AI is good at reading large amounts of text and producing a confident summary. For market intelligence that is both the attraction and the problem. A sentence such as “the company is planning a major expansion of its facility” reads the same whether it was stated in a regulatory filing or assembled from a job advert and a hopeful guess.
Commercial teams act on this kind of sentence. They make calls, commit partner time and occasionally say something to a prospective client that turns out to be wrong. So before relying on any AI-assisted intelligence—ours included—it is worth knowing how to test it.
Three kinds of statement
Almost everything a market-intelligence tool tells you falls into one of three categories.
- Fact. Something a source actually says. “The planning authority received an application for this site on this date.”
- Inference. A reasoned conclusion from the facts. “The application, together with the earlier acquisition, suggests a refurbishment is likely.”
- Unknown. Something that matters and has not been established. “No design team has been identified.”
All three are valuable. Inference is where most of the insight is, and an honest list of unknowns is often the most practical part, because it tells you what to go and find out. The harm comes when they are blended, and an inference arrives dressed as a fact.
Five questions to ask
1. Can I see the source?
Not a list of websites the system consulted, but the document behind this specific claim. If a statement cannot be traced to something you can open and read, treat it as an opinion.
2. Can I see the passage?
A link to a forty-page report is not evidence for a one-line claim. Good tools point to the exact words that support what they are telling you, so that checking takes seconds and actually happens.
3. Is the reasoning shown?
When the system moves from “this was announced” to “this work is likely”, the steps should be visible. You know your market better than any model does. If you can see the reasoning, you can correct it; if you cannot, you can only accept or reject the conclusion.
4. Does it say what it does not know?
A tool that never admits a gap is not more reliable, only less candid. Look for explicit unknowns, and for a system that will decline to draw a conclusion when the evidence does not support one.
5. What does a score mean?
A number feels precise. Ask what it measures and how it was arrived at. A fit or confidence score is usually a way of ranking what deserves your attention first. It is not a probability of winning, and it should not be read as one.
Why this matters more for early intelligence
The earlier you hear about potential work, the less of it is confirmed. That is the nature of being early. An opportunity built from an acquisition and two supporting signals is, by definition, mostly inference.
That does not make it less useful. It means the uncertainty has to stay visible, so that the action matches the evidence. A well-supported inference justifies some research and a conversation. It does not justify telling a prospective client that you hear they are building.
A practical test
Take one thing your current tools or processes told you last month and try to trace it back. Where did it come from? Which part was stated, and which part was concluded? What was not known at the time?
If that takes more than a couple of minutes, the intelligence is harder to trust than it looks, whether or not AI was involved in producing it.