Notes-to-Quote
- Pull the conversation thread + attachments into a clean transcript.
- Match named items against the live price book in Airtable; flag anything not found.
- Draft the quote into the team's existing template, line-items in the same order they always go.
- Land the draft in a review queue with a one-line "why this price" comment per line.
- Manager approves, edits, or rejects. Nothing leaves without her tap.
- ! Do NOT auto-send. Ever. The whole trust model breaks the first time a wrong quote leaves.
- ! The price book has to be the source of truth. If it's stale, the quotes are stale faster.
- ! Names matter. "10x panel" and "10-panel" need to resolve to the same SKU or the manager loses faith fast.
A few things worth saying out loud about how this play actually got built:
The smallest version we could ship
The first version was not the version above. It was a Loom of me dropping a single thread into Claude with the price book as plain text in the prompt, and asking for the quote in the team’s template. Manager watched the Loom, said “do that, but for real.”
We shipped the cheap version in three days. It was wrong about 30% of the time. The manager fixed those 30% in seconds and we logged what she changed. Two weeks later, the system was wrong less than 8% of the time — not because we tuned a model, but because we kept fixing the price book every time it surfaced an ambiguous SKU.
The trust unlock
The single most important design decision: the system has no send button. It only stages drafts. The manager’s role didn’t shrink — it sharpened. She stopped typing and started judging. She’s faster, but more importantly she’s not exhausted by 4pm.
Where it could go wrong
If you skip the human review and wire it to auto-send “for the easy ones,” the first wrong quote that lands at a real customer kills six months of compounded trust. Don’t.