Can AI Coach Sales Reps From Recorded Calls? Criteria With Review

Can AI Coach Sales Reps From Recorded Calls? Criteria With Review

Learn how to turn recorded sales calls into coaching criteria, using a sample, a written rubric and manager review before any feedback reaches the rep.

•

Want to apply this in your company?

Book a call with Juan →

To analyse sales calls with AI, start with a sample of recordings made with a written notice, a checked transcript and a written rubric with criteria anyone can verify in the audio or the text. AI helps apply the rubric and draft call reviews with quoted excerpts, but the manager confirms every review before talking to the rep. Without these elements, the analysis becomes opinion that looks like data.

Key takeaways

  • Start with the coaching decision: which team behaviour needs to improve and how it shows up on the call.
  • Build a varied sample of reps, lead types and outcomes, not just the longest or most recent calls.
  • Write observable criteria with described levels and real examples, avoiding judgments about personality.
  • Treat AI output as a draft review: it points to excerpts, and the manager decides what becomes feedback.
  • Calibrate the rubric with reviews done by managers before using the analysis with the team.

Why start with the coaching decision, not the recording

Many companies begin by asking AI to review the calls. The result is usually a well-written summary that helps little, because no decision was defined. Before opening any recording, answer three questions: what type of conversation needs to improve, at which stage it tends to stall, and what the rep would do differently if they knew that. The answers define which excerpts matter and which can be ignored.

It also helps to separate coaching from surveillance. Analysing calls to teach is different from monitoring every minute of work to enforce it. That difference shapes how the team receives the tool, and the topic is covered carefully in how to measure AI productivity without surveillance on your team.

How to build the call sample

A poorly chosen sample produces biased coaching. If the analysis uses only the closed deals, it teaches what worked on those days, not what repeats. Choose calls from different reps, different lead types (new contact, follow-up, existing customer) and with varied outcomes. Include short calls too, because the problem often sits in the first minutes, when the rep has not yet understood why the person was contacted.

Before trusting any evaluation, check the transcript. Product names, acronyms, amounts and regional expressions often come out wrong. Ask someone on the team to read excerpts of the transcript next to the audio and note the most common errors. If the transcript gets wrong exactly what the rubric measures, the analysis will judge a conversation that never happened.

Sample checklist:

  • Calls from more than one rep, with at least one case of each lead type the team handles.
  • A mix of outcomes: won, lost and still open calls.
  • Transcript checked in excerpts, with errors noted.
  • Customer personal data removed or masked when the analysis does not need it.
  • A record of who authorised the use and for what purpose.

Turning sales criteria into a rubric

The rubric determines the quality of the coaching. Each criterion must be observable: the rep asked about budget before presenting a price can be verified in the audio, while the rep is confident cannot. For each criterion, describe three levels, always, sometimes and not at all, with what you hear at each level, and add a real excerpt as an example.

A useful rubric usually has few criteria, because every new item dilutes the attention of both manager and rep. Pick the ones linked to the result you want to improve. If the team loses opportunities because no next step is agreed, the rubric should focus on how the call closes, not on how it opens. What was agreed needs to reach the CRM, and that record can follow the process described in how to update your CRM with AI after a sales meeting.

Illustrative scenario: a rubric for after-sales calls

Illustrative scenario: an equipment installation company wants to improve its after-sales calls, which today only confirm appointments and leave recurring questions unanswered. The manager sets four criteria: identifies the reason for the call before suggesting a solution; confirms the warranty information with the customer; logs any unresolved question for a follow-up; ends the call with a date or a named owner. Each criterion has clear levels and an example excerpt for each.

AI applies the rubric to a sample and returns, for each call, the suggested level and the excerpt that supports it. The manager reads it, partly agrees and adjusts. In some calls, the manager notices the rep asked the right question but in a different order, so the level description is changed to accept that variation. That adjustment is what makes the rubric reflect real work, not an ideal model nobody follows.

Calibration and manager review before feedback

Before using the analysis with the team, run a calibration. The manager scores a set of calls without looking at the AI suggestion, then compares the results. Where the gap is large, discuss the criterion, not the person. Many disagreements show that a level description is vague. When the reviews start to match most of the time, the rubric is ready for a pilot.

The most important rule is simple: no feedback goes from AI straight to the rep. The manager receives the draft with excerpts, decides what is relevant, corrects what the analysis misread and chooses one or a few changes to work on that week. Feedback with many points is rarely applied by anyone.

There are also limits the analysis cannot see. Tone of voice and body language are not always in the audio, the lead's context may sit in earlier WhatsApp messages, and an objection may have been handled in another channel. So the manager should read the analysis as a hypothesis about the conversation, not as a verdict.

Privacy, notice and fair use under LGPD

Call recordings contain personal data. Before starting, confirm with legal counsel or the person responsible for privacy in the company which notice will be used, what the stated purpose is and how long recordings are kept. For a similar case covering notice and legal basis, see recording meetings with AI: consent and LGPD in practice.

Use only what is necessary. If the criterion depends only on the rep's behaviour, there is no need to expose all of the customer's data to the analysis. If the tool processes information outside Brazil, the international transfer must be assessed before use. The ANPD publishes guidance on data protection that is worth reviewing. In addition, an analysis that could influence promotion or dismissal requires human review, because LGPD provides a right to review decisions made solely on the basis of automated processing.

Pilot plan in short steps

In the projects we run, the work follows four moves: map the current process and baseline, limit data and actions, pilot a real routine with logging and an agreed metric, then hand over the operation and decide whether to expand, adjust or stop. Applied to calls, this means starting with a small group of reps, measuring whether the chosen behaviour changes in later calls, and logging the cases where the analysis got it wrong. That error log is as useful as the wins.

Suggested plan:

  • Define the coaching decision and the behaviour you will observe.
  • Build the sample and check the transcript.
  • Write the rubric with levels and real examples.
  • Calibrate the rubric with manager reviews done without AI.
  • Run the analysis as a pilot, with mandatory review before each feedback.
  • At the end of the pilot, decide: expand the rubric, adjust criteria or stop.

Frequently asked questions

Do I need to tell reps that their calls will be analysed by AI? Yes, it is best to communicate clearly before starting. Explain the purpose, who will see the analyses, how long they are kept and how reps can ask questions. Without that transparency, the tool tends to be seen as surveillance, and team adoption drops.

Can I use the analysis to judge performance and decide promotions or dismissals? I do not recommend it. The rubric is designed for learning, with a limited sample and manager review. Using it as a performance score changes how the team behaves, and calls do not show context such as territory, portfolio or lead type. Career decisions need other criteria and human review.

Can AI send feedback directly to the rep? It can help draft the text, but sending should go through the manager. Coaching depends on trust, and generated text sent without review often sounds generic or unfair. The manager validates it, adjusts the tone and chooses what to prioritise.

Which tool should I use to transcribe and analyse calls? The choice depends on where the data is stored, who operates the tool and how the company records calls. Before deciding, answer the privacy questions and test transcription quality on real excerpts from your team. The rubric and human review matter more than the tool.

Sources

AI coachingSales callsCall analysisSales managementLGPD