Fact-check AI output in five minutes by marking every checkable claim (prices, numbers, dates, names, quotes and rules), checking each one against its original source rather than asking the AI, recalculating every sum and weekday, deleting what you can't source, and reading the result once as the customer will. Send anything legal-sounding to a person.
The dangerous errors aren't the obviously silly ones. They're plausible details that sound exactly like your business: last year's price, a date that's right but paired with the wrong weekday, or a rule stated with official-sounding confidence. You won't catch those by reading the draft and asking "does this sound right?", because it will. The routine works because it's mechanical: you stop judging the text and start checking each claim against something outside the AI.
The five-minute routine, minute by minute
Each step has a job and a way to prove you did it. Here's the checklist with the reason for each step and how to verify it.
| Minute | Do this | Why | How you know it's done |
|---|---|---|---|
| 1 | Mark every checkable claim: prices, numbers, percentages, dates, weekdays, names, quotes, rules, links, and words like "always", "never", "free", "guaranteed" and "by law" | You can't check what you haven't spotted, and absolutes are where invented rules hide | Every marked item is highlighted or listed |
| 2 | Check each claim against its source of truth: the price list, the booking system, the contract, the offer calendar | The AI isn't a source; it's the thing being checked | Each claim has a tick and the name of its source |
| 3 | Redo every sum, discount, total and weekday yourself | Arithmetic and calendars are where confident drafts slip most | You've recalculated, not re-read |
| 4 | Cut or soften anything you couldn't source in 30 seconds; rewrite legal-sounding lines as your own terms or send them to someone qualified | An unsourced specific is a liability; "by law" is a claim you probably can't back | No unticked claims remain |
| 5 | Read it once as the recipient: what will they believe, book or pay because of this? | Catches promises and implications that aren't single facts | You'd be happy to honour every promise in it |
A pocket version you can pin by the screen:
1. MARK every number, date, name, rule, "free", "always", "by law"
2. CHECK each one against the real source (not the AI)
3. REDO sums, discounts, weekdays
4. CUT anything unsourced; legal lines -> own terms or an adviser
5. READ as the customer: what will they act on?Now the routine on three realistic drafts. Each is illustrative, from a different kind of business, and each contains the sort of error that gets through a quick read.
Output one: a price in a nail salon's promotional post
The owner asked an AI assistant for an autumn social post about October's offer. The offer, from the salon's own offer calendar, is free nail art on full sets booked Monday to Wednesday, 1 to 31 October. The price list dated 1 September 2026 shows gel manicures at $38. The draft:
"Autumn's here! Our gel manicure is now just $35 (was $42), and every full set this month comes with free nail art worth $15. Book three visits and get your fourth half price!"
Minute 1, claims marked: $35; was $42; every full set; this month; free nail art; worth $15; fourth visit half price.
Minutes 2 and 3, checked:
- $35: wrong. The price list says $38. The AI had picked up $35 from an old caption the owner pasted in as a style example.
- "Was $42": invented. The salon has never charged $42. A "was" price you never charged misleads customers and can breach advertising rules.
- "Every full set": wrong. The offer is Monday to Wednesday only.
- "Worth $15": unsourced. Nail art is priced at the table, so there's no fixed value to quote.
- "Fourth half price": invented. There's no such offer.
Minute 4, cut and rewritten. Minute 5, read as a customer: would anyone book a Thursday full set expecting free art? Not any more.
"Autumn's here! Book a full set Monday to Wednesday this October and your nail art is on us. Gel manicures are $38. Book online or give us a call."
Time taken: about four minutes. Five problems in a 30-word draft is not unusual for promotional copy, because the AI is trying to make the offer sound generous. Fact-checking AI marketing copy before it goes live goes deeper on offers, comparisons and claims in ads.
Output two: dates in a yoga retreat email
A yoga studio asked for an email announcing its winter retreat. The booking system says Friday 12 March to Sunday 14 March 2027, two nights, with an early-bird price until 31 January. The draft:
"Join us for our Winter Reset Retreat, Saturday 14 March to Monday 16 March 2027: three nights of yoga, walks and good food. Early-bird pricing ends Friday 31 January, so book soon!"
Claims marked: Saturday 14 March; Monday 16 March; 2027; three nights; Friday 31 January.
Checked against the calendar, not memory:
- 14 March 2027 is a Sunday, not a Saturday, and the retreat starts on Friday 12 March anyway.
- 16 March 2027 is a Tuesday, not a Monday, and the retreat ends on Sunday 14 March.
- "Three nights" doesn't match either version: Friday to Sunday is two.
- 31 January 2027 is a Sunday, not a Friday. The date is right; the weekday is wrong.
Four errors, and every one is the kind that sends a guest to the wrong place on the wrong day, or makes them miss the early-bird price by two days because they trusted "Friday". The fix:
"Join us for our Winter Reset Retreat, Friday 12 March to Sunday 14 March 2027: two nights of yoga, walks and good food. Early-bird pricing runs until Sunday 31 January."
The habit that catches these: never check a weekday by thinking about it. Open a calendar and look. It takes five seconds a date.
Output three: a legal-sounding line in a wedding planner's FAQ
A wedding planner asked an AI to draft an answer to "Can I get my deposit back if we cancel?" for the website FAQ. The planner's contract says the $500 deposit is non-refundable if the couple cancels, and refunded in full if the planner cancels. The draft:
"Under consumer law, deposits are always non-refundable, so we are unable to return them in any circumstances. This is standard practice across the wedding industry."
Claims marked: "under consumer law"; "always non-refundable"; "in any circumstances"; "standard practice across the wedding industry".
Checked:
- "Under consumer law" can't be sourced to anything the planner holds, and consumer rules vary from place to place and case to case. It's a legal claim the business shouldn't make in its FAQ.
- "Always" and "in any circumstances" contradict the planner's own contract, which refunds the deposit if the planner cancels.
- "Standard across the industry" is an unverifiable generalisation that adds nothing.
The rewrite states the business's own terms and makes no claims about the law:
"Your $500 deposit secures your date and covers the planning work we start straight away. Under our terms it isn't refundable if you cancel, but if we ever have to cancel, we refund it in full. The full cancellation terms are in your contract, and we're happy to talk them through."
This is the output where the routine's step 4 matters most. If you're unsure whether your own terms are fair or enforceable, that's a question for a solicitor or adviser, not for the AI that wrote the draft. Legal-sounding lines are also where made-up specifics appear most confidently, the same pattern covered in catching made-up figures in AI-drafted proposals.
A quick fourth: names and a quote in a review reply
Names and quotes deserve their own mention because they're easy to skim past. An illustrative barber shop asked an AI to reply to a five-star Google review that said: "Great skin fade, friendly team, will be back." The draft reply:
"Thanks so much! [Barber A] loved doing your fade, and we're glad you enjoyed it. As one of our regulars put it, 'best cut in town.' See you in four weeks!"
Three claims to check. The booking system shows [Barber B] did the cut, not [Barber A]. The "regular" quote doesn't exist; the AI invented a testimonial, which is never acceptable in anything public. And "see you in four weeks" assumes a rebooking the customer hasn't made. The corrected reply thanks the customer, names the right barber, and drops the invented quote and the assumption. Thirty seconds of checking, and a public page that no longer misattributes work or fabricates praise.
Where each kind of draft usually goes wrong
After a few weeks of logging, most businesses see the same patterns. Use this table to decide where to look first in minute 1.
| Kind of output | The error it tends to contain | The check that catches it |
|---|---|---|
| Promotional posts | Invented offers, "was" prices, missing conditions | The offer calendar, line by line |
| Event and booking emails | Right dates with wrong weekdays; wrong durations | A calendar, opened, for every date |
| FAQ answers and policies | Absolutes and legal-sounding claims | The terms page; anything about the law goes to a person |
| Quotes and estimates | Sums that don't add up; items from a different job | Recalculate every line; compare with the job sheet |
| Review replies | Wrong staff names; invented testimonials; assumptions about rebooking | The booking record for that visit |
| Supplier messages | Wrong quantities or units; old delivery addresses | The purchase order and the delivery details on file |
| Summaries of long threads | A decision stated as final when it was only proposed | The original message in which it was agreed |
Build a source-of-truth card before you need it
Minute 2 only takes a minute if you know where the truth lives. Write a short card, one per business, listing the single source for each kind of fact. The yoga studio's, filled in:
| Kind of claim | The one source to check | Not these |
|---|---|---|
| Prices | Price list page, dated | Old posts, flyers, past emails |
| Class times and events | The booking system | The website's news page, which lags behind |
| Offers | The offer calendar in the shared sheet, with start and end dates | Anyone's memory of "the usual deal" |
| Policies | Terms page, current version | The FAQ, which should quote the terms, not replace them |
| People's names and roles | Current staff list | Old bios on the website |
| Anything about the law | Don't state it; link to your terms or ask an adviser | The AI's summary of "the rules" |
The card also shows you where your own records disagree. If the booking system and the website give different class times, the AI draft wasn't the only problem.
Why asking the AI to check itself falls short
It's tempting to finish a draft with "now check this for errors". Sometimes that catches a typo or a sum. But the model that produced a wrong detail can repeat it confidently, and it has no access to your price list, your calendar or your contract unless you've given them to it. In the nail salon example, the $35 came from text the owner supplied; asked to check, the AI would find the $35 in its own context and call it correct.
AI is useful in the routine in one place: minute 1. A prompt such as the one below produces a list of claims to check, which is handy for long drafts.
List every checkable claim in the text below: each price, number,
percentage, date, weekday, name, quote, link, rule or legal statement,
and every use of "always", "never", "free" or "guaranteed".
One per line. Don't judge whether they're correct.
[paste draft]Run on the retreat email, an illustrative output reads: "1. Winter Reset Retreat (name). 2. Saturday 14 March. 3. Monday 16 March. 4. 2027. 5. Three nights. 6. Early-bird ends Friday 31 January." That's a complete list, and not one item is marked wrong, because you told it not to judge. Then you check each line yourself. The split is simple: the AI finds, you verify. AI hallucinations explained for business owners covers why the verifying can't be delegated back to the model.
When five minutes isn't enough
The routine suits everyday output: a post, an email, a short web update. Some output needs more:
- Anything going to many people at once, such as a newsletter to 2,000 subscribers. One wrong date multiplies; have a second person run the routine independently.
- Quotes, invoices and contracts. Every line item gets checked against the source, not a sample.
- Anything legal, financial, medical or safety-related. The routine's job here is to spot the claim and route it to someone qualified.
- Automated output that sends without a person seeing it. If an automation drafts and sends, there's no moment for the routine. Add an approval step; adding human approval steps to AI automations shows how.
Turning the routine into a team habit
A routine one person follows is a habit; a routine a team follows needs a few supports:
- Put the pocket version where drafts are written: on the monitor, in the shared prompt library, at the top of the content calendar.
- Keep the source-of-truth card current. When a price changes, the card's source changes the same day.
- Log catches for a month. A simple tally of errors found by type (price, date, invented offer, legal claim) shows where your prompts need better inputs. If invented offers top the list, give the AI the offer calendar up front.
- Time it. Most short drafts take three to five minutes. If the routine regularly takes 15, the drafts are too long or the sources too scattered, and that's the thing to fix.
After a month, most teams find the same three or four error types recur, and a better prompt with the right facts pasted in prevents most of them before the check even starts.
Fact-checking AI drafts: quick answers
Can I use a second AI tool to fact-check the first one?
Use it to find claims, not to confirm them. Asking another assistant to 'list every checkable fact in this text' is a quick way to make sure you haven't missed a number or a date. But a second model can share the same wrong assumption, and it doesn't know your prices, calendar or terms. Each claim still has to be checked against your own source, such as the price list, the booking system or the signed contract.
What if I can't find a source for a claim in the draft?
Cut it or soften it. If you can't point to where a fact comes from within about 30 seconds, it doesn't go out as a fact. Replace a specific but unsourced figure with something you know is true, or remove the sentence. A shorter message with nothing wrong in it does its job; a longer one with a single wrong price can cost you the sale, a refund or a complaint.
Does every piece of AI output need this routine?
Anything that leaves the business or that someone will act on does: customer emails, social posts, quotes, invoices, web pages, supplier messages and anything with a number in it. Internal brainstorming, first drafts you'll rewrite, and notes to yourself don't. A useful test is to ask whether a customer, supplier or member of staff would be misled if a detail were wrong. If yes, run the routine.
Further reads
- What to Do When ChatGPT Gets Facts About Your Business Wrong — When public AI tools state wrong facts about your business.
- How to Check AI Translations Before Customers See Them — Extra checks when the AI output is a translation.
- How to Write a House Style Guide Your Team and AI Both Follow — A style guide that cuts errors before the check.
- Tone of Voice Examples: AI Copy Before and After Editing — Before-and-after edits of AI copy for tone.
- How to Review AI Call Transcripts for Quality and Compliance — The same discipline applied to AI call transcripts.
- How to Write AI Prompts That Sound Like Your Business — Prompts that give the AI your facts and voice up front.
- How to Set Up Human Review for AI Work Without Slowing Down — Four levels of human review matched to risk, how to make each check take under a minute, how many to sample, and when to relax or tighten.
- How to Check AI Is Doing Good Work, Not Just Fast Work — A rubric template, a 30-minute weekly sampling routine, a re-runnable test set and the warning signs, shown through a garden centre's plant-care replies.
- How to Run a SWOT Analysis of Your Business With AI — Six stages and copyable prompts for a SWOT built on your own figures and reviews, with a podiatry clinic example and a TOWS step that turns it into actions.
- How to Use ChatGPT as a Business Adviser Without Being Misled — Prompts and checks that turn ChatGPT into a useful sparring partner for business decisions, and stop it flattering a bad idea into a plan.
- How to Use AI on Your Own Admin First, Then Roll It Out — A six-week owner-first plan: pick three of your own admin jobs, log what AI saves and gets wrong, write a playbook, then hand one task to one person.
- ChatGPT Prompts for Bookkeepers: Client Queries, Chasers, Notes — Twelve copy-ready prompts for client replies, record chasers and file notes, each shown with an illustrative output and the fix it needs.
- How to Stop AI Inventing Case Law: A Small-Firm Checking Routine — A seven-step routine and a copyable citation log for catching AI-invented or misquoted authorities before anything reaches a court or client.
- How Marketing Agencies Use AI to Automate Client Reporting — A four-layer setup for automated client reports: data pipes, the right reporting tool, a commentary prompt that can't invent wins, and a five-minute check.
- Machine Translation Post-Editing: A Translator's Workflow — A post-editing routine for translators: triage by risk, pre-flight set-up, three focused passes, the errors fluent MT hides, and pricing from a sample.
- How Estate Agents Use AI to Prepare Valuation Appointments — A two-day prep routine for market appraisals: AI builds the property brief, comparables commentary and vendor pack, and the valuer sets the price.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: no vendor product facts are relied on; the examples are illustrative, and the weekdays were checked against a calendar. Checked September 2026.