Prepping a sales call with AI without inventing facts

Quick answer

Treat prep as three separate passes and never let them blend: gather facts with citations, synthesise from material you actually have, then write a call plan you can defend. Its value is not knowing the account; it is turning a folder of half-remembered notes into a one-page brief fast — and inventing an org chart just as fast if you let it work from vibes.

Why prep is the wrong job for a single vague prompt

"Tell me about Acme" returns a confident paragraph that may mix up two companies with similar names, restate last year's headcount as current, or attach a quote to the wrong executive. The failure is not an obvious guess you can spot; it is a plausible detail you repeat on the call. Prep compounds that risk, because everything downstream inherits it: the agenda, the questions, the posture you walk in with.

Pass one: gather with citations

Use a research tool that shows its sources, and ask for a short list rather than a narrative: what the company does, recent announcements, and the people who own the problem you sell into. Perplexity is built around cited answers, which is the right shape for this pass. Open every link and check the publication date: a two-year-old announcement presented as news is worse than no research.

Give the tool permission to fail: ask for public information only, sources linked, and "not found" where it cannot verify. A model allowed to say "not found" invents less. If an answer arrives tidy and unsourced, treat it as unverified.

Pass two: synthesise from material only you have

The second pass is where the return is real, because the input is your own record: the last call's notes, the support thread, the discovery answers in the CRM, the email where they explained their budget cycle. Paste that in and ask for structure rather than prose: what they said, what they did not say, what changed since last time. Claude and ChatGPT both hold long pasted threads; which one keeps your constraints better is something you can only judge on your own documents.

Two cautions before you paste anything. A personal subscription is not a data-processing agreement. If the notes carry customer names, pricing, or commitments, the question is what your organisation has accepted, not what the vendor's marketing says. What to ask before connecting an AI tool to company data is the checklist, and the AI usage policy checklist covers the internal rules. And redact what the call does not need: the brief needs the problem and the history, not personal details that happen to sit in the same note.

Pass three: write the plan you will actually use

Ask for a one-page plan with fixed slots, then edit it by hand:

One sentence on why this call is happening. If you cannot write it without the model, the call is not ready.
Three questions, ranked. The first should be answerable from their side, not yours. "How are you handling this today" beats "would you like our product".
One hypothesis, labelled as a hypothesis. "They are consolidating vendors after the reorganisation" is a guess until they confirm it.
What would change your mind. Naming the answer that would kill the deal stops you hearing only confirmation.
The next step you will propose. Specific and small: a scoped trial, or a technical call with one named person.

Keep the labels in the document. A plan where facts and hunches look equally confident is worse than a blank page, because you will read it in the ten minutes before the call and believe all of it.

Recording is the prep input most teams ignore

If your team already records calls with consent, the previous conversation is better material than anything public. Fathom, Fireflies, and Otter all separate summaries and action items from the transcript, which is the shape prep needs: what they asked for, who owned it, whether it happened. Recording rules are a policy question: consent, retention, and who can listen. The meeting-notes workflow and how to pick one notes tool cover that ground. Never paste a recording from an account whose retention terms you do not control.

Where automation helps, and where it quietly hurts

An automation in Make or Zapier can pull a call summary into the CRM record or assemble the prep document before the meeting. It should not send: auto-generated outreach built on an unverified detail puts a wrong fact in writing, where it becomes a correction instead of a question. Automate the assembly; keep the send human.

The five-minute check before you dial

Every factual line has a source you opened, or is labelled a hypothesis.
Names, titles, and the company's current name are right — rebrands and mergers are the classic failure.
Nothing in the document would embarrass you if the prospect read it.
The three questions are written down, not in your head.
You know the next step you will ask for.

Frequently asked questions

Can AI research a prospect without hallucinating?

Not reliably enough to skip verification. A cited research tool reduces invention and lets you check each claim, but it can still surface a stale page or an outdated figure. Treat its output as a list of leads to open.

Is it safe to paste CRM notes into a chatbot?

Only under terms your organisation has accepted for that tool. Customer names, pricing, and contract terms are exactly what those agreements exist for, and a personal plan is not one of them.

Should AI write the follow-up email after the call?

It can draft from notes you pasted, but a person should send it. Dates, commitments, and pricing in a follow-up are treated as contractual in practice, and a friendly detail nobody said is a problem you have to unwind later.