Using AI to transcribe and summarize discovery or sales calls so you never forget what a prospect actually asked for

Small business owners lose deals not because they have bad conversations, but because they can't remember what was actually said — and the cost of misremembering a prospect's key requirement can be losing the entire engagement. This post walks you through the fastest, lowest-cost way to automatically transcribe and summarize every discovery or sales call, starting with your next one. The setup takes under 15 minutes, the best entry point is free, and the output directly reduces the gap between what a prospect asked for and what ends up in your proposal.

What you need before you start

Fathom{:target="_blank"} — an AI notetaker that joins your Zoom, Google Meet, or Teams call as a bot, records and transcribes it automatically, then emails you a structured summary within minutes of the call ending. The free individual tier covers unlimited transcription and AI summaries — no credit card required, as of early 2026. If you need shared call libraries or CRM sync for a team, the Team Edition{:target="_blank"} starts at approximately $19/user/month.

ChatGPT{:target="_blank"} (GPT-4o) or Claude{:target="_blank"} 3.5 Sonnet — for secondary analysis when you need more than Fathom's built-in summary. Free tiers on both cover occasional use; ChatGPT Plus is $20/month if you're doing this daily at scale.

Time required: 10–15 minutes for basic setup (Fathom account + calendar connection). Add 20–30 minutes if you want CRM field mapping or a ChatGPT prompt library configured for follow-up emails.

Skill level: No technical background needed. You need a Google, Zoom, or Microsoft account and the ability to copy and paste text.

Set up AI call transcription so it runs without thinking

1. Create a Fathom account at fathom.video using your work email. The signup flow asks you to connect your Google or Microsoft calendar — do this immediately, because it's how Fathom knows which calls to join.

2. Authorize Fathom to join meetings by following the calendar connection prompts. Once connected, Fathom will appear as an invited bot participant on every meeting the tool detects on your calendar going forward.

3. Set your recording preferences in the Fathom dashboard under Settings → Recording. Choose whether Fathom joins automatically or only when you click "Record." For discovery calls, auto-join is the right default — you want zero friction at the moment a prospect books.

4. Run a test call with a colleague before your next real prospect call. Schedule a 5-minute meeting, let Fathom join, end the call, and wait for the summary email. You should receive it within 3–5 minutes. The email will include: a short paragraph summary, a list of action items, key decisions, and questions the other party asked.

5. Notify your prospect before the call starts. Announce verbally that you're recording the call and have an AI notetaker present. Do this at the top of every call — it is a legal requirement in many jurisdictions, not a courtesy. The legal details are covered in a section below.

Tip on speaker labels: Speaker diarisation (labelling who said what) is important for multi-person calls — without it, the transcript reads as a wall of undifferentiated text and summaries lose context on which objections came from the decision-maker versus a secondary stakeholder. This feature is available in Fathom, but availability varies by plan tier; free tiers may produce unseparated transcripts. Verify whether speaker labels are active for your account in your Fathom dashboard settings before relying on them for a proposal.

Skipping the calendar connection and manually inviting Fathom to each call is a reliable way to miss calls. One forgotten discovery call with a high-value prospect, and the setup time pays for itself several times over.

How to summarize sales calls with ChatGPT or Claude

Fathom's built-in summary covers the basics well. Where it falls short is depth — it doesn't know your sales process, your proposal format, or what specific language a prospect used that signals urgency or hesitation. That's where a second-pass with ChatGPT or Claude earns its keep.

1. Open the Fathom summary email and copy the full transcript section (not just the summary paragraph — the verbatim transcript).

2. Open ChatGPT (GPT-4o) or Claude 3.5 Sonnet and paste the transcript into a new conversation.

3. Use a structured prompt. Here is the one I'd recommend for discovery calls:

You are analyzing a sales discovery call transcript for a small service business. Using only the content in the transcript below, answer these four questions:

1. What specific problems or pain points did the prospect describe?
2. What outcomes or results did they say they want?
3. What objections, concerns, or hesitations did they raise?
4. What next steps or commitments were mentioned?

Be specific. Quote the transcript directly where it supports your answer. Do not invent anything not in the transcript.

[Paste transcript here]

4. Review the output against the transcript. Both GPT-4o and Claude 3.5 Sonnet handle long transcripts well and follow structured prompts reliably, but they occasionally conflate speaker turns on dense crosstalk sections. Spot-check any quote that drives a proposal decision.

5. Use the output to write a follow-up email or update your CRM. Copy the pain points and next steps directly into your CRM notes field, or ask the model to draft a follow-up email: "Using the answers above, write a concise follow-up email that confirms what I heard and outlines proposed next steps. Professional tone. Under 200 words."

The value of this second-pass step is not speed — it is specificity. A generic Fathom summary tells you "prospect mentioned budget concerns." A ChatGPT-analyzed transcript tells you "prospect said 'we have about $3,000 set aside for this but my partner needs to sign off on anything above $2,500.'" Those are two different proposals.

When something goes wrong

The Fathom bot doesn't show up on the call. Root cause: your calendar connection dropped or the meeting was created from a different calendar account than the one you connected. Fix: check Fathom's dashboard under Upcoming Meetings to see if the call appears. If it doesn't, manually invite your Fathom bot email address (shown in your account settings) as a calendar guest.

The transcript is accurate but the speaker labels are wrong. Root cause: Fathom assigns speaker names based on meeting invitation data. If a participant joined from an unrecognized email or a dial-in number, their name won't resolve. Fix: after the call, open the transcript in the Fathom dashboard and manually reassign speaker labels before using the transcript for a proposal or CRM update.

ChatGPT produces a summary that includes details not in the transcript. Root cause: GPT-4o occasionally fills gaps with plausible-sounding inference, especially on short or low-audio-quality transcripts. Fix: include "Do not invent anything not in the transcript" explicitly in your prompt (already included in the template above), and cross-check any dollar figure, timeline, or named stakeholder against the source transcript before using it in client-facing materials.

The legal side before you start recording

Recording laws vary significantly by jurisdiction, and "I didn't know" is not a legal defense.

In the US, federal law follows a one-party consent standard — meaning one participant (you) consenting to the recording is sufficient. However, approximately a dozen states — including California, Florida, and Illinois — require all-party consent. If your prospect is in one of those states, they must be informed and must consent before you hit record. Check the Reporters Committee's state-by-state guide{:target="_blank"} for a reliable reference.

In the UK and EU, GDPR{:target="_blank"} requires explicit informed consent before recording. This means telling the person clearly what you're recording, why, how long you'll keep it, and what you'll do with it.

The practical fix is simple: open every call with a one-sentence disclosure. Something like: "Before we start — I use an AI notetaker so I can be fully focused on our conversation rather than taking notes. Is that okay with you?" Most prospects agree immediately, and many appreciate the transparency.

Limitations to know before you rely on this

AI transcription accuracy for English in quiet environments typically runs 90–95% for tools like Fathom, Otter.ai{:target="_blank"}, and Fireflies.ai{:target="_blank"}. That number drops noticeably with heavy accents, multiple simultaneous speakers, or poor audio — a phone call on a bad connection can produce transcripts that miss entire sentences. Never use an AI transcript as a legal or contractual record without human review.

Otter.ai's free plan caps transcription at 300 minutes per month with a 30-minute per-conversation limit — sufficient for occasional use, but constraining if you run multiple discovery calls weekly. Its Business plan{:target="_blank"} runs around $20/user/month and adds CRM integrations with Salesforce and HubSpot. Fireflies.ai's free plan includes 800 minutes of storage and basic summaries; paid plans start at approximately $18/seat/month and include their "Ask Fred" AI assistant, which lets you query a transcript directly — "What budget did the client mention?" — without a separate ChatGPT session. Both are legitimate alternatives to Fathom, but neither matches Fathom's free tier for individual users at zero cost.

What to do next

Once this workflow is running, the natural optimization is feeding your call summaries into a CRM automatically rather than pasting them manually. Fathom's Team Edition includes CRM sync; Fireflies handles this on paid plans as well. If you are not yet using a CRM, this is a good moment to start — the structured output from these tools maps directly to contact record fields.

For building the next layer of automation around your sales workflow, see how to automate client follow-up emails with AI and how to use AI to write better proposals from client notes.

FAQ

Is there a genuinely free AI transcription tool that doesn't require a credit card? Yes. As of early 2026, Fathom's individual plan offers unlimited transcription and AI summaries for Zoom, Google Meet, and Teams at no cost, with no credit card required at signup. Otter.ai also has a free tier, but it caps at 300 minutes per month and 30 minutes per conversation — a real constraint if your discovery calls run longer. Pricing checked February 2026; verify at each tool's site before committing.

How do I summarize a client call with ChatGPT if I already have a transcript? Paste the full transcript into a new ChatGPT (GPT-4o) or Claude 3.5 Sonnet session and use a structured prompt that specifies exactly what you want extracted: pain points, objections, next steps, budget signals. Unstructured prompts like "summarize this" produce generic output. The more specific your extraction criteria, the more usable the result for writing a proposal or drafting a follow-up email.

What does AI call transcription actually cost for a small team of five? For a 5-person team, Fathom's Team Edition would run approximately $95/month ($19/user/month as of early 2026). Fireflies Pro comes in at roughly $90/month ($18/seat/month). Otter.ai Business runs approximately $100/month ($20/user/month). The honest answer is that for a team where each person runs even two sales calls per week, the time saved on manual note-taking — conservatively 20–30 minutes per call — represents a return that dwarfs the monthly subscription cost at any reasonable billing rate.

Does AI transcription work for phone calls, or only video meetings? Most AI notetaker bots — Fathom, Otter, Fireflies — work by joining video meetings as a participant, which means they require Zoom, Google Meet, or Teams. They do not natively record traditional phone calls. For phone-based businesses, Otter.ai has a mobile app that can record in-person or phone conversations directly; accuracy on phone audio is lower than on video calls due to audio compression.

What if my prospect refuses to be recorded? Respect the refusal — this is non-negotiable both legally and professionally. The alternative is to use the call summary workflow in reverse: take lightweight notes during the call (just keywords and numbers), then immediately after hanging up, dictate or type a brief voice memo or text summary and feed that into ChatGPT for structuring. It is less accurate than a full transcript but significantly better than relying on memory 30 minutes later.