AI in the DMs: what an AI sales assistant should and shouldn't say

Why the DMs became a sales channel

For many experts, the first real conversation with a buyer now happens in a direct message. Someone sees an ad or a post, replies with a keyword or a question, and expects an answer quickly.

Speed matters here. Harvard Business Review reported in 2011 that firms which contacted online leads within an hour were nearly seven times as likely to qualify them as firms that waited even an hour longer.

An AI assistant can reply in seconds at any hour. The question is no longer whether it can answer, but what it should be allowed to say when it does.

What a good assistant should do

Think of the assistant as a polite, well briefed front desk. Its job is to move a genuine enquiry one step forward, not to close a sale.

• Greet the person by name and acknowledge what they asked about.
• Answer common questions from approved information only: what the programme covers, how long it runs, who it is for and how calls work.
• Ask two or three short qualifying questions, one at a time, in plain language.
• Offer the next step, usually a booking link or a short application, once the person has shown real interest.
• Hand the conversation to a human whenever it goes beyond what the assistant has been told.

That is a narrow job, and it should be. The narrower the brief, the fewer ways it can go wrong.

What it should never say

This list matters more than the first one. Write it down before you write a single prompt.

1. No result promises. Nothing like "you'll hit $20K a month" or "clients double their revenue". Even if a past client did, the assistant cannot know this person's situation.
2. No invented prices or terms. If pricing is only discussed on calls, the assistant says so. It never guesses a number, a discount or a payment plan.
3. No pretending to be human. If someone asks whether they are talking to a person, the answer is honest. Some places, California for one, have laws requiring sales bots to disclose that they are bots.
4. No professional advice. No tax, legal, medical or investment guidance, even when the question seems simple.
5. No pressure. No fake deadlines, no "only two spots left" unless that is true and confirmed by a human.
6. No arguing. If someone is rude, sceptical or asks to stop, the assistant stops politely.

Write the rulebook before the prompt

Most problems with AI in the DMs come from vague instructions. "Be helpful and book calls" leaves the assistant to fill gaps with guesses.

A better approach is a short rulebook the assistant must follow, written in plain sentences:

• Only state facts that appear in the approved information below.
• If you are not sure, say a team member will reply, and flag the conversation.
• Never mention price. Say pricing depends on the person's situation and is covered on the call.
• Ask a maximum of three questions before offering the booking link.
• If the person mentions a refund, a complaint or a payment problem, hand over immediately.

Keep the approved information in one document you update as the offer changes. When the facts change, the assistant's answers change with them.

An assistant that says "let me get someone for you" is not failing. It is doing its job.

A sample exchange

Here is how a well briefed assistant might handle a typical enquiry.

Prospect: "Saw your ad. How much is the programme?"

Assistant: "Thanks for getting in touch. Pricing depends on where your business is now and what you need, so it's covered on a short call. Can I ask what you sell at the moment and roughly how many clients you have?"

Prospect: "Coaching, about eight clients."

Assistant: "Got it. And what would you most like to change in the next six months?"

Two questions, no number, no promise, and the person is still talking. If they push for a price again, the assistant offers to have a team member reply directly.

Hand over, then review

The handoff to a human should be clear to the prospect. Something like "I'll pass this to [Name], who'll reply here shortly" sets an expectation the team must then meet.

Then read the transcripts. Once a week, go through a sample of conversations and look for three things: answers that stretched the facts, questions the assistant couldn't handle and moments where a person dropped off.

Each one is a line to add to the rulebook or the approved information. An assistant improves through that weekly review, not through clever prompting on day one.

Track a few simple numbers alongside the reading: how many conversations reached a booking link, how many were handed to a human and how many went quiet after the first reply. If handovers keep rising, the approved information is missing something people want to know.

The aim is not an assistant that can say anything. It is one that says a small number of true, useful things quickly, and knows exactly when to step aside.