85%
Reduction in post-meeting logging effort
100%
Meeting capture rate — up from ~40%
3
Structured output types per note
0
Client data sent to third parties
The problem

Meetings happened. The CRM didn't know.

The system

Say it. Transcribe it. Structure it. Log it.

Four stages. The rep records and moves on; only a two-minute review stands between voice and CRM.

  1. Stage 1

    Capture

    After a meeting or call, the rep records a voice note — phone recorder, WhatsApp, any voice memo app. No new tools or training. The audio is submitted to the pipeline via a simple upload or shared inbox. The rep moves on immediately.

    Friction removed: no typing, no form fields, no CRM login at the point of capture.

  2. Stage 2

    Whisper transcription

    Whisper transcribes the audio on-premise — no audio or transcript leaves the organisation. The information locked inside the recording becomes searchable, structured text.

    Privacy: on-premise STT — client data never reaches a third-party API.

  3. Stage 3

    Open-source LLM classifier

    An open-source LLM reads the transcript and classifies every segment: action items (with owner and deadline if mentioned), CRM field updates (deal stage, next meeting, budget, contact changes), and general notes (sentiment, context, preferences). Output is structured JSON — not a blob of text.

    Triggers Stage 4: structured output sent to the rep for confirmation.

  4. Stage 4

    Rep reviews & syncs

    The rep receives a structured summary: action items with suggested owners and due dates, a preview of CRM fields about to update, and the general notes. They edit anything wrong, confirm, and the CRM is updated. Total time: under two minutes.

    Result: action items auto-assigned; CRM data current the same day.

The outcome

Same reps, same meetings, different discipline.

Metric Before After
Post-meeting logging time 15–20 minutes manual typing per meeting ~2 minutes — review structured output, confirm
Meeting capture rate ~40% — most meetings never logged under time pressure 100% — voice note takes 2 minutes at the car door
Action item tracking Verbal, buried in email threads or forgotten entirely Auto-extracted, assigned with due dates, visible in CRM
CRM data freshness Updated at weekly review — always a week stale Updated same day — pipeline reflects today's reality
Data privacy Notes typed into cloud SaaS — client data on third-party servers On-premise — audio and transcript never leave the org
Logging consistency Variable — depth depended on time the rep had Consistent structure every time — same fields, every meeting
The design principle

The note should take less effort than saying it.

"The note should take less effort than saying it."

The constraint was not capability — reps knew what needed logging. The constraint was friction. Fifteen minutes of structured typing at the end of a full day is fifteen minutes most reps don't have. So it doesn't happen.

The voice note removes friction at the point of capture; the STT and LLM pipeline does the structuring the rep would otherwise do by hand. What's left for the human is judgement — reviewing, correcting, confirming. The open-source LLM wasn't a cost decision, it was a data decision: client meeting content is sensitive, and running classification on-premise means it never touches an external API.

Applied to

Same pipeline. Different conversations.

See what this looks like in your business.

90-minute AI Opportunity Audit. No pitch — just findings you can act on.

Book a call