Summarize Sales Call Transcripts into CRM Notes

Tier: Core · Hands-on time: ~30 minutes · Tools: Claude, ChatGPT, Google Sheets

The problem with CRM notes isn't that reps won't write them — it's that every rep writes different ones. One logs objections, another logs vibes, a third logs "good call, following up". A fixed extraction schema run over the transcript makes every call comparable, and comparable notes are what make forecasts arguable with evidence.

When to use — and when not to

• After every recorded discovery or demo callthe transcript exists (Zoom, Meet, Teams, or a notetaker) and the CRM field is blinking.
• Backfilling a neglected pipelinea quarter of recorded calls, zero notes, and a pipeline review on Friday.
• Handoversa rep leaves or a deal moves from SDR to AE, and the next owner needs the deal's factual history, not a folder of hour-long recordings.

And when not to:

• The call wasn't recorded with consent. Recording and transcription consent rules vary by region and company — if the transcript's provenance is doubtful, don't process it, full stop.
• Your CRM already has native AI call summaries. If your stack (Gong, HubSpot, Salesforce Einstein) writes structured summaries into the record, use those and spend this recipe's effort on the schema-tuning idea in "Make it yours".
• The call was a negotiation with legal-sensitive content. Pricing concessions and contract terms need the rep's own careful notes, reviewed, not an automated pass.

The recipe

1. 1.

   Define the schema once (5 minutes, once ever). Agree the fields your pipeline review actually uses. A solid default: pain, current solution, budget signal, timeline, decision process/people, objections, competitors mentioned, next step (with date and owner), deal risk. Write them down — this exact list goes in every prompt, and its stability is the whole value.

2. 2.

   Run the extraction with quote-anchoring rules. Paste the transcript with this prompt:

   Copy

   You are a sales operations assistant. Extract CRM notes from this
   call transcript into EXACTLY these fields:
   
   - Pain: the problem in the prospect's own words (short quote).
   - Current solution: what they use today.
   - Budget signal: anything said about money, quoted verbatim.
   - Timeline: any dates or urgency signals.
   - Decision process: who decides, who else is involved, stage.
   - Objections: each concern raised, with the prospect's wording.
   - Competitors mentioned: names only.
   - Next step: the SPECIFIC commitment made — what, who, by when.
   - Deal risk: one sentence, based only on what was said.
   
   Rules:
   - Every field must be supported by the transcript. If the call
     didn't cover a field, write "Not discussed" — never infer.
   - Quotes verbatim, with the speaker. Never put words in the
     prospect's mouth.
   - Distinguish what the PROSPECT said from what the REP claimed.
   - Keep the whole output under 200 words. It's a CRM note, not
     an essay.
   
   TRANSCRIPT: [PASTE]
   

   A good output looks like this:

   Copy

   Pain: "we spend two days a month fixing the export by hand"
   (Maria, Ops lead).
   Current solution: manual Excel process + retiring legacy tool.
   Budget signal: "under $20k we don't need a committee" (Maria).
   Timeline: wants pilot before fiscal year end (Oct 1).
   Decision process: Maria recommends; CFO signs >$20k; IT security
   review required — Not discussed who runs it.
   Objections: data residency ("has to stay in-region"); migration
   effort.
   Competitors mentioned: Flowline.
   Next step: Rep sends security docs by Thu; demo with CFO week of
   the 18th (rep to propose dates).
   Deal risk: security review timing could push past fiscal year end.
   

3. 3.

   Verify the load-bearing fields (5 minutes). You were on the call — check the three fields money depends on: budget signal, timeline, next step. Confirm quotes are real (search the transcript for a phrase). Models are most tempted to "helpfully" upgrade a vague "maybe next quarter" into a firm timeline; that upgrade is how pipelines rot.

4. 4.

   Paste into the CRM — fields, not blob. Map each schema field to its CRM field where one exists; the rest goes into the notes field under its label. Resist pasting the whole thing as one paragraph into "Notes" — the structure is what makes next quarter's reporting possible.

5. 5.

   Batch mode for backfill. For a stack of transcripts, process them one per message in the same conversation ("Same schema. Next transcript:"), then finish with: "Output all calls as one table — one row per call, columns = schema fields — as tab-separated values." Paste that into Google Sheets and you have an instant pipeline-review grid: every deal, same fields, sortable by next-step date.

Make it yours

• Claude for long calls and batches: hour-long transcripts fit easily, and it is disciplined about "Not discussed" when the prompt demands it. Set the schema as the start of a Project so every future paste inherits it.
• ChatGPT with the audio file: if you have a recording but no transcript, transcribe first (voice upload or Whisper-based tools), then run the schema. Never extract from memory of a call — the transcript is the evidence chain.
• Google Sheets as the review layer: the batch table plus a filter on "Next step" containing no date is the fastest deal-hygiene report you'll ever build — every row that shows up is a deal drifting.

Quality check

• Every quote is findable in the transcriptspot-check by searching a phrase.
• "Not discussed" appears where it shoulda note with all nine fields confidently filled from a 20-minute call is a red flag, not a win.
• Next step has all three partswhat, who, by when. "Follow up" is not a next step.
• Prospect statements and rep claims are not blended"they need it by Q4" must mean the prospect said so.
• Under 200 wordsif it's longer, it's a summary of the call, not a CRM note.