The 12 questions to ask your AI provider before signing

These 12 questions for your AI provider are designed to be asked before signing, when they still cost zero euros. After signing, each one can cost between a thousand and several thousand: the one about unasked maintenance appears as a surprise invoice, the one about data ownership appears when you want to change providers and discover that you can't. This list comes from contracts we've seen go well and those we've seen go wrong, and each question includes its "why": what the answer reveals and what response should make you raise an eyebrow.

A piece of advice: don't send them by email as a questionnaire. Ask them in conversation, in this approximate order, and observe both what they answer and what makes them uncomfortable to answer.

On method (questions 1-4)

1. Which process would you automate first in my company and why? This matters because it separates those who diagnose from those who sell. The only serious answer before measuring is "we don't know yet: first we need to see your processes and their numbers." Anyone who answers with certainty in the first meeting is reciting their catalog, not analyzing your company.
2. How do you measure the result? What baseline will you take and when? This matters because without a baseline there is no demonstrable saving, only feelings. A serious provider proposes measuring hours, volume, and process errors before building anything, and commits to comparing against those figures after three months. If you had to be the one to mention the word "metric," we're off to a bad start.
3. What part of my process will you NOT automate? This matters because the honest answer to a real process is never "everything." Exceptions, ambiguous cases, and sensitive decisions must remain in human hands, and the provider must be able to tell you which ones they are. Anyone promising 100% either doesn't know the trade or is inflating the expected savings.
4. When will the first agent be in production with my real data? This matters because the date betrays the methodology. A first agent on a bounded process should be running in 4-8 weeks after diagnosis; a nine-month plan without production in sight is an eternal pilot under construction. Demand a date and numerical success criteria for that date.

On money (questions 5-8)

5. What is the total cost for the first year: development, maintenance, and model consumption? This matters because the quoted price is almost never the real cost. You must add monthly maintenance and consumption of AI models, which depends on volume, to the development cost. A serious provider gives you all three numbers; a weak one only gives you the first, and you discover the other two in the invoices.
6. What happens if the volume grows? How does the price scale? This matters because agents are hired for today's volume and used with tomorrow's. Ask for the formula: what part of the cost is fixed and what part grows with use. Almost all unpleasant surprises in AI projects live in this unasked question.
7. Does the project go by phases with exit points, or is it all or nothing? This matters because phases are your main risk control tool: if the diagnosis is disappointing, you lose weeks, not months. Be wary of large, "all-in" contracts signed at once "to take advantage of synergies": synergies never arrive, but the commitment clause does.
8. What guarantee do you give on the result? This matters because the answer calibrates the provider's confidence in their own method. No one serious guarantees exact savings figures—it depends on your process and your team—but they can commit to measurable success criteria per phase and not bill for the next one if they aren't met. "AI cannot be guaranteed," flatly, is an evasion.

On ownership and exit (questions 9-12)

9. Who owns the data, prompts, and business logic when we finish? This matters because that's where your operations live. The agent configuration, decision rules, and process knowledge must be yours by contract, in writing. If the provider hesitates here, everything else is secondary: you are buying dependency with a renewal date.
10. How does it integrate with my current systems and what happens if I change my ERP or CRM? This matters because the value of the agent lies in its connections. Ask for technical details: how it connects with your tools—ideally through native connectors with your systems, not by copying data by hand—and what the cost would be to redo the integration if you change a piece of the stack tomorrow.
11. What training will my team receive and who will be able to adjust the system without you? This matters because a system that only the provider knows how to touch is a perpetual rental. Knowledge transfer—documentation, training sessions, internal capacity to make minor adjustments—must be in the proposal with names and hours, not just as an intention.
12. Give me two clients I can call. This matters because it's the cheapest question on the list and the one that returns the most information per second. Don't ask for logos: ask for phone numbers. The provider's reaction to this request—naturalness vs excuses—is worth as much as the calls themselves.

How to read the answers as a whole

No provider fails all twelve, and passing eleven is no guarantee. What you're looking for is the pattern: if evasions are concentrated in the money block, expect surprise invoices; if they focus on ownership and exit, expect dependency; if they fail the method questions, the whole project is a gamble. And there is a meta-signal that summarizes everything: a serious provider enjoys these questions because they allow them to differentiate themselves from smoke-and-mirrors salespeople. The one who gets uncomfortable is answering all of them at once.

These twelve questions cover the conversation prior to signing; the underlying criteria for comparing providers—specialization, cases, team, methodology—can be found in our guide on how to choose an AI consultancy. And if you want to know how we answer them, the details of our method are on our artificial intelligence consultancy page: diagnosis first, phases with exit points, metrics against baseline, and knowledge transfer by contract.

In fact, the best way to come prepared to that conversation is to have your own numbers before sitting down: which processes hurt, how many hours they cost, and what savings are at stake. Our diagnosis gives you exactly that map—and with it, the twelve questions stop being a checklist and become a negotiation with the upper hand.

Frequently Asked Questions

What is the most important question before hiring an AI provider?

If you can only ask one: "who owns the data, prompts, and logic when we finish?". Operational ownership determines whether in two years you will have a system of your own or a dependency with a renewal fee. Everything else can be corrected; this, if signed wrong, cannot.

Which answers should immediately disqualify a provider?

Three deal-breakers: promising to automate 100% of the process, refusing to provide references with phone numbers, and not being able to say who will own the business logic at the end. Each reveals, respectively, commercial inflation, lack of real cases, and dependency by design.

Is it normal for the provider not to guarantee results?

It is normal not to guarantee exact savings figures, because they depend on your process and your team's adoption. It is not normal to refuse to commit to measurable success criteria per phase and exit points if they are not met. The difference between the two separates prudence from evasion.

When should I ask these questions: before or after requesting a quote?

Before accepting any quote and, ideally, in the first or second meeting. The answers change what the proposal should include (phases, metrics, training, ownership), so a quote prior to these questions is almost always incomplete.