What AI should and should not do in a sales deal

The line

AI prepares, practises and checks. A person decides, speaks and sends. Everything in this post follows from that sentence.

The reason is simple. Anything said to a buyer is a commitment from your company, and a person has to stand behind it. An AI model can be wrong and sound certain. If its work goes to the buyer unread, the buyer finds the mistake before you do.

So the question for any AI task in sales is where its output lands. If it lands in front of your rep as a draft or a flag, good. If it lands in front of the buyer, stop. Here is how that works across the eight moves of The Salecopy Method.

Signal, Anchor and List: AI prepares

The first three moves are about knowing the deal before you negotiate it. AI is good at reading what you already have and laying it out.

• Signal. AI reads the notes and the call transcript and proposes the why-now in one sentence. The rep confirms or rewrites it. It must not supply a deadline the buyer never stated.
• Anchor. AI builds the cost-of-doing-nothing calculation from the buyer's stated inputs and shows every input, so the rep can check it before it is said out loud. It must not fill a missing input with a guess, and the figure does not go to the buyer until a person has checked the arithmetic.
• List. AI compares the contacts on the deal with the buying roles and flags the empty ones, for example a deal in negotiation with no budget holder attached. It must not reach out to those people. The rep asks the buyer who they are.

Evidence, Concessions and Options: AI checks and practises

The middle three moves are where deals lose money. Here AI works as a checker that reads the record before the call and tells the rep what is not ready.

These are also the three moves worth rehearsing. AI can play the buyer so the rep says the trade out loud before the real call. There is a full setup in how to practise a negotiation with AI.

• Evidence. AI pulls objections out of call transcripts, matches each to the evidence you already have, and lists the ones still open before the next call. It must not invent a reference, a number or a test result, and it must not mark an objection closed. The buyer closes an objection.
• Concessions. AI checks a proposed concession against the give and get list and the approval limits, and tells the rep which trade to offer and who must sign it off. It must not approve the concession or offer it.
• Options. AI refuses to call the deal ready for a pricing call until the walk-away is filled in, and drafts two smaller-scope versions from the price list. It must not choose the walk-away. That number is a business decision, and a person writes it.

Paper and Yes: AI drafts, a person sends

The last two moves produce things the buyer will read. This is where the line matters most.

One more limit. An AI summary of a contract is not a legal review. For contract terms, have your lawyer review.

• Paper. AI drafts the mutual plan from the deal's stage and the buyer's stated process, and flags any step that is past its date. It must not agree dates with the buyer or chase them. The plan is agreed between people.
• Yes. AI writes the follow-up from the call, with the agreed next step and date at the top, for the rep to approve and send. It must not send it. The rep reads every line, because the rep was on the call and the AI was not.

Keep a person in the loop

A person in the loop means a named person approves before anything leaves your company. It should be built into the setup, so it does not depend on anyone remembering.

• Every AI output arrives as a draft or a flag on the deal record. Nothing goes straight to an inbox outside your company.
• The AI has no permission to send email, change a deal's stage, change the amount or apply a discount.
• Each output shows what it was based on: the transcript, the fields, the price list. If the rep cannot see the inputs, the rep cannot check the answer.
• The record shows who approved the draft and when.

Cost every AI call

Each time an AI model reads a transcript or writes a draft, it costs money, usually based on how much text goes in and comes out. The amounts are small one at a time and easy to lose track of in total. Before you switch a task on, write down three things: what one run costs, how many runs you expect a month, and what work it replaces.

Here is an example with invented numbers. Say a transcript check costs 4 cents a run and your team holds 300 calls a month. That is 12 dollars a month for that task. Now you can decide whether it is worth it, and you will notice if it runs ten times more often than planned.

Costing also tells you when to run a task. A check that runs when a deal changes stage costs far less than one that runs every hour, and it tells you the same thing.

Three checks before you switch on a new AI task

Ask these of any AI feature, whether it comes inside your CRM (the system where your deals and contacts live) or as a separate tool.

If any answer is no, leave it off until it is yes. If you want help setting this up on the tools you already own, that is what our AI automation work covers.

• Does it prepare, practise or check? If it decides, speaks or sends, redesign it so a person does that part.
• Is there a named person who approves the output before the buyer sees anything?
• Do you know what one run costs and how often it runs?

What to add to your CRM

A prompt to sort your AI tasks onto the right side of the line

You are helping me review how my sales team uses AI. The rule is: AI prepares, practises and checks. A person decides, speaks and sends. You only analyze and report. Do not contact anyone, and do not draft messages to buyers.

Here is every task we use AI for in sales, or plan to: [list each task on its own line, for example: summarize call transcripts, draft follow-up emails, suggest discount levels]

For each task:
1. Say which it is: prepares, practises, checks, decides, speaks or sends.
2. Say where the output lands: in front of a rep, on the deal record, or in front of the buyer.
3. If it decides, speaks or sends, or if the output reaches the buyer without a person approving it, rewrite the task so that AI produces a draft or a flag and a named role approves it.
4. Name which of the eight steps it supports: Signal, Anchor, List, Evidence, Concessions, Options, Paper or Yes.
5. List what I would need to know to cost it: how often it runs, and roughly how much text goes in and comes out each time.

Do not estimate prices. Do not assume how a task works if I have not told you. Finish with a list of what is missing from my descriptions that you would need in order to give a firmer answer.