How to Review Sales Calls Without Gong (3-15 Reps)

You need four things to review sales calls without Gong: transcripts of your calls, a scorecard of 5 to 8 yes/no questions, a weekly sample of calls, and a fixed weekly slot for the sales manager. By my estimate the slot costs about 25 minutes per rep per week once the transcripts exist: checking the verdicts on two calls plus one 15-minute one-to-one. A language model can read each transcript first and give a verdict on every question. Each verdict comes with lines from the call that you can check. The manager decides what to coach.

Gong's pricing page lists no prices. It says licenses are priced per user and a platform fee comes on top. To get a number you fill in a form. Apart from the transcripts, the routine below needs only a document for the scorecard and a language model for the first pass.

First, count how many of your calls anyone listens to. Six reps making 8 calls a day produce 240 calls a week. A manager who listens to three of them has heard about 1% and coaches from a few stories. The steps below raise that to 12 calls a week, 5% of the total, each with quotes as evidence.

Step 1. Pick one behavior to change this month

Choose one step of your sales process that loses you the most deals when reps skip it. If you have no written playbook, pick the habit you most want to hear on calls. An example is booking the next step before the call ends.

You score 5 to 8 things but coach one. One behavior per month keeps the coaching specific.

Step 2. Write a scorecard of 5 to 8 questions

Each row is a step of your sales process turned into a yes/no question. Every question needs a pass condition you can check from the words on the call. The last column says which part of the call to read when the step is missing. Here is an example. Adapt it to your playbook.

Leave out tone, energy and rapport. They are hard to score the same way twice, and scoring them from audio raises an EU legal question. Article 5(1)(f) of the AI Act prohibits AI systems that infer the emotions of a person in the workplace, except for medical or safety reasons. The Commission's guidelines on prohibited AI practices give the example of a call centre using webcams and voice recognition to track employees' anger, and say that inferring emotions from written text is outside the prohibition because it is not based on biometric data (C(2025) 5052, paragraphs 250 and 251). The guidelines are not binding, and this is not legal advice. A scorecard about what was said, applied to a transcript, infers no emotions.

Step 3. Get recordings and transcripts

Zoom can make an audio transcript of cloud recordings. Zoom's help page says you need a Pro, Business, Education or Enterprise account with cloud recording enabled. An account owner or an admin with the right privileges turns it on in the account or group settings. The same page says audio transcription supports English only and applies only to cloud recordings.

Google Meet offers transcripts on some Google Workspace editions.

For phone calls and for other languages, export the audio from your dialer or CRM and use a transcription service. A developer can run an open-source speech model instead.

Tell the other side at the start that the call is recorded and reviewed with AI tools. Consent rules differ by country, and in the US they also differ by state. Check yours with a lawyer.

Step 4. Sample

Review two calls per rep per week. Take one from a deal that stalled or was lost. Take one from a deal that was won. With six reps that is 12 calls.

Step 5. Let a model do the first pass

Paste the transcript and your scorecard into a language model with the instruction below. The transcript needs timestamps and speaker names. It also carries customer names and prices, so use a business plan or an API that offers a data processing agreement. Under GDPR, Article 28 requires a contract with any provider that processes personal data for you.

Below are a sales call transcript and a list of steps from our sales process.
For each step, give one of three verdicts.
FOUND: the rep did it. Copy the exact line, with timestamp and speaker.
MISSING: you searched the whole call and no line shows it. Say which part of the call it belongs in, and copy the three lines nearest to that place.
UNCLEAR: the transcript is garbled or cut off where the step would be.
Copy lines word for word. Do not paraphrase.
A FOUND verdict must rest on a line the rep said. If the step also needs the customer's answer, copy the customer's line as well.
For the first MISSING step, write one sentence the rep could say next time.

Then check each verdict.

1. FOUND. Search the transcript for the quote. If it is not there, discard the verdict. Models drop filler words and change case, so search for three distinctive words of the quote before you give up. If a quote runs across two speakers, search each part. Check who spoke: a customer saying "let's meet Thursday" does not prove the rep booked it.
2. MISSING. The model says nothing was said, so there is no quote to search for. Read the three lines the model copied. Then search the transcript for two or three words that would signal the step, such as a weekday, "calendar" or "invite" for the next step. If you find them, read the lines around them: the model may have missed the step.
3. UNCLEAR. Listen to that part of the audio.

These checks catch invented quotes and some wrong MISSING verdicts. They cannot tell you whether a real quote means what the model says it means. Read each remaining quote in context before you coach.

Here is a MISSING verdict from a made-up example call. A person wrote that review by hand to show the format, so a model's output will look plainer. At 04:12 the customer says "we need it installed by the 15th". The rep sends a PDF and says "let me know". Verdict on the next step: MISSING. Line to try: "The 15th works if we confirm by Friday. Can we lock 15 minutes on Thursday at 10?"

Step 6. Run a weekly rhythm

Monday: read the verdicts, check the quotes and pick one fix per rep. Then hold a 15-minute one-to-one with each rep. Show the quote and the line to try.

The next week, look at the same question on that rep's new calls. Two calls a week is a small sample, so judge progress over several weeks. When you disagree with a verdict, change the pass condition in the scorecard so the next review follows your rule.

Time. Reading the verdicts and checking the quotes takes about 5 minutes per call: 60 minutes for 12 calls. Six one-to-ones of 15 minutes add 90. The total is 2.5 hours a week for six reps, about 25 minutes per rep. I have not timed this on a team. The 5 minutes is a guess for reading six verdicts, searching the quotes and reading the lines for each missing step. Getting the 12 transcripts into the model and listening to audio are extra. Time your first three calls and use your own number.

Pitfalls

• Too many questions. With more than 8, the verdicts take longer to check and the one fix gets lost among them.
• A scorecard the reps have not seen. Show it to them before the first review. They should know what is checked.
• Poor transcripts. Phone audio and people talking at once cause errors, and a quote can match a wrong transcript. Listen to the audio behind five verdicts before you trust the pattern.

When a tool starts to pay off

Sampling stops working when you want every call reviewed or when the weekly review takes more time than you have. That is when to ask Gong and other tools for a price and compare it with the hours you spend.

If you want someone else to do the first pass

I run Oyendo, which checks recorded calls against your own playbook. It started as a call review system I built for a manufacturer's sales team, and its engine has the same shape as the routine above. A model reads the transcript, scores each step of the playbook and quotes the line. Then code checks the result: it compares each quote and timestamp with the transcript, confirms that the right person said the line and looks for contradictions inside the review, such as a step marked missed next to a high score. A second model pass rechecks the calls the code flagged and a sample of the rest. I test the checks with Claude by planting invented quotes in reviews and counting how many get caught.

Building it changed this routine in three places. Short replies from the other person split quoted sentences, and a strict text match flagged real quotes as invented, so the check now matches each quote against the whole transcript. Sales reps at the manufacturer rejected the first version over one score they found unfair, and accepted it once each score showed the quote from their own call. And with a handful of calls a week, the share of calls that pass a step moves a lot from chance alone, so Oyendo shows a rep's progress only after a large change across many calls.

To see a review of your own call, send one recording at oyendo.eu/review. The form asks you to confirm that the callers were told the call is recorded. Oyendo is built for sales teams of 3 to 15 reps. On a 30-minute call we go through your draft scorecard question by question, and I tell you which questions a model can check from a transcript and which it cannot. Book at cal.com/salta.team or write to artem@tastare.com.