How to Choose an AI Consultant: A Buyer's Checklist

If you want to know how to choose an AI consultant, the short answer is: ask a handful of blunt, specific questions and judge them on the answers, not the pitch. The market is full of people who can talk fluently about AI and far fewer who have shipped a working system that survives contact with real users. This checklist gives you the filtering questions that separate the two — and a fair warning that the right answers are sometimes uncomfortable.

Why the usual selection process fails

Most buyers evaluate AI consultants the way they evaluate any vendor: credentials, case studies, a polished proposal, a confident demo. The problem is that AI is unusually easy to demo and unusually hard to operate. A slick prototype proves almost nothing about whether the same system will hold up under production load, stay within budget, and remain accurate as your data shifts. So the standard process selects for presentation skill, when what you actually need is operational track record.

Being slightly contrarian here is healthy. The best signal is not enthusiasm about what AI can do; it is specificity about what goes wrong and how they handle it. A consultant who only talks about upside is either inexperienced or selling. We have written before about why a working MVP beats a deck of slides, and the same logic applies to choosing who builds it.

The filtering questions to ask

Use these in a first conversation. You are not looking for perfect answers; you are looking for concrete, specific ones. Vagueness is the tell.

1. What did you ship to production in the last six months?

This is the single most useful question. Anyone can describe an architecture; far fewer have put one in front of real users recently. Ask for something specific they shipped, what it did, and what broke. If the answer drifts to pilots, prototypes and proofs of concept that never went live, you have learned something important. Shipping to production is a different discipline from building a demo, as we cover in our note on running AI agents in production.

2. Will you give me a fixed price for the first 90 days?

A consultant who has done this work before can usually scope a meaningful first phase and stand behind a fixed price for it. Open-ended time-and-materials from day one is a way of transferring all the delivery risk to you. A fixed 90-day price is not magic — it forces both sides to agree on what "done" means, which is exactly the conversation you want to have before money changes hands. Compare their answer against our published pricing so you know what fixed-scope work typically costs.

3. Who owns the code and the models?

You should own what you pay to have built. Some consultants retain the code, lock you into their hosting, or build on a proprietary platform you cannot leave. Ask directly: at the end of the engagement, do I have the source, the prompts, the configuration and the right to run it myself or hand it to someone else? If the answer is anything other than a clear yes, you are buying a dependency, not an asset.

4. How do you monitor cost and accuracy in production?

This question quietly separates the people who have operated AI from the people who have only built it. Models cost money per request and their accuracy can drift as inputs change, so a serious operator has a concrete answer about how they observe both — what they log, what alerts they set, how they catch a quality regression before your customers do. If monitoring sounds like an afterthought, it will be one. This is core to a credible production AI stack.

5. Walk me through your last AI bug.

Ask them to tell you about a recent failure in detail: what went wrong, how they noticed, how they diagnosed it, what they changed. The quality of this story tells you almost everything. People who genuinely operate AI systems have war stories and tell them readily; people who do not will deflect or speak in generalities. A consultant who claims nothing ever goes wrong is the biggest red flag of all.

Vendor-neutrality matters more than it sounds

Be wary of anyone whose recommendation always lands on the same product, platform or model, regardless of your problem. A consultant tied to a single vendor — whether by a partnership, a reseller margin, or simply familiarity — will tend to shape your problem to fit their tool. Vendor-neutral advice means the recommendation follows from your requirements, not from whose logo is on the consultant's website. This is harder to verify, but you can probe it: ask when they would not use their usual stack, and listen for a real answer.

Green flags worth looking for

• Specific, recent production references they can describe in detail.
• Willingness to fix a price for a clearly scoped first phase.
• Plain language about limits — what AI will not do well for your case.
• A clear answer on ownership of code, prompts and configuration.
• Monitoring as a built-in, not a line item added later.

Red flags worth walking away from

• Demos that dazzle but no production references.
• Refusal to commit to any fixed scope or price.
• Ownership terms that lock you into their platform.
• Only upside, never a frank discussion of failure modes.
• The same product recommended regardless of the question.

A note on matching the right service to your stage

The right engagement depends on where you are. If you are still deciding whether AI fits a process, a short audit is the sensible first step. If you have a candidate use case, a proof of concept lets you see real behaviour on your data before committing to a build. Only then does it make sense to move to a full production engagement. Our AI implementation work is structured this way deliberately, and you can see how the stages map to cost on our pricing page.

How to use this checklist

Run two or three consultants through the same five questions and compare answers side by side. The differences will be stark, and they will tell you more than any proposal document. You are not trying to catch anyone out; you are trying to find the small group of people who have actually done the work and will tell you the truth about its limits. As a rough guide, an audit starts from around €2,500, a proof of concept from around €20,000, and production work from €50,000 — figures we publish openly so you can benchmark anyone you talk to. If you would like to put us through the same questions, book a free consultation and ask away.

Frequently asked questions