Turn Recorded Calls Into CRM Updates and Deal Awareness

Any recorder that produces a transcript works. Gemini meeting notes, Zoom transcripts, Fireflies.

Raw speaker-labeled transcripts produce the best extractions. Notetaker summaries still work, and the workflow lowers its confidence scores to match, because secondhand notes are already an interpretation.

The baseline this replaces is familiar. Updating records after each conversation takes a rep about thirty minutes per client, so the records go stale and deal awareness lives in whoever happened to be in the room.

No quote, no item. The workflow extracts what was said. It never invents. Every item carries the verbatim words that created it, so anyone can audit any entry back to the source in one click.

Talk the meeting explicitly deferred becomes a follow-up, never a task. Here is the shape of one extracted item.

{
  "title": "Send pricing summary to their CFO",
  "owner": "us",
  "estimate": "S",
  "commitment": "committed",
  "scope": "in",
  "tier_proposed": "review",
  "source_quote": "Can you get the pricing
    summary over to our CFO by Friday?"
}

Watch for the item you cannot quote. Cut it. A phantom commitment costs more than a missed one.

Meetings wander. The workflow judges two things separately for every item, because they are different facts. Was it agreed, and is it inside the agreed scope. The route follows from the pair.

• Committed, in scope. An owner and a yes. A work item this cycle.
• Committed, out of scope. Agreed in the room. It proceeds, and it always carries a scope note, so scope stays governed.
• Explored. Musing with no owner and no date. Parking lot, no task.
• Non-goal reversal. Something already scoped out came back. Flagged with the original exclusion so a person re-decides.

The tell is an owner and a yes. Enthusiasm alone is explored. When commitment is unclear, choose explored and let a person promote it.

The parking lot is expansion signal, and promotion out of it is a human act. Raising a twice-mentioned topic back later is where future work comes from.

Owner is the first fork. Items owned by your side are candidates for execution. Items owned by the other side are tracked as waiting-on entries and never executed.

That single rule keeps the workflow from quietly doing work that belongs on someone else's plate, and it lets tomorrow's brief say "waiting on Dana since the 14th."

Rules decide the tier. The stricter rule always wins. Reversible, scoped, testable work is eligible to run unattended. Everything else waits for review.

Deterministic overrides sit above the model's proposal, so a misclassification can never grant itself more autonomy. Large items wait. Low-confidence items wait.

And anything a person outside your team will see, in their words or yours, waits for a person on your side. Always, regardless of size.

Watch for the tier assigned by vibes. If you cannot name the rule that decided it, the rule layer is missing.

Approved output lands where your team already works. Record updates drafted into the CRM. Tasks with an owner, a size, and the source quote. A channel brief a teammate reads in a minute.

When the meeting updates prior work, the workflow matches the existing record instead of creating a duplicate.

The front end is a door, and you should name your own. Slack teams get a Slack brief, Microsoft teams get Teams.

Supervised first. Graduation is measured, never assumed. At the start, a person reviews every extraction before anything writes to a live system.

Score each pass on five failure modes. Invented items, missed asks, wrong owners, wrong routes, and unfaithful quotes, because a fabricated or trimmed quote fails the pass even when the item is otherwise perfect. The workflow graduates to writing on its own only after consecutive clean passes, and only for the tiers the rules already allow.

Trust in this design is a property you measure, with a date attached.