How to choose an AI consultant (without buying theatre)

The market has a theatre problem

Every consultancy became an AI consultancy about eighteen months ago. Some of them have genuinely run AI in production; many have run a webinar. For a small or mid-sized business, the cost of choosing wrong is not just the fee — it is a year of momentum and the internal credibility of the next attempt.

I am an AI consultant, so read this with that in mind. But these are the questions I would ask anyone in this market, me included.

Five questions to ask before you sign

What specific problem will be fixed, and by when? A serious answer names a workflow, a metric and a date. 'Unlock AI-driven transformation' is not an answer; it is a mood.

Do you use this yourself? Ask to see the consultant's own operation: their agents, their knowledge base, their review process. Someone selling AI-native working should be running an AI-native business. If the proof is a slide, keep looking.

Who actually does the work? A common pattern in consulting: senior people win the deal, junior people deliver it. Ask who you will be working with, by name, for the duration.

What happens when the AI is wrong? Every real system produces bad output sometimes. You want to hear about tests, human approval gates, logs and a definition of good — governance. If the answer implies the output will simply be trusted, the risk lands on you.

What does the engagement end with? The right answer is a working system your team can run without the consultant, plus the training to run it. The wrong answer is a strategy document and a proposal for phase two.

Red flags

Certainty before diagnosis. Anyone promising specific results before understanding your business is selling, not consulting. The honest version of early-stage advice is a diagnosis with evidence, not a guarantee.

Tool-first pitches. If the recommendation arrives before the questions do — especially if the consultant resells the tool being recommended — you are the distribution channel, not the client.

No mention of your people. AI adoption fails on the human layer more than the technical one. A consultant who never asks who owns the workflow, who checks the output, or how the team feels about it has not read the failure statistics.

Allergic to 'no'. Sometimes the right answer is that AI is the wrong fix for this workflow, or that your data is not ready. A consultant who has never told a prospect 'not yet' is optimising for their pipeline, not your outcome.

Start small, on purpose

Whoever you choose, structure the start the same way: a short, fixed-price diagnosis of one function — sometimes called an AI readiness assessment or AI audit — that produces evidence you own either way. You should leave it knowing which workflows are worth fixing, what is missing, and what a build would cost. Then decide about the bigger engagement with the evidence in front of you.

That is exactly what my AI Reality Check is, and yes, other consultants offer equivalents. Whichever you pick, the principle stands: two weeks of evidence before anyone builds anything. The consultants worth hiring will happily agree to that shape. The theatre merchants will try to skip it.