The problem
The company trains sales teams in its own methodology, but that training decayed as soon as it ended. Reps returned to live calls with no structured feedback, and managers could only review a small sample of conversations by hand. A prospect had asked for a product that could deliver coaching continuously rather than only during training sessions, and a UI-only prototype existed with no working system behind it.
Sales methodology training was delivered once and then decayed, with no mechanism to reinforce it against real conversations.
Reviewing call transcripts manually meant only a small fraction of conversations received any coaching feedback.
Feedback, when it arrived, was generic rather than measured against the company's specific sales methodology.
Reps had no structured answer to the question of what to do next to advance a particular opportunity.
Coaching ignored the history of the account, treating each call as if it were the first conversation with that client.
What we designed
We encoded the client's own sales methodology as a scoring rubric rather than relying on a model's generic idea of good selling, and required every judgement to cite supporting quotes from the transcript so coaching is traceable to what was actually said. We deliberately kept the input to transcripts only for the first release, and were explicit that win-likelihood is a reasoned estimate from the language model rather than a trained predictive model, so that nothing in the product was oversold.
Call scorecard
Discovery depth
“What is driving the timeline on your side?”
Objection handling
“We can come back to pricing later.”
Next step agreed
“Let us book the technical review for Tuesday.”
Recommended next step
Address the pricing objection directly at the start of the next call.
Estimate
Win likelihood: moderate. A reasoned estimate, not a prediction.
Rubric engine
Encodes the client's sales methodology as structured scoring criteria
Transcript analyser
Evaluates a call against the rubric and returns structured output with supporting quotes
Account memory
Feeds summaries of prior calls with the same client into the analysis so coaching reflects deal history
Coaching generator
Produces three strengths, three improvement areas, the recommended next step and key questions for the following call
Signal detection
Surfaces missed buying signals and unaddressed objections
Follow-up drafting
Generates a follow-up email based on what the call actually covered
Coaching dashboard
Presents scores, guidance and win-likelihood with the reasoning behind it
How it works
From trigger to result.
How the system works
7 stages
01Transcript upload
TriggerPasted or uploaded and linked to a client account.
02Methodology rubric
DataThe company's own sales methodology as scoring criteria.
03Account memory
DataSummaries of earlier calls with the same account.
04Scored analysis
AIStructured output. Every score carries a quote from the transcript.
05Coaching outputs
ResultStrengths, improvements, next step, key questions, missed signals, follow-up email.
06Dashboard
ResultScores and guidance against the deal timeline.
07Memory update
LogicThe analysis is summarised and stored for the next call.
Win-likelihood is a reasoned estimate produced by the language model with its justification shown, not the output of a trained predictive model. We stated this explicitly rather than presenting it as a prediction.
Our approach
How we worked.
Prototype audit
Reviewed the existing UI-only prototype and separated genuine requirements from features invented during prototyping.
Approach validation
Established that rubric-based scoring with structured output and quote citation was the correct method rather than rule-based analysis.
Architecture
Designed the backend, data model, account memory and scoring pipeline.
Scope control
Locked the first release to transcript input and reuse of the existing frontend, deferring everything else.
Risk
Identified documentation of the sales methodology as the critical dependency.
The outcome
What the client received.
We audited the existing prototype and identified screens and features that had been invented during design and were not in the actual requirements, preventing the build from inheriting them.
We insisted the coaching be scored against the client's documented methodology rather than a generic model of good selling, which is what makes the output defensible to a sales trainer.
We required supporting quotes for every judgement, so a rep can see exactly what in their call produced the feedback.
We stated plainly that win-likelihood is a reasoned estimate rather than a trained prediction, rather than allowing it to be presented as a forecast.
We identified that coaching quality depends entirely on the client supplying their methodology in documented form, and flagged it as the primary delivery risk.
Questions
What people ask.
Can AI coach sales calls against a company's own methodology rather than generic best practice?
Yes, and this is the difference between a useful tool and a novelty. The methodology is encoded as a structured scoring rubric, and the model evaluates each call against those specific criteria. Without that, you get generic sales advice that any trainer would reject.
How do you stop AI coaching feedback from being vague or invented?
Require evidence. Every score and every piece of feedback must cite a direct quote from the transcript. A rep can then see precisely which moment in the conversation produced the guidance, which makes the feedback both verifiable and actionable.
Can AI predict whether a deal will close?
Not reliably, and it should not be presented that way. A language model can produce a reasoned estimate with its justification shown, which is genuinely useful for prioritisation. A trained predictive model requires a large volume of historical outcome data most companies do not have. We recommend stating which one you are delivering.
What determines the quality of AI coaching output?
The methodology documentation, more than the model. If the sales methodology arrives as a clear written framework, the rubric is straightforward and the coaching is sharp. If it exists only in trainers' heads, that extraction work has to happen first and it is the single largest risk to the timeline.
Does the AI need to see previous calls with the same client?
It should. Coaching on an isolated call ignores where the deal actually stands. Summaries of prior conversations with that account are fed into each new analysis, so the recommended next step reflects the relationship rather than treating every call as a first contact.