An AI tool answers in the same confident tone whether it is right, partly right or wrong. That is the main reason people get into trouble with it: the text looks finished, so it is copied into an email, a report or a decision without anyone looking at it twice.
This article is for employees and team leads who already use AI tools at work. It does not argue for or against using them. It gives five questions you can ask in a minute or two, and a rule for how much checking is enough. If you also need rules for the whole team, see writing a simple usage policy for AI tools; this article is about the check of a single output.
Why a good-looking answer is not a checked answer
Language models produce text that is likely to follow your question. They are not looking facts up in the way a search for a document does, unless the tool is explicitly connected to sources. The result can be accurate, but it can also contain invented details, mixed-up names or outdated statements, all written in the same smooth style. We describe this problem in more detail in AI hallucinations.
So the question is never "does this sound right?" but "what would I need to see to rely on it?"
The five questions
1. Where does this come from?
Ask the tool what the statement is based on, and then look at the answer with suspicion. If it names a document, law, study or web page, open it yourself. A reference that looks real can still be wrong or not exist. If the tool gives no source at all, treat the statement as a draft idea, not as a fact.
2. How current is it?
Many tools work with knowledge that ends at a certain date, and some can search the web while others cannot. Prices, rules, product features and people in roles change. If the answer depends on something that changes, check the date of your source, not the date of the answer.
3. Do the numbers and names check out?
Numbers, dates, percentages, quotations and proper names are the most common places for small errors. Take each one that matters and compare it with the original document, spreadsheet or website. Do not ask another AI tool to confirm it; two tools can repeat the same mistake. Also check simple arithmetic, since totals and shares are not always calculated correctly.
4. Does it answer my question, or a similar one?
The tool may answer a more general question than yours, or assume a country, industry or company size that is not yours. Read the answer against your real situation: your market, your contract, your system. If you left out important context in the question, the answer fills the gap with an assumption, and you may not see which one.
5. What happens if this is wrong?
This question decides how much effort the other four deserve. A brainstorm of meeting titles can be used as it is. A paragraph that goes to a customer, a number that goes to leadership or a statement about a legal or financial obligation needs a real check, and often a person with the right expertise. Never let an AI answer be the only basis for a decision that is hard to reverse.
Match the checking to the risk
You do not have to verify everything to the same depth. A practical rule:
- Low consequence (internal notes, ideas, first drafts): read it once, fix what looks off.
- Medium consequence (customer-facing text, internal reports): check every number, name and source that appears in the text.
- High consequence (contracts, finance, safety, personal data, legal or regulatory statements): use the AI output only as a starting point and have the owner of that area review the final text.
Also think about what you paste into the tool. Do not enter confidential or personal data unless your company has agreed that the tool may be used for it.
A short example
Suppose a colleague asks a tool to summarise the main differences between two suppliers' offers for a customer proposal. The summary says supplier B offers a 24-hour response time and includes training.
Going through the questions: the source is not named (question 1), so you ask the tool which document it used and find that it only saw one of the two offers. The response time is a number (question 3); in the original offer it is 24 working hours, not 24 hours. The training is mentioned only as an option (question 4). The proposal goes to a customer (question 5), so every figure is compared with the offers before it is used.
The check took a few minutes and removed two statements that would have been wrong in a document sent outside the company.
Make the check repeatable
Write one or two lines next to the output: what you verified, against what, and when. If the text is shared, mark which parts came from an AI tool and which parts a person confirmed. This helps a colleague who reviews the work later, and it builds a habit that does not depend on one careful person.
If you are at the stage of deciding where AI belongs in your work in the first place, an AI audit is a sensible first step.
What to do this week
Pick one AI output you used recently at work. Go through the five questions and note which one you would have skipped. Then choose a rule for your own team: for example, "any number in an external document is checked against its original source." One rule applied every time is more useful than a long checklist used once.
Sources
- Our related articles linked above.
- The five questions and the risk levels are a general working approach, not a standard; adapt them to your own company and tools. The example is illustrative and does not describe a real client. No figures or client results are claimed in this article.