All field notes

Question ·

Can AI coach sales reps from their recorded calls?

AI can coach reps from recorded calls when it scores each call against the team's own scorecard, checks every quote and orders calls the way experts do.

AI can coach sales reps from their recorded calls when it scores each call against the team's own scorecard, quotes only what was said, and has been tested on whether it orders calls the way trained experts do.

The short version

The leading AI models can read a call transcript, find the moments that matter and explain what to do differently. The hard part is consistency. A model asked to score the same call twice can give different answers, and a model that sounds confident can be wrong about what happened on the call.

We tested this at length. Our research program, run with PeerLab's blind human marking, ran more than 5,000 tests and more than 250,000 scoring decisions across 15 AI models, scoring real enterprise discovery roleplays against a 20-skill capability scorecard. The full findings are in the field guide Can AI judge a sales call?. The answer was yes, with conditions.

What makes AI call coaching trustworthy

Four things separated the setups that worked from the ones that did not.

From a recorded call to coaching a rep can use
  1. 01The rep's callA transcript or a connected recorder
  2. 02Score against the team's scorecardYes/no checks per skill, with marker reasons
  3. 03Check every quoteFeedback points at words the buyer said
  4. 04One change for the next callTalked through by voice or on a call

It scores against the team's scorecard with clear checks. The final method uses a checklist model that answers yes/no questions for each skill, carrying the reasons trained assessors gave for their marks. Only 1.7% of its scores changed when the same calls were scored again (the checklist model alone). Combined with a second, blind whole-call score, the method ranked calls close to the order trained experts did.

It learns from real marker reasons. Adding the reasons trained assessors gave for their scores was the biggest single gain in the whole program. Textbook examples added nothing. Rewriting the prompt was risky: one redesigned system prompt moved up to 40% of scores on a re-run.

It is judged on ranking. A scorer that never read the transcript landed a lower average error than the best real model. Average error flatters a lazy scorer. The useful test is whether the AI puts calls in the right order, from strongest to weakest.

It checks its own quotes. In cheap evidence extraction, a fifth of extracted quotes were not in the call. Coaching that quotes the rep or buyer has to verify every quote against the transcript, or it will eventually invent one.

Model choice matters less than people expect. The most expensive model finished eighth. And newer models were not more generous: each generation marked harder than the last.

An example

This example is illustrative. A rep at Arrowmere, a 600-person logistics firm, brings the transcript of a discovery call to an AI sales expert. The expert scores it against the team's own scorecard and flags one moment: the buyer said late deliveries "cost us a lot," and the rep moved on to the demo.

The feedback quotes that line, checked against the transcript, and suggests one question for next time: roughly what did late deliveries cost last quarter? The rep then talks it through with the expert by voice on a call, practicing the question twice before the next meeting. The manager sees the same moment ahead of the 1:1.

What AI coaching from calls cannot do

AI coaching from recordings depends on the recording. It works from the transcript the rep brings or a connected recorder, so calls that are not recorded are invisible to it. It also cannot replace a manager's judgment about the person: what they are ready to hear, and what is going on outside the call.

It is also different from conversation intelligence. A call recording platform records and searches calls; an AI sales expert coaches the rep on what to do next in the team's own method. See AI sales coaching vs conversation intelligence for the comparison, and the AI call scoring guide for how to evaluate any scoring tool.

Where Replicate Labs fits

Replicate Labs builds AI sales experts that coach in a team's own method. Caty works with any method, from MEDDPICC to a house method. Start a call with Caty in one click, talk the deal through by voice or video on a call, and bring the recording or transcript for feedback. More on how it works, or talk to us.

FAQ

Can AI score sales calls accurately?

AI can score sales calls consistently when it uses clear yes/no checks against the team's scorecard and reasons from trained assessors. Our tested method ranked calls close to the order trained experts did. See can AI score a sales call.

Is AI call coaching better than coaching from a manager?

AI call coaching and manager coaching do different jobs. AI can review every call and find the moments quickly, while the manager knows the rep and decides what to work on. They work best together.

What does AI need to coach from a recorded call?

AI needs a transcript of the call, the team's scorecard or method, and ideally the reasons trained assessors gave for their marks. A connected call recorder saves the rep from uploading each call.

Do newer AI models score sales calls more generously?

Newer AI models did not score more generously in our research. Each generation marked harder than the last, and the most expensive model finished eighth of the models tested.

Put an expert on your deals.

Talk to Keenan about a live deal for free, or talk to us about experts for your team, your product or your coaching business.

Talk to us