What to ask before hiring an AI consultant
Eight questions, including the one that rules me out.
Most AI consulting goes wrong before any work starts, in a conversation where nobody asked an awkward question. These are the seven I would ask, and I am one of the people you would be asking.
1. “Who does the work?”
Not who is in this meeting. Who writes the deliverable.
In larger firms the person selling and the person delivering are frequently different, and the gap between them is where quality goes. There is nothing wrong with a team — there is something wrong with not knowing whose name is on the output.
2. “What is the deliverable, and how would I know it was wrong?”
The second half is the useful half. Anyone can describe a deliverable. Being able to describe how it could be falsified means it contains claims rather than impressions.
A weak answer sounds like you'll have a clear roadmap and alignment across stakeholders. Nothing there can be wrong, which means nothing there can be right.
3. “Show me something you shipped that is still running.”
Not a case study. Something in production, and what it does now.
The AI consulting market filled up quickly. RAND puts the AI project failure rate above 80%, roughly double conventional IT, so a consultant with a long client list and nothing running is statistically ordinary rather than unlucky.
4. “What would make you tell me not to build this?”
This is the one I would weight most heavily. Someone who cannot name the conditions under which they would talk you out of the work has an incentive problem, and you will be paying for it.
A good answer is specific: if a rule or a query would do it better, if you cannot name the number that should move, if nobody senior owns the outcome after I leave.
5. “What does this cost to run after you have gone?”
Evaluation, observability, governance, model updates, drift, security review, human quality assurance. If none of that appears in the business case, the return in the proposal is arithmetic rather than forecast.
6. “Who owns this when the engagement ends?”
An AI capability with no internal owner degrades quietly. Nobody kills it; it just stops being maintained, and eighteen months later it is switched off during a migration.
Knowledge transfer should be in the scope, not offered as an extra.
7. “What is the smallest version of this engagement?”
A consultant confident in their work will happily sell you a small piece of it first. One who insists on a large committed programme before demonstrating anything is managing their risk with your budget.
8. “Can you actually staff this?”
I have put this last because it is the one that rules me out, and a list of questions that only ever produces the answer hire this person is not a list worth reading.
I work alone. If what you need is a team of six for eighteen months, an on-call rota, or deep ML engineering at a scale that requires more than one person, I am the wrong call and I will say so on the intro call rather than after the contract. Some work genuinely needs a firm.
The version of this question that separates people honestly: ask what they cannot do. Anyone who answers that fluently has thought about where their work stops. Anyone who cannot name a single engagement they would turn down is telling you their scope is whatever you are willing to pay for.
The pattern in the weak answers
The weak answers all share a shape: they cannot be checked. Roadmap, alignment, transformation, strategic clarity. None of it is false. None of it is verifiable either, and a proposal you cannot verify is a proposal you cannot hold anyone to.
The questions that work are the ones with a wrong answer available.
Common questions
What should you ask before hiring an AI consultant?
Who actually does the work, what the deliverable is and how you would know it was wrong, what they have shipped that is still running, what would make them advise against building, what it costs to run afterwards, who owns it when they leave, what the smallest version of the engagement is, and what they cannot do.
What are the red flags when hiring an AI consultant?
Answers that cannot be checked. Roadmap, alignment and strategic clarity are not false, they are unverifiable — and a proposal you cannot verify is one you cannot hold anyone to. Refusing to sell a small first engagement is the other.
Should an AI consultant ever tell you not to build something?
Yes, and being unable to name the conditions under which they would is the strongest signal against hiring them. If a rule or a query would do the job better, or nobody can name the number that should move, the honest answer is not to build it.
How do you check an AI consultant's track record?
Ask for something in production rather than a case study, and ask what it does now. With AI project failure rates above 80% per RAND, a long client list with nothing running is statistically ordinary.
Cutting enterprise change requests 15 to 7 minutes — an engagement where the unpopular recommendation — cut the scope in half — was the one the numbers came from.