sure of itself where you meant to be careful.
• Something carried over from an earlier thread or document that does not belong in this message.
• A commitment you did not mean to make.

None of these look like errors at first glance. That is the problem. A clean, confident draft is easy to trust.

Read it the way the other person will

Before you send, read the whole thing once, slowly, as the person receiving it. Not a skim of the first line. Ask yourself:

1. Is every fact in here true, and could I point to where it came from?
2. Does it say what I mean, and nothing I do not mean?
3. Would I be comfortable if this were forwarded to my manager, or to the customer's manager?
4. Is there anything in here that should not leave the company?
5. Does it sound like me?

If the answer to any of these is no, fix it or write it yourself. That usually takes a minute or two. Untangling a wrong message after it goes out takes much longer.

Be careful what you put in

Reading the output is half of it. The other half is what you paste into the tool: customer details, information about coworkers, pricing, contracts, anything private. Follow your company's rules about which tools are approved and what can go into them. If there are no rules yet, ask before you paste, not after. For company work, use the business account the company provides, not a personal one.

Some messages should be yours alone

A draft is fine for routine replies. It is a poor fit for the messages that carry weight: an apology to a customer, hard news for a colleague, anything about someone's job, anything you would be upset to see quoted back to you. Those deserve your own words, even if they take longer.

For owners and managers

This is a habit, and habits get taught. A short session on the real messages your team sends each week will do more than a policy nobody reads. Make the expectation plain: AI drafts, people approve, and whoever presses send owns the message. It is the same rule as in What your people should be doing now, applied to the place it matters most.

Make it safe to say the tool got something wrong. The people who catch mistakes are doing the job well. If they feel foolish admitting the assistant nearly sent something wrong, they will stop mentioning it, and you will hear about it from a customer instead.

And go first. If leaders send drafts without reading them, everyone else will too.

On Monday

Before you send anything an assistant drafted, read it once all the way through as the person getting it. If you lead a team, pick three real messages from last week and walk through them together: what the tool would have written, and what you would have changed.

Have a question? Ask it here.