How to choose an AI consultancy: 10 questions before you sign

Type "AI agency" into a search bar and you get two hundred results promising the same thing: transform your business, automate everything, results in weeks. None of them explain how. Choosing an AI consultancy has turned into an act of faith when it should be an act of judgement: figuring out who actually stands behind the result, who's just selling a nice-looking demo, and who disappears the moment the first invoice clears.

You don't need to be technical to get this right. You need to ask the right questions, in the right order, and recognize an evasive answer when you hear one. That's what follows: ten questions that separate a serious consultancy from someone selling smoke, the warning signs that should end a conversation without further discussion, and what you should demand in writing before signing anything.

Why is it so easy to get this wrong?

Because the sector has no barrier to entry. Anyone with an OpenAI account and a PowerPoint template can call themselves an "AI consultancy" by this afternoon. There's no professional body, no mandatory certification, no simple way to verify experience before you hire. The result: companies paying for projects that never reach production, or that reach it and nobody knows how to maintain.

The other reason is that the buyer — you, the decision maker — tends to evaluate the only things they can evaluate without being an expert: the demo, the rapport, the price. And those three are exactly what a mediocre provider is best at manufacturing. A demo always works; the problem shows up with your real data, your real systems, your real volume.

That's why choosing an AI consultancy shouldn't look like buying software. It looks more like bringing on a partner: someone who will have access to your processes, your data, and — if it goes well — a share of your company's future. Measured against that bar, everything you ask changes.

The 10 questions to ask before you sign

Ask these in the first or second meeting. The quality of the answer — not just its content — tells you more than any case study on their website.

1. Which specific process in our company will you improve, and how will you measure it? If the answer is generic ("we'll make your business more efficient with AI"), there's no project yet, just an intention. 2. Who actually sits with us day to day? Ask for the name and profile of the person doing the work, not the partner who signs the proposal. Many consultancies sell with seniors and deliver with unsupervised juniors. 3. What happens to our data? Where it's processed, with which models, whether it's used to train third-party systems, and what confidentiality clauses they sign. A vague answer here is a compliance risk you'll inherit, not them. 4. What do we get if the project doesn't work out as expected? Every AI project carries real uncertainty. An honest provider admits it and explains how they detect failure early, instead of you discovering it six months in. 5. Who maintains this once the project ends? A pilot nobody maintains degrades on its own. You need to know if that falls to your team, theirs, or nobody. 6. Can you show us a similar project, with someone we can actually call? Not a screenshot. A real reference willing to talk to you for ten minutes. 7. How is cost calculated if the scope changes mid-project? In AI, scope almost always changes, because the first contact with real data reveals things nobody knew. You want the rules of the game before the game starts. 8. What tech stack do you use, and why? You don't need to understand the details, but you should notice whether the answer is tailored to your case or the same slide they show everyone. 9. How much of your own work do you automate with AI so you can charge us less? A consultancy using AI to work faster should pass some of that efficiency on to the price, or at least explain why it doesn't. 10. What happens if we stop working together? Code ownership, access to the data generated, delivered documentation. Ask this while the relationship is still in its courting phase; it's the worst moment to negotiate it mid-breakup.

A consultancy that can answer these ten questions without needing two weeks to "check internally" is already telling you something about how it works.

What red flags should end the conversation?

Some signals don't need further investigation. If they show up, it's reasonable to close the conversation:

• Promises results without knowing your data. Nobody can promise a 30% cost reduction without having seen how you work today. That's a marketing number, not an estimate.
• Only talks about technology, never about process. If the whole meeting goes by without them asking how your team works today, they're not going to fix your problem — they're going to install theirs.
• Doesn't have a single client willing to talk to you. Anonymous success stories ("a leading company in the sector") are free to write. Verifiable references cost reputation, which is exactly why they're worth something.
• Sales team and technical team are different people who don't agree on anything. It's the clearest sign that what's being sold isn't what gets built.
• Pushes you to sign fast with an "end of quarter" discount. A project that's going to live inside your company for years shouldn't be decided by a quarter-end offer.
• Never once mentions GDPR. If they're going to touch personal data, the absence of this conversation is a problem you'll pay for, not them.

What should you demand in writing before signing?

None of this is excessive; it's the reasonable minimum to protect your company:

• Scope with measurable success criteria, not just deliverables. "Cut reconciliation time from 4 hours to 40 minutes" is a criterion; "implement an AI system" is not.
• A data plan: what's used, where it lives, who has access, what happens to it when the contract ends.
• A timeline with review milestones, not just a final delivery date. AI projects are far easier to correct at week three than at month three.
• An ownership clause covering code, fine-tuned models and generated documentation.
• An explicit maintenance plan: who's responsible if something breaks in production two months after the project wraps up.

Senior consultancy, project agency, or freelancer: which fits your case?

Not every format suits every project. This table sums up when each one makes sense:

| Format | When it fits | Main risk | |---|---|---| | Consultancy with senior partners | Projects touching critical processes, sensitive data, or several areas of the business at once | Higher cost than a small agency | | Project-delivery agency | A bounded, well-defined need (a chatbot, a specific integration) | Little support after delivery | | Freelancer or independent professional | Very specific technical tasks with small scope | Depends on one person; no backup if they're unavailable | | Internal team with no prior experience | Never, as a first choice, for your first AI project | Learning curve paid for with production errors |

The practical rule: the more critical the process and the more sensitive the data involved, the more it pays to bring in a partner with accumulated experience and the ability to stand behind the result, not just deliver it.

What does a well-run selection process look like?

A serious selection process usually takes this shape, even when it only lasts two or three weeks:

1. A short diagnosis before any proposal. If you get a closed proposal on the first call, they didn't diagnose anything — they copied a template. 2. Conversations with at least two or three providers, even if you already have a favorite. It calibrates pricing and surfaces differences in judgement you only notice by comparing. 3. Verification of at least one real reference, via a call, not a written testimonial. 4. A bounded pilot before the big commitment, when the project allows it: a four-to-six-week trial on one concrete process, before signing an annual contract. 5. A joint review of the ten questions in this article, with whoever actually decides the purchase, not just whoever managed the meetings.

Frequently asked questions

Is an AI consultancy the same as an AI agency? Not necessarily. Consultancy usually implies diagnosis, ongoing support and accountability for the result over time; agency usually focuses on delivering one specific project and closing it out. Neither term is regulated, so what matters is what each provider actually does, not what they call themselves.

How much does a project with a serious AI consultancy cost? It varies so much by scope that any generic figure would be misleading. What matters isn't the isolated price, but the cost per measurable outcome: which process improves, by how much, and in what timeframe. A serious consultancy helps you build that comparison before asking you to sign.

Can I evaluate an AI consultancy without technical knowledge? Yes. The ten questions in this article are designed for exactly that: they don't require you to understand models or architectures, they require the other party to clearly explain what they'll do, with what data, and with what accountability.

What do I do if I already signed with a provider and I'm seeing warning signs? Check the contract for available review milestones and push for a follow-up meeting with measurable criteria on the table. If there are neither milestones nor criteria, that's the first thing you need to negotiate going forward, even if the contract is already signed.

The honest summary

Choosing an AI consultancy doesn't require predicting the future of the technology; it requires asking the uncomfortable questions before the contract is signed and the budget committed. Any serious provider welcomes being asked, because they're the same questions a team that knows what it's promising asks itself.

If you want a second opinion before deciding, or simply don't know where to start evaluating proposals, the audit gives you an independent diagnosis of your processes and data before anyone sells you anything. Or if you'd rather tell us about your case directly, let's talk.