How do you tell an AI consultant is bluffing?
Short answer
A bluffing AI consultant cannot explain how their approach could fail, gives you a specific improvement number with no baseline behind it, and gets vague the moment you ask about data handling. A genuine specialist will tell you what is out of scope, what they are unsure about, and what they would flag as a risk before starting. If every answer sounds confident and none of them sound cautious, that is the tell.
The AI field moves fast and the vocabulary is easy to pick up without the underlying skill. That makes it an easy place to sound competent without being competent, which is exactly why the usual signs of bluffing show up so often here.
Ask what happens when the approach fails. A real practitioner has failure modes ready to describe, because they have hit them before. Someone who has never actually built and shipped anything tends to answer with reassurance instead of specifics.
Watch for firm numbers with no baseline. A promise of a fixed percentage improvement in efficiency, accuracy or revenue, offered before anyone has looked at your data, is not a forecast. It is a guess dressed up as one. A credible answer says what the number depends on and what would need measuring first.
Ask about data. Someone doing this properly will want to know where your data lives, who is allowed to see it, and what happens to anything sent to a third-party model provider. If that question gets waved away, treat it as a warning rather than efficiency.
Ask what is out of scope. People who know the limits of their own method can describe them without being pushed. People who are bluffing tend to say the approach will handle everything, because admitting a limit feels like admitting weakness.
Jargon density is a weak signal on its own but a useful cross-check. Ask the person to explain their method in plain language, one step at a time. If the explanation only survives in jargon, that usually means there is no concrete method underneath it.
A short paid pilot with a defined, checkable outcome is one of the better tests available. It costs less than a long contract and it forces claims into the open, because either the thing works on your data or it does not.
It is reasonable to ask whether someone works to a recognised framework, such as the NIST AI Risk Management Framework, and to ask them to explain how they apply it rather than just naming it. Confident, specific, checkable answers are the difference. Vague confidence is the bluff.
Related questions
Is it a bad sign if a consultant admits uncertainty?
No, the opposite. Admitting uncertainty about outcomes before doing the work is normal for anyone honest about how AI projects actually go. Total certainty upfront is the thing to be wary of.
Should you be suspicious of a fixed-price quote?
Not automatically, but ask what it is fixed against. A fixed price for a clearly scoped pilot is reasonable. A fixed price for an open-ended outcome, agreed before any data has been seen, is harder to justify.
How do you actually check references?
Ask to speak to a previous client directly, not just read a quote. Ask that client what went wrong during the work, not only what went right. A consultant confident in their own record will not block that conversation.