How AI Can Analyze In-Store Sales Conversations: Capture, Transcribe, Score, Coach

The pipeline in one view

Consent and capture on the advisor's phone

Nothing starts until the customer knows. The entrance and the counter carry signage, and the advisor says a short line before tapping record. The tap is the only thing the advisor does; there is no lanyard device to charge or counter microphone to point. The recording is made on the advisor's own phone or a store handset, with the store, zone and time attached.

Indian shop floors lose signal in basements, back corners and at 6 pm on a Saturday. Capture is offline-first: audio is written to the phone in chunks as it records, queued locally, and uploaded when a connection returns. A conversation recorded with no network at all reaches the server intact an hour later. This matters more than any model, because a pipeline that drops the conversations from the busiest hour has analyzed the wrong sample.

Transcription in Hindi, English and Hinglish

A real store conversation switches language mid-sentence: the price in English, the doubt in Hindi, the product name in whatever the advertisement used. The transcription model has to handle that mix rather than one language at a time, and it has to keep speaker turns so the advisor's question and the customer's answer are not run together. Borentis transcribes Hindi, English and Hinglish, and adds regional languages as the network needs them, with an honest word error rate reported per language rather than a claim of every language on earth. Once the transcript exists, the audio is deleted.

Scoring the playbook step by step

1. The retailer writes the playbook as named steps: greeting, discovery, demonstration, offer, objection, close, contact. Each step has a plain description of what counts as done.
2. For each step, the AI reads the transcript and asks one question: did this happen, and where. It answers with a score and the transcript line that earned it, or the absence that lost it.
3. Offer accuracy is scored against this month's scheme: the amount, the terms, the eligibility. A wrong number is a fail with the wrong line quoted.
4. Objections are marked where the customer raised them, and marked answered or unanswered by what the advisor said next.
5. The conversation gets a total, but the total is the least useful number. The step scores are what a coach uses, because each comes with its evidence.
6. A manager who disagrees with a score can read the line and dispute it. Scores that cannot be traced to a line are not scores; they are opinions with decimals.

Clustering, coaching and the follow-up draft

Across a network, the same objection is raised in a hundred wordings. The AI groups them: the exchange offer that sounds low, the EMI that needs a card the customer does not have, the rival's price in the next lane. Each cluster is ranked by store and week, with three example lines, so head office can answer one objection centrally rather than a hundred times.

Coaching starts from the most-skipped step per advisor. The task is short: here is the step, here are two lines from your conversations where it was missed, here is the best example from the network. The trainer assigns it; the advisor sees their own score before anyone else does.

Follow-up starts from what the customer said. If they wanted the red variant by the fifteenth, the draft says that, in the customer's language, addressed to the number they gave. The draft sits in the advisor's queue. A person edits it and sends it. Nothing goes to a customer automatically.

What the AI does not do

• It does not record without consent, and it does not keep audio after transcription.
• It does not identify customers. A phone number enters the record only when the customer gave it for follow-up.
• It does not send messages. Every follow-up is a draft a person reviews.
• It does not decide incentives or discipline. Scores are evidence for a manager, not verdicts.
• It does not replace the CRM. It feeds the CRM the leads and the reasons.