Skip to content
3 role hubs · 8 tasksData privacy

AI tools for answering questions at work: check where the answer comes from

Sort question-answering AI tools by where the answer comes from: model memory, a web search or your own documents. Each needs a different check, and anything that decides a figure, tax position or employment matter goes to the primary source and a person who signs it off.

Short answer: before you ask “which tool?”, ask “where will this answer come from?” The type of tool tells you what to check. For some questions the right move is to skip the chat and open the official page.

Three kinds of tool that answer questions, and where each answer comes from

Memory. You type a question and the tool replies from what the model learned in training. It looks nothing up. The reply is quick and reads well, and it has no source at all.

Web search with citations. The tool runs a search, pulls back pages and writes a summary with links or numbered citations. The footnotes make it look as though the working is being shown, which is exactly why each one needs opening before you rely on it.

Your own documents. You give it a policy, a handbook or a contract and ask it to answer from those files. Attaching a file does not on its own limit the tool to that file. Microsoft’s support page for Copilot, for example, says it can draw on content you attach, your other work data and the web as well as its general training, and that if it cannot find relevant sources it may give a more general answer. So check which sources the tool is set to use, and say what you want in the prompt (Microsoft suggests wording such as “Use this file only”). Even then, if the handbook is out of date, the answer is out of date too, and the tool will still sound sure of it.

All three write smoothly whether they are right or wrong. That is the position this post takes: a fluent answer is not evidence. The check has to come from somewhere other than the tone.

Which kind to use for which work question

Memory tools suit questions where being slightly off costs nothing: rewording an awkward email, reminding you what a term means, suggesting a structure for a report. These are questions about wording, so a shaky memory does no harm.

Search tools suit “what’s out there on this?” questions. Use the summary as a reading list and do the reading yourself.

Document tools suit questions about your own paperwork. “Where does our expenses policy mention mileage?” is a good one, because you can go straight to the clause it points to. “Does our contract allow this?” needs a person who reads the whole clause and owns the answer.

For the wider picture of matching tools to jobs and to the data involved, see the post on choosing AI tools by task and data. This post is narrower: it covers only where the answer comes from and how you check it.

A worked example: checking an answer against gov.uk (fictional question)

Here is a made-up scenario. A colleague asks a chatbot: “What’s the VAT registration threshold, and when do we have to register?” The chat replies with a confident figure and a tidy line saying you must register “straight away” once you pass it.

Now trace it. Open the gov.uk Register for VAT page and read the live text. When we checked it on 28 October 2026, it said:

  • you must register if your total taxable turnover for the last 12 months goes over £90,000, or if you expect it to go over £90,000 in the next 30 days
  • if you went over in the last 12 months, you have to register within 30 days of the end of the month when you went over
  • your effective date of registration is the first day of the second month after you go over

Thresholds change, so read the page on the day you need it rather than relying on this post. Then compare. Does the chat’s figure match the page? Does its timing match? In our fictional case, “straight away” is wrong: the page gives a 30-day window and a separate effective date. Even if the chat had matched the page, all you would have confirmed is that it agreed with the source. Either way, the source did the work.

Use the same method for an employment question. Suppose someone asks the chat how much notice a fictional employee with seven years and three months’ service is owed on dismissal. Go to Acas. Its dismissal or redundancy notice page sets out the statutory minimum for employees: 1 week for 1 month to 2 years’ service, 1 week for each full year from 2 to 12 years, and 12 weeks for 12 years or more. So the statutory minimum here is 7 weeks. The same page says an employer may give more notice than the statutory minimum but cannot give less (this is contractual notice), and it lists exceptions such as gross misconduct. A chat answer that quotes one number and stops has left out the contract, the written statement and the exceptions.

If anyone asks where the answer came from, you can point to the Acas page. The notice for a real person still comes from their contract and your HR records, and someone who owns the decision signs it off.

Questions to keep away from a chatbot altogether

Some questions should go straight to the primary source, whatever type of tool you have:

  • anything that sets a figure: a threshold, a rate, a deadline, a payment
  • anything that fixes a tax position, such as whether a rule applies to you
  • anything that decides an employment matter, such as notice, dismissal or pay

This site’s rule: numbers and decisions come from the system of record (the ledger, the HR system, the bank) and from official guidance. AI drafts words and structure. A person checks and approves.

For finance work, the AI for accountants hub and the post on picking AI for accountants by task and what it sees go further. For people questions, start at AI for HR.

What to check before you rely on any answer

Ask three things, in this order.

  1. Where did this come from? If the tool cannot show you, treat the answer as coming from memory, which means treating it as a guess.
  2. Can I open the source myself? A citation you have not clicked is decoration. Click it, find the line, and check it says what the summary claims. Sometimes the link is real but the claimed sentence is nowhere on the page.
  3. Who signs this off? If the answer feeds a number or a decision, a person with the authority to own it reads the primary source. If nobody is named, nothing has been checked.

For document tools, add a fourth: how current are the files? An answer taken faithfully from an out-of-date policy is still wrong. If you are drafting or refreshing that policy, AI for HR policies shows how to keep a person in charge of the content.

Accuracy also becomes a legal matter once personal data is involved. The ICO’s guidance on AI and data protection says data protection’s accuracy principle “applies to all personal data, whether it is information about an individual used as an input to an AI system, or an output of the system”. The ICO notes that this guidance is under review following the Data (Use and Access) Act, so read the current version on ico.org.uk.

What to paste, and what to leave out of a question

The less you paste, the less can go wrong. Ask in general terms where you can. You rarely need a real employee’s name, salary or start date to ask how a rule works, so use placeholders such as [employee] and [years of service].

Leave out personal data, client details and anything confidential unless your organisation has approved that tool for that data. Whether a provider uses business-tier chats to train its models depends on the provider, and terms change, so check the current terms for the tool you have been given. If you are not sure, ask whoever approves your tools before you paste anything.

Questions

Questions people ask. Every answer open.

Short answers, drawn from this post. The sources are listed above.

Something missing, or out of date? Tell the author.

Email hello@usingaias.com

This post 4

What is the best AI tool for answering questions at work?
There is no single winner. Use the tool your organisation has approved for the information involved, and pick the type to suit the question: a memory tool for wording, a search tool for a reading list, a document tool for finding things in your own paperwork. For figures, tax positions and employment matters, go to the primary source.
Can I trust an AI answer if it gives sources?
Only once you have opened them. Click each citation, find the line and check it says what the summary claims. Sometimes the link is real but the claimed sentence is missing from the page. For thresholds and employment rules, read gov.uk or Acas directly.
Is it safe to ask an AI tool questions about my company's documents?
That depends on the tool, its current terms and whether your organisation has approved it for that data. Check before you upload, and leave personal data out unless you are sure it is allowed. Your document stays the record, so check the answer against the clause itself, and check the file is the current version.
Can AI answer HR or tax questions for me?
It can help you phrase the question or tidy the wording of a reply. The answer itself comes from the primary source (gov.uk, HMRC guidance or Acas), your own records, and a person who owns the decision and signs it off.