Case study · Conversation intelligence

Every call heard, not a sample.

Consumer durables · contact centre and field service. Every call scored against the company's own rubric, the revenue moments flagged in the same pass, and the model calibrated line by line against the client's own auditors.

The situation

Quality checked on a sliver, revenue leaking in the dark

A contact centre and field-service network handling a very large volume of calls a month

What we walked into

Quality was scored on a small manual sample; everything outside it was unread. Nobody could answer what customers asked for and didn't get, which agents were drifting from the script, or how often a renewal was on the table and went unmentioned. The calls also arrived in several Indian languages, which most off-the-shelf tooling quietly drops.

What we built

A pipeline that transcribes every call in the language it actually arrives in, scores each one against the company's own audit rubric, and in the same pass marks the revenue moments — AMC and warranty renewal, exchange, spares and consumables. Supervisors get per-agent scores; the full transcript archive becomes searchable in plain English. Where the client had already hand-scored a set of calls, we calibrated against it so machine and auditor can be compared attribute by attribute.

How it's wired

Recording in, scored and searchable out.

A read-only pull from the call platform; nothing written back to it. The whole path runs inside the client's environment.

Source

Call recordings contact centre & field service The existing audit rubric Hand-scored calls where they exist

↓

1 · Transcribe

ASR Indian vernacular languages, not English-only Speaker separation

2 · Score

Against your rubric attribute by attribute Revenue moments AMC, warranty, exchange, spares Calibrated to your auditors

3 · Surface

Per-agent scorecards Flagged calls to supervisors Plain-English search over every transcript

The results

Coverage from a sample to everything.

Figures from the live engagement; client name withheld.

2 in 3

calls carry a genuine revenue opportunity a manual sample never surfaced

48%

AMC capture on the calls the system flagged

~50%

of those opportunities closed with no pitch at all — the customer asked and was answered

100%

of calls audited and searchable in plain English

What changes Agent scores stop being a lottery of which calls got picked, and the audit pass starts paying for itself — the same listen that checks compliance finds the renewals nobody logged. Because a set of calls was hand-scored first, agreement with your auditors is measured on a shared sample every month, not asserted once during the sale.

The stack, and where your data lives

Runs inside your environment.

Built on

Whisper / ASR Vernacular language models LLM scoring self-hosted or your endpoint Vector search over transcripts Your warehouse Your repository

Data boundary Call audio is sensitive, so this is a self-hosted deployment by default — transcription and scoring run on open-weight models inside your VPC or on-premise, and no audio or transcript leaves your boundary. Access is through your own IAM and every action is logged.

Have a contact centre scored on a sample?

In a five-day discovery we run your own calls through the pipeline and show you what the sample has been missing — the roadmap is yours to keep either way.

Book a 30-minute call