Can ChatGPT Coach Me on My Sales Calls? An Honest Answer, With the Prompts

Yes, partly. Enough that you should do it. Not enough that it will keep working.

That is the whole answer, and almost nobody writes it down, because the people who publish about sales coaching are selling either a platform or a training program and neither wants to admit that a free chat window does a real part of the job.

So here is the useful version: the prompts that actually work, then the four specific places they stop working, so you can tell which wall you have hit when the feedback starts feeling thin.

What a general model is genuinely good at

Three things, and they are not small.

Naming what happened. Paste a transcript and ask what the buyer was actually worried about, and a frontier model will usually find it. It is reading comprehension, and reading comprehension is the thing these models are best at. It will spot the sentence where the buyer's tone changed while you were busy answering the question they asked instead of the one they meant.

Rehearsing something specific. Ask it to be a skeptical VP of Finance who has been burned by a tool like yours, and it will hold that character well enough to be uncomfortable. For a call tomorrow that you are dreading, this is worth twenty minutes.

Rewriting. Your follow-up email, your discovery questions, the two sentences you fumbled. It is a fast, patient editor and it never gets bored of your fourth draft.

The prompts

Use these directly. They are the ones worth having.

To get a read on a call:

Here is a transcript of a sales call I ran. Do not summarize it. Tell me the single moment where the conversation stopped going well, quote the exact lines, and tell me what the buyer was signaling that I responded to as if it were something else.

To find what you skipped:

Read this transcript. List the things I now know about their problem, then list what I would need to know to write a business case for this purchase. Show me the gap between those two lists.

To rehearse:

You are the VP of Operations at a mid-size logistics company. You agreed to this call because your CFO asked you to look at options, not because you want to change anything. Be polite, be busy, and do not volunteer information. I will start.

To pressure-test a deal:

Here is what I know about this opportunity. Argue that it will not close this quarter. Use only what I have told you, and tell me which of your arguments you are least sure about.

That last one is the highest-value prompt in the list, and it is the one people skip.

The four places it stops

1. It has no memory of your deals

Every conversation starts from zero. You paste the transcript, you get a read, and then that read is gone. Next week, on the same account, it does not know what it told you or whether you did it.

You can fight this with a running document you paste in each time. That works for about a month, and then the document is four thousand words of context you have to maintain by hand, and you stop maintaining it. The failure is not dramatic. You just quietly go back to not doing it.

Coaching is a sequence. A single brilliant read on a single call is a book review, not coaching.

2. There is no standard underneath it

Ask it to score your call out of ten and it will give you a number. Ask it again in a new window and you will get a different number, from different criteria, invented on the spot to suit the transcript in front of it.

That is not a scoring failure, it is the absence of a scorecard. Without a fixed standard, "an 8" means nothing, because there is nothing it is an 8 against, and you cannot see whether your calls are getting better because the ruler changes length every time you use it.

3. It is generous, and generosity is the enemy here

General-purpose models are tuned to be agreeable. You will get "great question to open with" for a question that was not great. It will find the positive framing because being encouraging is what it was trained to do.

You can push against this, and you should. Add "be harsh, assume I am experienced, do not soften anything" to every prompt and the feedback gets noticeably better. But you are correcting for a bias in the tool with every single message, and the correction fades as the conversation goes on and it drifts back toward agreeing with you.

The uncomfortable version of this: the coaching you most need is the coaching you will least enjoy asking for.

This one is measurable, and we measured it. Six models wrote a cold email from one identical brief, then judged each other blind while playing three skeptical buyers. Eighteen verdicts. The coached model won 17; four of the frontier models won none. The gap was not knowledge, it was that the general models wrote about the seller and the coached one wrote about the buyer. Method and raw judge reasoning are published, so the number is arguable rather than asserted: the benchmark.

4. It cannot hear the call

Text loses the pause before the answer, the drop when you said the price, the moment two people started talking at once. If you are working from your own notes rather than a transcript, it is worse again: you are asking for a read on your own account of what happened, which is exactly the version that already omits what you missed.

How to tell which wall you have hit

• The read is good but you are not acting on it, and next week's call has the same problem → that is the memory wall.
• You have no idea whether you are improving → that is the standard wall.
• It keeps telling you that you did well and you do not believe it → that is the generosity wall.
• It cannot explain why the call felt wrong when the words on the page look fine → it cannot hear the call.

Three of the four are fixable with better prompting for a while. None of them are fixable permanently, because they are properties of a general assistant with no memory of your pipeline and no opinion about what good looks like.

What we built, and who it is not for

Replicate Labs is a coach that holds the other side of all four: it remembers your deals across sessions, scores against a fixed standard so the number means something in week six, tells you what is wrong without hedging, and works from the call rather than your account of it.

It is not for you if you want a rehearsal partner for one call tomorrow. Open ChatGPT, use the prompt above, close the tab. Genuinely. Paying for that would be silly.

It is for you if you have done the prompt thing for a month and noticed you stopped.

The fastest way to see the difference is to bring a call you have already had. Not a scenario, not a roleplay. A real one, with the awkward part still in it.

Get the read on your last call