Marta
What not to ask it
Questions that look like all the others and where the answer is not worth the same.
With a tool that answers almost anything, the hard line is not what it knows: it is which kinds of question its answer is worth as much as it looks. These four are answered as fluently as the rest and should not be used the same way.
The four
| Question | Why it fails | Who to ask |
|---|---|---|
| "Does this comply?" | It depends on your case, sector and country, and it changes | Your adviser, with the document in front of them |
| "Should I accept this contract?" | It is a decision with consequences, not a fact | Whoever answers for that decision in the company |
| "What will happen if…?" | Predicting is not reading: there is no fact to look up | Nobody: that is decided, not consulted |
| "Is this supplier trustworthy?" | It mixes data with judgement about people | Your own history, read by you |
Important
All four share what makes them dangerous: **the answer sounds exactly as confident as when it tells you what expires next week**. There is no visible marker separating a looked-up fact from a constructed opinion, so you have to make that distinction when asking, not expect it in the reply.
The same question, rephrased to be useful
- 1
Instead of "does this comply?", ask what it contains
"What does this document say about deadlines and penalties?" is lookup-able and useful.
- 2
Instead of "should I accept?", ask what to look at
"Compare it with the previous contract and tell me what changed" prepares the decision without taking it.
- 3
Instead of "what will happen?", ask for history
"How many times was this supplier late last year?" does have an answer.
- 4
And instead of judgements, ask for dated facts
Facts you can show; a judgement you cannot.
Watch out
The case needing most care is the regulatory one, because it tempts most. A reasonable-sounding answer about whether something complies can lead you not to consult and to decide on it — and if it turns out that in your specific case the answer was different, "the assistant said so" covers you with nobody.
Where it is very good
That is the ground where it genuinely saves hours, and it is much wider than it seems once you get used to asking about facts rather than opinions.
Worth knowing
Asking consumes, including when the answer is useless. Rephrasing a badly framed question three times costs three; framing it as a factual query first time usually gives a better answer and one action.
›What if I insist on asking something legal?
It will answer, and that is precisely why you should be clear what weight to give it.
›Can it cite where something came from?
For what is in your documents, yes. For a general opinion, there is no source to cite.
›Does more context help?
For factual queries, a lot. For decisions, context was never the problem.
A real case
The situation
Someone asks whether a supplier contract meets what a client requires.
What you do
- Rephrases: asks what the contract says on those specific points
- Takes that summary and the document to their adviser
What you get
The adviser's consultation takes ten minutes instead of an hour, and the decision is taken by whoever should.
The situation
It is asked for a decision that belongs to a person.
What you do
- Uses the answer as support and decides
What you get
Responsibility stays where it should.
The situation
It is asked for a legal interpretation.
What you do
- Consults the adviser for that
What you get
Each question goes to whoever can answer it.
The situation
It is asked about something not in the documents.
What you do
- Checks whether the information exists
What you get
Expectations are set.
The situation
It is asked to invent a missing figure.
What you do
- Records the gap rather than filling it
What you get
What is missing shows as what it is.
The situation
It is asked about data from outside the company.
What you do
- Checks its scope
What you get
The answer is read knowing what it covers.
This article answers
- can i ask the ai legal questions
- will the ai tell me if this complies
- asking the assistant for opinions
- limits of what it can answer