You have three proposals for an AI project and they look alike. The same promises, the same words, a price that ranges from one to three. The real difference does not show in the sales pitch. It shows six months later, when you need to change the agent, switch model providers, or explain to a client why the answer was wrong.
Choosing an AI agency is therefore not about comparing catalogues. It is about getting your contact to say what they will do when the system gets it wrong, who will own the code, and what running it will cost in a year. The questions below serve that purpose. A serious provider answers them plainly, before even sending a quote.
Key Takeaways
- An AI quote can only be compared when scope, volumes and maintenance are described the same way.
- The decisive questions cover the problem to solve, the data, code ownership, exit terms and maintenance.
- Ask for a written, dated, costed answer wherever the salesperson answers with a demo.
- A provider who refuses to discuss leaving the contract, or does not know what happens to your documents, should be ruled out.
- A short paid first step beats a yearly commitment signed on a hunch.
Price comes last, not first
Comparing two quotes that do not describe the same work leads nowhere. The cheaper one may include maintenance while the other bills it separately, or handle a single process while the other claims four. Before looking at the amount, get the scope, the number of users, the volume of documents processed and what remains on your side spelled out.
This caution is not gratuitous distrust. In 2024, Gartner predicted that at least 30% of generative AI projects would be abandoned after proof of concept by the end of 2025, citing poor data quality, inadequate risk controls, escalating costs or unclear business value. Four causes, and none of them about the quality of the model. All of them about framing and contracts.
Maturity is uneven across French small and mid-sized companies. According to Insee (July 2026), in 2025, 18% of French companies with 10 or more employees reported using at least one artificial intelligence technology, 8 points more than in 2024. The rate reaches 15% for companies with 10 to 49 employees and 58% for those with 250 or more. In other words, your agency rarely works for a market where everyone already knows what they are doing. Framing questions are not a large-company luxury: they compensate for experience your team does not have yet.
A short audit on one specific process lets you cost the starting point before ordering anything. It is also a way to test a provider's method on a modest commitment: commission a focused audit.
What the agency must describe before talking about tools
A provider who opens with the name of a model or a platform has not listened to your problem yet. The three descriptions below should fit on one page and be approved in writing.
The problem, stated in measurable terms
A solid agency restates your need in everyday figures: files handled per week, time spent on each, errors observed, current response time. Then it names an indicator to move and a date to check it.
Ask directly: which indicator must change, by how much, and when will we measure it together? An answer such as "you will save time" is not a target, it is a slogan. If nobody can name the indicator, the project has no finish line and you will never be able to say whether it succeeded.
The scope of the first deliverable
Ask what is included and, above all, what is excluded. A first batch should cover one process, one team and a limited volume. Formulas that promise a full transformation in a single phase are a warning sign: they make acceptance impossible to declare and the exit impossible to negotiate.
What happens when the system gets it wrong
This is the question few executives ask, and the one that separates providers. A probabilistic model makes mistakes. What matters is how the error is handled: does the system detect that it does not know, hand over to a human, log the decision, alert a manager?
Ask for the error rate they accept, how it is measured, and the escalation procedure. An agent that answers everything without ever referring to a person is more dangerous than one that stops. That design choice belongs at the start, not as a fix after go-live: design an agent that escalates to a human.
Data and the real cost of running it
Where documents live, and who can access them
Your documents, client exchanges and internal files become inputs to the system. You need to know where they are stored, with which host, for how long, and who can read them. Ask for the list of subcontractors and the clause covering transfers. An answer such as "it is secure" with no document is not an answer.
The most useful question is also the simplest: is my data used to train a model? The answer should be no, or clearly framed and accepted in writing. If your contact hesitates, assume the answer is yes.
Usage cost, kept separate from build cost
An AI project has two prices. The build price, which appears in the quote, and the usage price, which depends on the volume processed and on model tariffs. The second is often left out of the budget, and it does not disappear at delivery.
Ask for a simulated monthly invoice at three volumes: your current volume, three times that, and ten times. You will immediately see whether the economics hold as usage grows, and whether the agency has actually measured cost per document processed. Also ask who pays the model provider subscription: you, directly, or the agency with an undisclosed margin.
What the answers should contain
| Question | Reassuring answer | Answer that should alert you |
|---|---|---|
| Monthly cost | a figure per usage volume, with the calculation | "it is included", with no written cap |
| Data | list of subcontractors, hosting location, retention period | "it is secure", with no document |
| Errors | measured rate, escalation procedure, decision log | "the model improves on its own" |
| Exit | export of code, data and documentation | "we will see when the time comes" |
To place these answers against market offers, a comparison grid helps more than a feature catalogue: compare two AI solutions with the same grid.
The contract: ownership, exit, maintenance
Who owns the agent
Three objects must belong to you or be clearly licensed: the code, the instructions given to the model, and the chains that do the work. A contract granting a right of use does not make you the owner. The difference is paid on the day you want the agent modified by someone else, or simply hosted at your premises.
Have the assignment of rights written down, along with the list of third-party components and their licences. If the agency relies on a proprietary platform, ask what remains usable if the price changes or the service shuts down.
How you get the system back
Reversibility is prepared before signing, not at the moment of divorce. Ask for the exact procedure: which files are handed over, in what format, within how long, and at what cost. The right test is simple: can the system be rebuilt in a clean environment from what you were given, without the agency's help?
Many projects fail that test without anyone ever asking for it. An agent that only runs in a provider's account, with credentials they alone hold, is not your tool. It is their service.
Maintenance, costed from the start
Models change, interfaces evolve, providers deprecate versions. An agent delivered without maintenance degrades slowly, and nobody notices before a user reports an absurd answer. Ask for a maintenance contract with its scope, its response time, and how model provider price changes are handled.
If the agency answers that "it runs on its own", ask a second question: who warns the users the day the provider changes its API? Silence is usually the answer.
Compliance, people and warning signs
Your obligations, not only theirs
An AI system processing personal data remains subject to the GDPR, and the European AI regulation adds obligations that apply in stages depending on the use case and the level of risk. Ask who writes the system documentation, who carries out the impact assessment when one is required, and how the provider supports you during an audit or a client complaint.
You remain the data controller. An agency that treats compliance as an administrative formality leaves you alone with the risk.
Who actually works on the project
Ask for names and roles, the share that is subcontracted, the share produced by generative tools, and above all who will still be there in six months. An AI project depends as much on knowledge of your business as on technical skill. If the team changes after signing, the carefully built framing evaporates.
Ask about training as well. The Insee figures show it: among companies that do not use AI, 54% cite a lack of expertise, and more than half of those that do use it say the same gap holds them back. A tool delivered without adoption produces no lasting gain. Building skills is part of the deliverable: train the team that will use the tool.
Three answers that should stop the conversation
- "Your data is none of our business, the provider handles it." Nobody then carries responsibility for the processing.
- "No need for a maintenance contract, it runs on its own." The day the model changes, you have no recourse.
- "We will train you after go-live." The team discovers the tool when it has to use it under pressure.
Conversely, a credible provider agrees to show you a comparable system in operation, gives you references you can call, and offers a short paid first step. To judge that kind of work, nothing replaces seeing a process that is already automated: browse concrete use cases.
FAQ
Should I pay for an audit before signing a quote?
A short, billed audit with a fixed time frame prevents you from signing a quote built on assumptions. It should deliver a description of the process, a volume, an estimated running cost and a recommendation. Refuse the free three-day audit: it is really a sales study, and its scope is calibrated to end in a proposal.
How long should a first AI project last?
Long enough to measure an effect, short enough to stop without major loss. Ask the agency to describe an end date, a success criterion and a stop criterion. If it cannot name the point at which the project would be considered a failure, you do not have a project, you have a subscription.
What if the provider refuses to hand over the code?
Look for the reason. A proprietary platform can justify a licence rather than an assignment, but the absence of a negotiated exit cannot. At the very least, ask for ownership of your data, the instructions written for your business, and a guarantee of a full export. If the refusal is total, move to the next provider: the cost of switching will always exceed the price of signing.
Can an AI agent run without maintenance?
No. Models evolve, interfaces change, prices move and uses shift. Maintenance can be light, it cannot be zero. Plan at minimum a periodic review of the error rate and a named contact for incidents.
How do I check an AI agency's references?
Ask for two references on projects close to yours by business, not by technology, and call them without the agency. Three questions are enough: is the system still running, who maintains it, and what would you do differently? A reference that cannot answer the third question has not really lived through the project.
Conclusion
Choosing an AI agency means choosing a method. The questions to ask before signing are not meant to trap the salesperson: they check that someone has thought about scope, data, usage cost, ownership and exit. A provider who answers precisely on those five points will probably save you time, even if their quote is not the cheapest.
At NexeAI, the work starts with an audit of the process and the data, before any agent is built. That step is there to decide whether the project deserves to be launched, and under what conditions. If you are preparing a first project and want to test your questions against a structured grid, let us talk about your case.
