# Meeting Notes to CRM Workflow

Turn a raw meeting transcript into a CRM-ready update in one pass: a tight summary, specific field changes, and a follow-up task list. Works with any capable AI assistant (ChatGPT, Claude, Gemini). No plugins, no integrations, no code.

## What you need

- A meeting transcript from your video call tool, or detailed notes. Transcripts work better because the model can quote them.
- The list of fields your CRM actually uses for a deal or contact: stage names, next step, close date, whatever you track.
- Any capable AI assistant. This recipe is plain text and does not care which one.

## One-time setup

1. **Write down your field map.** Open your CRM and copy the exact field names and, for pick-lists like deal stage, the exact allowed values. The prompt below has placeholders in [brackets]; replace them with yours. If the model can only choose from your real stage names, it cannot invent one.
2. **Decide your no-go fields.** Some fields should never be set from a transcript (contract value before a quote exists, for example). Delete them from the prompt so the model never touches them.
3. **Save the finished prompt** wherever your team keeps snippets: a doc, a text expander, or a saved prompt in your AI tool. The value is in reusing the same prompt every time so the output format stays consistent.

## The prompt

Paste this, then paste the transcript at the bottom.

```
You are a sales operations assistant. I will give you a meeting
transcript. Extract what belongs in our CRM. Use only what is in the
transcript. Never infer or invent. If something is unclear or missing,
put it under "Flags" instead of guessing.

Our CRM fields (edit these to match your CRM before using):
- Deal stage (allowed values: [Discovery / Proposal / Negotiation /
  Closed-won / Closed-lost])
- Next step (free text, one sentence)
- Expected close date (only if stated or clearly implied in the call)
- Contacts (name, role, company - only people actually named)
- [Add or remove fields to match your CRM]

Return exactly this structure:

1. SUMMARY - five bullets max: what the meeting was about, what was
   decided, what changed since the last touchpoint.

2. FIELD UPDATES - one line per field above, in the format:
   Field name: new value | supporting quote from the transcript
   If a field should not change, write "no change".

3. FOLLOW-UP TASKS - every commitment made by either side, as:
   [Owner] - [task] - [deadline, only if one was stated]
   Include commitments we made to them AND commitments they made to us.

4. FLAGS - anything ambiguous, contradictory, or missing. Examples:
   the customer's stated timeline conflicts with our recorded close
   date, or a decision-maker was mentioned but never named.

Transcript:
[PASTE TRANSCRIPT HERE]
```

## Run it after each meeting

1. Open your AI assistant, paste the saved prompt, paste the transcript, send.
2. Read the field updates first. Every proposed change should carry a supporting quote. No quote, no update.
3. Copy the summary into your notes field, apply the field changes you accept, and create the tasks.
4. Read the flags. That is where the interesting material lives: mismatches between what the customer said and what your CRM believes.

## QA checklist

Do this every time. It takes a minute.

- Every field update has a direct quote behind it. Spot-check at least one against the transcript.
- Names and companies are spelled the way the customer spells them, not the way the transcript mangled them.
- No task has an owner or deadline that was not actually said. "By Thursday" in the output should exist in the transcript.
- The stage value is one of your real stage names, not a plausible-sounding invention.
- Every flag got a human decision, not a shrug.

## When this outgrows DIY

This recipe is a manual loop: you paste, you review, you copy into the CRM. That is the right shape while meeting volume is low and one person owns the pipeline. It stops being the right shape when transcripts pile up unprocessed, when several reps each run their own drifted version of the prompt, or when you want the CRM updated without anyone remembering to do it. At that point the fix is not a better prompt. It is a system that watches for new transcripts, writes to the CRM through its API, logs every change, and routes only the flagged calls to a human. That is a build, not a snippet.

---

Free recipe from agentclaw (agentclawhq.com). Use it, share it, adapt it.