How to Turn Sales Call Recordings Into Coaching Notes With AI

Record calls with the customer's permission, get a speaker-labelled transcript from a note-taker (Fathom, Fireflies, Otter, or the recap built into Zoom, Teams or Google Meet), then run it through one fixed prompt that scores the call against your own scorecard and quotes timestamps as evidence. A manager checks the notes, then coaches one change per week.

The scorecard is what makes this work. Ask an AI "how did this call go?" and you get polite, generic praise ("great rapport, clear explanation of services"). Ask it whether the salesperson found the budget before quoting, with the exact line where it happened, and you get something you can coach from. So the first hour goes on deciding what a good call looks like in your business, before anyone presses record.

Follow me on Instagram@sagnikteaches

Write the scorecard before you record anything

A useful scorecard has five to seven behaviours, each one observable in a transcript. "Built rapport" is not observable. "Asked what the couple are most worried about" is. Each behaviour gets a simple 0-2 score: 0 not done, 1 partly done, 2 clearly done.

Connect on LinkedInSagnik Bhattacharya

Here is a filled-in scorecard for a small wedding planning business, for the first consultation call with a couple:

Subscribe on YouTube@codingliquids

Two things are deliberately not scored. Talk ratio (how much of the call each side spoke) is reported as a number for context, because in a consultative sale the planner may rightly talk more when explaining process. And "tone" is left to the human listener, because AI sentiment scores are not reliable enough to coach from.

Let your own won and lost calls shape the scorecard

If you already have old recordings, AI can help you draft the scorecard from evidence rather than instinct. Pick five calls that led to a booking and five that didn't, anonymise the names, and ask for differences in what the salesperson did:

Attached: 10 consultation transcripts. Files starting WON led to a booking;
LOST did not. Compare what the PLANNER did in each group.
List up to 8 behaviours that appear clearly more often in WON calls,
with how many WON and LOST calls show each one and one quote per behaviour.
Do not explain why; just count and quote.

An illustrative answer might say that a dated next step appeared in 5 of 5 won calls and 1 of 5 lost calls, and that asking about the budget before describing packages appeared in 4 of 5 won and 2 of 5 lost. Treat this as a list of candidates, not proof: ten calls is a tiny sample, and a lost call may have been lost on price or date availability, whatever the planner said. Keep the behaviours that also match your experience, and drop anything that looks like a coincidence, such as "mentioned the weather".

Recording with consent without making the call awkward

Recording rules differ from place to place. In some, one party's consent is enough; in others, everyone on the call must agree. The habit that works everywhere is to ask, clearly, at the start, and to note that you asked. Something like:

"Before we start, I record these calls so I can send you accurate notes afterwards and so our team can keep improving. Is that alright? Happy to switch it off if you'd rather."

Most people say yes. If someone says no, stop the recording and take notes by hand. Put the same sentence in your booking confirmation so nobody is surprised, and check with a solicitor if you record phone calls routinely or work across different jurisdictions. Bots that join video calls announce themselves in the participant list, which helps, but it does not replace asking.

Picking a note-taker for coaching, not just minutes

Almost any note-taker produces a summary. For coaching you need three things: speaker labels you can correct, timestamps, and a transcript you can export or copy. Prices below are list prices per user in USD, checked in September 2026; confirm on each vendor's pricing page before you buy.

If you see older guides mention "Zoom AI Companion", Zoom retired that brand in June 2026; the same features are now simply named for what they do, such as meeting summary. For a fuller comparison of these tools for general meetings, see AI meeting note-takers compared for small teams.

Built-in summaries are fine for minutes, but they follow the vendor's template, not your scorecard. The practical setup for most small teams is: the note-taker records and transcribes, and a chat assistant does the coaching analysis using your prompt. On a business plan, keep the prompt and scorecard in a shared ChatGPT or Claude Project so every manager runs exactly the same instructions.

The coaching-notes prompt

This prompt does most of the work. Paste the scorecard in once, as the project's instructions, and then paste each transcript as a new chat.

You are reviewing a recorded first consultation for a wedding planning business.
Use ONLY the transcript provided. Speakers: PLANNER and COUPLE (may be two voices).

For each behaviour in the scorecard:
- Score 0, 1 or 2 using the definitions.
- Quote the exact line(s) as evidence, with the timestamp.
- If the behaviour did not happen, write "Not observed" and say where in the
  call it could have happened (timestamp).
Never paraphrase a quote. Never invent a line. If the speaker label looks
wrong, say so instead of guessing.

Then write:
1. One thing the planner did well, with the quote.
2. The ONE change that would most improve the next call, and a suggested
   line they could use instead.
3. Talk ratio as a percentage, for context only.
4. Any factual detail from the couple that the proposal must include
   (date, venue, guest count, budget range, priorities).
Keep it under 300 words. Plain text.

Two instructions carry the weight. "Quote the exact line with the timestamp" lets the manager check any claim in seconds. "Not observed, and where it could have happened" turns a missing behaviour into a specific moment to coach.

A sample output, and what the manager corrected

Here is an illustrative output for a 28-minute consultation:

Agenda and time check: 2. [00:41] "We've got about half an hour; I'll ask a few
questions first, then show you how we work."
Qualifying facts: 1. Date [02:10], venue [02:55] and guest count [03:30] asked.
Budget range not asked before packages were described at [09:12].
Priorities and worries: 2. [05:48] "What's the bit you're most dreading?"
Process explained: 2. [12:20-15:05] first three months described step by step.
Budget concern handled: 2. [21:40] Couple: "It's more than we'd pictured."
Planner linked fee to "not having to manage suppliers yourselves" [21:52].
Next step with a date: 0. Not observed. Call ended at [27:30] with "I'll send
something over." Could have set a date at [26:50].
Did well: the "most dreading" question, which surfaced the family-seating worry.
One change: before ending, propose a date: "I'll send the proposal by Thursday.
Could we speak Monday evening to go through it?"
Talk ratio: planner 58%, couple 42%.
Proposal must include: 14 June, venue booked, about 50 guests, seating plan
help with separated parents.

When the manager checked the quotes against the recording, two corrections came up. The guest count was "fifteen", not "fifty": the couple were planning a small wedding, which changes the whole proposal. Transcription tools often confuse numbers that sound alike, so any number that will go into a quote should be checked against the audio. Second, the "budget concern handled" score was generous; the planner linked the fee to a priority, but also said "we can probably knock a bit off", which the prompt missed because it stopped reading after the first good line. The manager changed that score to 1 and added a line to the prompt: "Check the whole call for any discount offered."

That second fix is typical. Your prompt improves every week for the first month as you notice what it misses. Keep a short change log so everyone runs the current version.

Turning notes into a 20-minute coaching session

Notes that nobody discusses change nothing. A simple weekly format works for small teams:

1. Before the session (5 minutes, the salesperson). Read the AI notes for your calls that week. Mark anything the transcript got wrong.
2. Start with the strength (3 minutes). Play the timestamped clip of the thing done well. People repeat what they hear praised specifically.
3. Pick one change (5 minutes). Choose the lowest-scoring behaviour that appeared in more than one call. One call is an accident; two is a habit.
4. Practise the moment (7 minutes). Read the transcript excerpt, then role-play the better version twice.
5. Set the target. "On next week's calls, agree a dated next step every time." Next week's notes show whether it happened.

A running tally per behaviour makes progress visible. For the planner above, over four weeks, it might look like this (illustrative):

A three-planner business over six weeks

Take an illustrative wedding planning business: the owner plus two planners, running about ten first consultations a week between them. Before AI, the owner listened to two recordings a week, at roughly 45 minutes each including notes. That is 90 minutes of the owner's week covering one call in five, and coaching conversations were based on whichever calls happened to be picked.

With the process above, every call gets AI notes within minutes of ending. The owner spends about five minutes per call checking quotes against the audio at the flagged timestamps (50 minutes a week) and 20 minutes coaching each planner (40 minutes). Total: 90 minutes, the same as before, but covering all ten calls instead of two, with evidence for every point.

Costs are modest. Two planners and the owner on Fathom Business at the annual rate would be $75 a month; on Otter Pro annually, about $25 a month, with the scorecard analysis done in a chat assistant the business already uses. The setup took roughly three hours: an hour for the scorecard, an hour to test the prompt on five old recordings, and an hour to agree the coaching format.

What to expect in results is harder to promise. In this illustration, the share of calls ending with a dated next step went from about four in ten to nine in ten within a month, because that was the behaviour coached first. Whether bookings follow depends on price, season and fit. Compare consultation-to-booking rates over a full quarter against the same quarter last year before drawing conclusions.

The same method for shorter, simpler sales calls

A personal trainer's free consultation

For a personal trainer selling 12-week packages, the scorecard shifts: ask about injuries and medical history before recommending anything, ask about schedule constraints, show one example session, and agree a start date. In an illustrative review of three calls, the notes showed the trainer recommending the 12-week package at around minute six, before asking about a knee injury that came up at minute eleven. Nobody had noticed on the calls themselves. The one change for that week was a single line: "Before I suggest anything, is there any injury or condition I should know about?"

A driving school's phone enquiries

A driving school's sales calls are three-minute phone enquiries, which makes a six-behaviour scorecard overkill. Cut it to three: asked about experience so far and any test date, offered a specific first-lesson slot, and took a mobile number. Instead of reviewing each call, the office manager pastes a week's transcripts into one chat and asks for a tally: "For each call, did we offer a specific slot? Quote the line." In one illustrative week, 11 of 19 enquiries ended with "I'll check availability and call you back", and five of those never booked. The fix was a shared view of instructor availability during calls, not a better script.

Where AI call analysis misleads you

• Numbers and names. "Fifteen" and "fifty", "Sian" and "Shaun", "the 14th" and "the 4th". Check any figure that goes into a quote or proposal against the audio.
• Swapped speakers. When two people talk over each other, the transcript can give the couple's line to the planner. The prompt's "say so if the label looks wrong" instruction helps, but a human check at each quoted timestamp is still needed.
• Inconsistent scores. The same transcript can come back scored slightly differently on two runs. Fixed definitions, an example of a 0, 1 and 2 for each behaviour, and a human who owns the final score keep this in check. Lowering the model's randomness is not a dependable fix on current models.
• Coaching to the scorecard. If a behaviour scores well but bookings fall, the scorecard is measuring the wrong thing. Review it every quarter against calls that booked and calls that didn't.
• Surveillance feel. Staff who think recordings are for catching them out get guarded and scripted. Say plainly what the notes are for and who sees them.

Privacy needs its own line. Sales conversations contain budgets, family situations and sometimes health details. Use a business plan that does not train on your content by default, or switch off model training in a consumer account's privacy settings, and set automatic deletion for recordings once coaching is done. Reviewing AI call transcripts for quality and compliance goes further on what to keep and what to strip out.

Checking the process is paying off

After a month, answer four questions. Are notes produced for every recorded call without anyone chasing? Has the lowest-scoring behaviour improved across the team? Do salespeople find corrections in the notes less often than in week one (a sign the prompt has matured)? And is the owner spending the same or less time on coaching than before? If the answer to the last one is "more", cut the review step back to flagged timestamps only.

The same transcripts can do more once the coaching routine is settled. The next steps each call agrees can flow into your CRM (see updating your CRM automatically after sales calls), and recurring objections are good raw material for call scripts that don't sound scripted. Before any AI summary goes to a customer as "notes from our call", run it through a five-minute fact-check routine.

Questions about coaching from recorded calls

Do I need a dedicated sales coaching platform?

Not for a team of two to five. A note-taker that produces speaker-labelled transcripts, plus a chat assistant running your scorecard prompt, covers the coaching. Dedicated conversation-intelligence tools earn their cost when you have enough calls that trends across the team matter more than individual notes, or when you want scores logged automatically against deals in a CRM.

Should staff see their own AI scores?

Yes, and before the manager discusses them. Seeing the notes first lets the person check for transcription errors and come to the session with their own view. Keep the scores for coaching only. If they start feeding pay or discipline decisions, people learn to perform for the scorecard, and the calls get worse even as the numbers rise.

Can this work for consultations held in person?

Yes, with care. Some note-takers record in-person meetings from a phone or laptop, and Google Meet's note-taking can now handle in-person meetings on eligible plans. Ask permission at the start as you would on a video call, place the device where both voices are clear, and expect more speaker-label mistakes than on a video call.

How long should I keep sales call recordings?

Only as long as you need them for coaching and follow-up, often a few weeks. Keep the coaching notes, which contain the quotes you need, and delete the audio and full transcript on a schedule. Set the retention period in the tool itself rather than relying on memory, and check with your data-protection adviser if calls contain sensitive personal details.

Further reads

• How Much Does an AI Note-Taker Cost per Seat? — What note-takers really cost once the whole team has a seat.
• AI Call Summaries: Log Every Phone Enquiry in Your CRM — Log every phone enquiry in your CRM without typing it up.
• How to Automate Sales Follow-Ups With AI Without Being Pushy — Act on the next steps your calls agree, automatically.
• How Small Businesses Use AI in Sales: 10 Real Examples — Other ways small firms use AI across the sales process.
• Zero Data Retention: What It Means When You Choose an AI Tool — What retention settings mean before you upload call transcripts.
• AI vs Human Transcription: Which Is Worth Paying For? — When a human transcript is worth paying for instead.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Fathom pricing page; Fireflies pricing page and its July 2026 pricing guide; Otter.ai pricing; Microsoft Learn (Teams Premium licensing and intelligent recap); Google Workspace Help (Take notes for me).