Fact-Checking AI Output

A smooth answer is not a checked answer

An assistant writes a wrong fact in the same calm voice as a right one. You cannot tell them apart by reading. So the habit is not to read more carefully. The habit is to check before you use.

This lesson is about the answer you already have. The earlier lesson, Sources First, Prompts Second, is about what you feed the assistant. Feeding it good sources lowers the number of errors. It does not bring them to zero, and the ones that remain are the ones a buyer will notice.

Six mix-ups to look for

Most errors in account research come from a short list. Know them and you will find them faster.

The look-alike company is the one that hurts most on named accounts. You chose this company on purpose. A brief about the wrong company, sent to the right person, tells them you did not do the work.

Step one: split the answer into claims

A paragraph is hard to check. A list of claims is easy. Take the assistant's answer and write one line for each thing that could be right or wrong: a name, a title, a date, a number, a quote, an event.

Claim log (paste into a spreadsheet)

Claim	Type	Goes to buyer?	Source I opened	Date of source	Result
[the claim, in one line]	[name / title / date / number / quote / event]	[yes / no]	[link]	[date]	[keep / rework / bin]

Fill the first three columns now. Fill the rest as you check. The log is also your record, so a colleague can see what was verified and when.

Step two: decide how hard to check each claim

You do not have the time to check everything, and you do not need to. Ask one question about each claim: what happens if this is wrong?

• It goes in front of the buyer. A name, a quote, a number, an event you will mention. Check every one at the source. No exceptions.
• It shapes how you approach. A guess about priorities, or about who matters. Check it once, or turn it into a question.
• It is background. Context that you will not state. Leave it, but do not let it turn into a statement later.

The best way to cut the work is to cut the claim. If a message does not need the risky fact, remove the fact.

Step three: check against the primary source

Open the original. Do not ask the same assistant whether it was right. It will usually agree with itself.

A second assistant can help you find errors you missed. Use it to point you at claims to check, then open the source yourself. The Fact-Check AI Research Prompts pack has ready prompts for that.

Look-alike check (paste with the answer)

Below is a research answer about [company], whose website is [domain] and which operates in [country]. List every company, product, person and place named in it.

For each one, tell me whether it could be confused with another organisation of a similar name, such as a parent, a subsidiary, a competitor or a firm in another country. Flag any claim that might belong to a different organisation than the one I mean. Do not tell me whether claims are true. Tell me what I need to check.

Answer:
[paste]

This prompt does not decide anything. It raises flags for you to check, which is what it is good at.

Keep, rework or throw out

Every claim ends up in one of three places.

Throwing out feels like waste. It is the cheapest outcome. A fact you cannot source is a risk you have not yet been charged for.

Where this sits in named account work

On a long list nobody checks the facts, because there are too many companies and each message is generic anyway. On a named list the sums are different. You have a short list of accounts, one message per person, and each message can only be as good as its weakest fact.

So make the check part of the routine, not an extra:

• Do it before any new person at the account is contacted.
• Log the date. A fact checked in the spring is a guess by the autumn.
• Check names and titles again before each call, since people move.
• Keep checked facts in the account notes, apart from anything you have not checked.