The problem
Why this needed to exist.
Recordings preserve a conversation, but useful deal evidence often disappears between the call and the CRM. Generic summaries do not show what was proven, what is missing, or what the seller should do next.
Defining decision
Treat the transcript as the evidence boundary: every finding must point back to the call or clearly state that the information is missing.
Approach
How the system was shaped.
- Designed one structured analysis contract covering every MEDDICC category, discovery quality, objections, risks, actions, notes, and follow-up.
- Required findings to cite transcript evidence and label absent information instead of filling gaps with plausible language.
- Separated analysis, validation, transcription, persistence, and interface layers so each can be tested independently.
Tradeoffs
What the design chooses—and gives up.
- A strict schema creates more predictable outputs but limits free-form coaching.
- Local history reduces infrastructure and privacy exposure while leaving team collaboration and CRM write-back outside the MVP.
Evidence
What can be inspected.
- A complete synthetic sample report is available without an API key.
- Pydantic validates scores, enums, required outputs, and complete MEDDICC coverage.
- Tests mock provider responses and make no network calls.
Limitations
What this does not claim.
- Generated notes remain decision support and require seller review before CRM entry or customer delivery.
- Live audio transcription sends content to the configured provider; consent and retention rules still apply.
Lessons
What carries into the next system.
- Useful AI sales coaching depends less on confident prose than on traceability, missing-evidence handling, and a clear next decision.