Call Intelligence
AI Call Summary Prompts: 5 Templates That Produce Usable Notes
Aug 28, 2026
Part of AI call summary software
Five AI prompts for call summaries, from a one-line recap to a CRM-ready record, with the rules that stop a summary drifting into fiction.
Most AI prompts for call summaries fail in the same way: they ask for a summary. The model obliges with three paragraphs that read well and leave out the one detail somebody will need next week, the time the customer said they would call back, the price that was quoted, the thing they were promised. A useful summary prompt does not ask for prose. It asks for specific fields, in a fixed order, and tells the model what to write when a field is missing.
The five templates below are the ones we start from when we configure AI call summary software for a new deployment. Each is short enough to paste, and each is built around the same three rules: name the fields, forbid invention, and keep the summary shorter than the call.
What a Good Call Summary Prompt Has to Do
- Name the fields: caller, reason, outcome, next step, owner, deadline. A model given fields fills them; a model given "summarise this" writes an essay.
- Forbid invention: "If the transcript does not state it, write 'not stated'." Without this line, the model guesses, and a guessed deadline is worse than none.
- Fix the length: a word cap per field. Summaries are read in the gap between calls; a recap longer than the call is not a recap.
- Quote what was promised: prices, dates and commitments are copied verbatim from the transcript, not paraphrased, because paraphrase is where disputes start.
- Tag, don't editorialise: sentiment and urgency as single words from a fixed list, so they can be counted across a month of calls.
Template 1: The One-Line Recap
For the inbox, the Slack channel and the manager skimming forty calls. One sentence, fixed shape, no adjectives.
“Summarise this call in one sentence of at most 30 words using this shape: [Caller name or 'unknown caller'] called about [reason]; outcome: [booked / resolved / transferred / message taken / no action]; next step: [action and owner, or 'none']. Use only facts stated in the transcript.”
Template 2: The CRM-Ready Record
The workhorse. Each line maps to a CRM field, so the output can be written straight to the contact's timeline without a person reformatting it.
“Extract the following from the transcript, one per line, in this order. Write 'not stated' for anything the transcript does not contain. Caller: full name and phone number. Company: if mentioned. Reason for call: one sentence. Outcome: one of booked, rescheduled, cancelled, resolved, transferred, message taken, no action. Commitments made by us: quote verbatim. Commitments made by caller: quote verbatim. Next step: action, owner, deadline. Sentiment: one of positive, neutral, frustrated. Urgency: one of low, normal, high.”
Template 3: The Sales and Lead Summary
For inbound enquiries and outbound follow-ups, where the point of the summary is whether the lead is real and what it will take to close it.
“From this sales call, record: Lead: name, company, role. Need: what they are trying to solve, in their words. Timeline: when they need it, verbatim if stated. Budget: any figure mentioned, verbatim, or 'not stated'. Decision process: who else is involved. Objections raised: list each in the caller's words. Qualification: one of qualified, needs follow-up, not a fit, with one sentence of reasoning drawn only from the transcript. Agreed next step: action, owner, date.”
Template 4: The Support and Ticket Summary
For service lines, where the summary becomes the ticket and the next agent needs to pick up without asking the customer to repeat anything.
“Write a support ticket from this call. Customer: name and account identifier if stated. Issue: one sentence, in the customer's words where possible. Steps already tried: list, from the transcript only. Resolution on the call: what was done, or 'unresolved'. Promised follow-up: what, by whom, by when, verbatim. Priority: one of low, normal, high, using the customer's stated impact. Do not speculate about causes not discussed on the call.”
Template 5: The Compliance and Dispute Summary
For regulated lines and any call that might be revisited: what was disclosed, what was consented to, and what was said in the words that were used.
“Produce a factual record of this call for audit. Participants: names and roles as stated. Identity verification: what was checked and the result. Disclosures read: list each, with whether the caller acknowledged it. Consents given or refused: quote the caller verbatim for each. Requests made by the caller: verbatim. Actions taken: list, with timestamps if present in the transcript. Anything unclear or inaudible: note it as such. Do not summarise tone or intent; record only what was said and done.”
Rules That Keep Summaries Honest
The templates do most of the work, but three operating rules decide whether people trust the summaries enough to act on them.
- Summarise from the transcript, not the audio description: the model should see the words, with speaker labels. A summary built from a lossy intermediate inherits its errors.
- Keep the recording one click away: a summary is a pointer to the truth, not a replacement for it. Every summary carries a link to the recording and the full transcript.
- Review a sample weekly: ten summaries against their recordings, every week. Drift is gradual and invisible until somebody checks.
Prompts for Call Summaries vs. Built-In Summaries
If you are pasting transcripts into a chat window, the templates above are the whole solution. If the summaries are being written automatically after every call, the prompt is one setting among several: which fields map to which CRM properties, who receives the recap, and what happens when a field comes back as 'not stated'. Automatic call notes that land in the right record without anyone copying them are what turns a good prompt into a process.
FAQs
What is the best AI prompt for call summaries?
- One that names the fields, forbids invention and caps the length. The CRM-ready template above, caller, reason, outcome, commitments quoted verbatim, next step with owner and deadline, sentiment and urgency from fixed lists, and 'not stated' wherever the transcript is silent, produces summaries a team will actually rely on.
Can an AI summarise a call from the recording alone?
- It should summarise from a transcript with speaker labels, produced from the recording. Summarising directly from audio, or from a description of the audio, compounds transcription errors with summarisation errors. Keep the recording linked from every summary so a detail can always be checked.
How long should an AI call summary be?
- Shorter than the call, and usually under 120 words for a routine call. The one-line recap is for scanning; the structured record is for the CRM. If a summary is longer than that, the prompt is asking for prose rather than fields.
Can the summary be written straight into HubSpot or Salesforce?
- Yes, when the summary is produced by the call platform rather than pasted into a chat tool. Each field in the template maps to a CRM property, and the summary, outcome and a link to the recording land on the contact's activity timeline automatically after the call.
