How to review sales calls with ChatGPT

Prepare the recording first

1. Get a transcript. Most phone systems produce one; otherwise any transcription service will do.
2. Strip surnames and phone numbers. The model has no need for your customer's personal data.
3. Mark who is the rep and who is the customer, or the review will confuse the two.
4. Add two lines of context: what you sell and roughly what it costs.

You are an experienced head of sales. Below is a transcript of a call between a sales rep and a customer.

Context: [what we sell, approximate price, where the customer came from]

Review the call and return exactly these blocks:

1. WHAT THE CUSTOMER WANTED. One sentence, in their words.
2. WHAT THE REP ESTABLISHED BEFORE TALKING ABOUT THE PRODUCT. List the questions they asked. Mark separately which were about the customer's situation and which were about product features.
3. PRICE. Was it quoted? If yes, on which of the rep's messages. If not, write "never quoted".
4. OBJECTIONS. For each: the customer's exact words, whether the rep asked a clarifying question before answering, and what they answered.
5. HOW IT ENDED. Was a next step with a specific date agreed? Quote it.
6. WHERE MONEY WAS LOST. Maximum three. For each: a verbatim quote from the transcript, why it is a loss, and what to say instead next time.

Rules:
- Give a verbatim quote from the transcript under every finding. No quote, no finding.
- Do not assess tone, politeness or rapport. Actions and words only.
- Do not praise. If the call was good, write "no critical losses" and explain what worked, with a quote.
- Be brief. No preamble, no closing summary.

What this prompt gives you

• a list of the moments where the conversation stalled, with quotes
• an answer to whether price was quoted and when
• visibility into whether the rep explored the situation or went straight to pitching

What it does not give you

• Amounts. The model does not know your margin or deal cycle, so it cannot price the losses.
• A picture of the team. One call is one case. Patterns need twenty or more.
• Comparison between reps. That requires equal traffic and identical criteria.
• Truth without checking. The model misreads context: it calls a question unanswered when the answer was a voice note, and misses that the customer asked to be called back later.

Going from one call to the whole team

1. Run twenty recordings through the same prompt.
2. Put the answers in a table: price quoted yes or no, at which message, next step dated or not.
3. Calculate the percentages. Those, not the individual reviews, show what is breaking systematically.
4. Verify five findings by hand: listen to the recording at the quoted moment.

Then you hit volume. Twenty recordings is an evening; two hundred a month is not. Between February and June 2026 weekly active Codex users in sales grew 41x as people moved from asking for advice to delegating the work, and that is exactly the volume at which running it by hand stops being possible.