What you'll get
- 4 copy-ready blocks
- 6 min read
AI research is fast and confident, and confidence is not accuracy. Before any AI-assembled fact reaches a buyer, it needs the treatment below: extract the claims, demand the sources, cross-examine the brief. Three prompts, a risk tier to decide how hard to check, and a routine that costs five minutes per account.
Prompt 1: Extract the claims
Run this on any AI-produced brief, email draft or account summary:
Below is a research brief that may contain errors. Extract every
factual claim into a numbered list. A claim is anything checkable:
names, titles, numbers, dates, events, quotes, product facts,
relationships between companies.
For each claim, add:
- WHERE it could be verified (the company's site, a filing, a news
source, LinkedIn)
- RISK if wrong: HIGH (goes in front of the buyer), MEDIUM (shapes
our approach), LOW (background colour)
Do not assess truth yet. Just list, locate and tier.
Brief:
[paste]The tiering is the efficiency: HIGH claims get human verification, every time; MEDIUM gets prompt 3; LOW gets left alone.
Prompt 2: Demand the sources
For research done in a tool with web access, run this as the standing follow-up:
For claims [numbers] above: give the specific source for each, with
a working link and the date of the source material. Quote the exact
sentence from the source that supports the claim.
Rules:
- If you cannot produce a source, say "unsourced" plainly.
- If the source says something adjacent but not quite this, say
"stretched" and show the gap.
- An old source is an answer too: flag anything over 12 months old.The three honest outputs (unsourced, stretched, stale) are the prompt's real product. Sources-first research habits, which prevent half of this cleanup, are the method in the account research prompt pack and the lesson Sources First, Prompts Second.
Prompt 3: Cross-examine the brief
Run in a fresh conversation, ideally a different model, with no shared context:
You are a sceptical reviewer. Below is a research brief about
[company]. Attack it:
1. Which claims are most likely to be hallucinated or stale, based on
how these errors usually look? Name them specifically.
2. What is internally inconsistent: numbers that do not add up, dates
out of order, titles that conflict?
3. What is suspiciously convenient for the seller using this brief?
4. What is missing that a competent human researcher would have
checked: leadership changes, recent news, the obvious counter-fact?
Be specific and ungenerous. End with the three claims you would
verify first.A second model with no stake in the first model's answers catches a different error class: the internally consistent fiction. Its "three claims to verify first" plus your HIGH tier is the human checking list.
The five-minute routine per account
Minute 1 Prompt 1 on the brief. Read only the HIGH rows.
Minutes 2-4 Verify every HIGH claim yourself at the source: the
company site, the filing, the profile. Not via the
model; via the actual page.
Minute 5 Prompt 3's top three, if any survived into your draft.
Fix, or cut: an email never needs the risky fact.The last line is the discipline that makes this fast: when verification is expensive, deletion is free. A first line built on a verified fact beats a richer paragraph built on three unverified ones; the annual-report routine exists precisely because primary documents cut the checking bill.
Worked example (illustrative, not a real company)
Dana Whitfield has an AI brief on Calder Cold Chain before a Tier 1 first call. Prompt 1 finds 14 claims, four HIGH. Prompt 2 returns one "stretched": the "recent acquisition of a rival" traces to a trade-press rumour from last year, never confirmed. Prompt 3, in a different model, flags the operations director's tenure as a likely confusion with a namesake. Three minutes of human checking confirms both problems; the acquisition line dies, the tenure gets corrected from the actual profile. The call happens with eleven verified facts and no landmines, and the email afterwards quotes only what Dana checked personally. All names are invented.
Five mistakes to avoid
- Asking the same model "are you sure?" It will be sure. Confidence calibration inside one conversation is theatre; the cross-examination needs a fresh context or a different model.
- Checking everything equally. Flat-rate scepticism is why teams abandon checking. The risk tiers put the five minutes where the damage lives: claims a buyer will see.
- Accepting a link as a source. Models produce real-looking links to pages that do not say the thing. The "quote the exact sentence" rule exists because the quote is checkable in seconds and the link alone is not.
- Shipping the risky fact anyway, hedged. "I believe you recently acquired..." is not a hedge, it is a wrong fact with an apology attached. Cut it; the email survives.
- Treating stale as true. Last year's revenue, the old job title, the pre-pivot strategy: the commonest error class is time, not invention. The 12-month flag in prompt 2 catches most of it.
Why this works on named accounts
Common questions
How do you fact-check AI research quickly?
Tier it: extract every claim, mark what a buyer will actually see as high-risk, verify those few at the primary source yourself and run a sceptical second-model pass for the rest. Five minutes per account covers it, because deletion handles whatever verification cannot.
Why does AI research sound right when it is wrong?
Models produce the most plausible continuation, and plausible is what stale or invented facts look like: a believable revenue number, a sensible-sounding acquisition. Internal consistency is no signal either, which is why the cross-check runs in a fresh context where the first model's fiction cannot defend itself.
Which AI research errors matter most in sales?
The ones a buyer can falsify from their chair: their own numbers, their colleagues' names and titles, their company's events. Getting the market context slightly wrong is survivable; getting the buyer's world wrong in the first email is a credibility event, which is why those claims are HIGH tier and always get a human check.
More from the Academy
15 AI Prompts for Account Research
For sellers who use Claude, ChatGPT or Perplexity to prepare for an account.
You leave with a set-up prompt, 15 research prompts in working order, a fact check that catches invented details and a sample of its output.
AI Prompt Pack: Read an Annual Report in 10 Minutes
For sellers researching a large account before first contact.
You leave with 7 prompts, a 10-minute timetable, a one-page brief format and one brief written out.
ChatGPT and Claude Prompts for Writing Cold Emails
For sellers who use ChatGPT or Claude to draft outreach and get emails that sound like everyone else's.
You leave with a set-up prompt, a research block, 10 prompts that draft, cut, criticise and fact-check, and one worked draft.
Want this in your inbox?
Optional. We use your email only to send you this, nothing else.