
An AI provider does not install like a plumber. It plugs a model into your quotes, your customer files, sometimes your mailbox. From the day you sign, part of your IT risk sits on their side, and you stay accountable to your customers and to the regulator. The questions fit on one page. You still need to ask them in the right order, and know what a good answer contains.
Key Takeaways
- Ten questions are enough, across six families: data, access, hosting, subcontractors, incidents, exit.
- Small and mid-sized companies made up 48% of the victims of attacks recorded by ANSSI in 2025, against 37% in 2024.
- The supply chain has become an attack vector: one compromised provider contaminates its customers in a single operation.
- A vague answer on encryption, access or incident notification is a refusal in disguise.
- Ask for written evidence: data processing agreement, data flow diagram, incident procedure, audit right written into the contract.
Why your AI provider's security is your problem
The risk has changed scale. In its cyber threat panorama, ANSSI states that "the number of attacks doubled between 2020 and 2025". Small organisations are hit first: "very small enterprises (VSEs), small and medium-sized enterprises (SMEs) and mid-sized enterprises (ETIs) are the most affected: 48% of victims in 2025 (against 37% in 2024)".
The agency also points at the vector that matters here. "By targeting a provider or a widely used piece of software (for example, the cloud), attackers can reach a large number of victims in a single operation." In plain terms for a business owner: how seriously your AI supplier takes security is part of your own security setup. You do not delegate the risk, you share it.
The legal framework reinforces that link. When a provider processes personal data on your behalf, it becomes a processor under the GDPR and you remain the controller. The French data protection authority makes the point in its recommendations on developing AI systems (2025): the processor must "ensure that a personal data processing agreement has been signed", "strictly follow the instructions of the controller" and "not use the personal data for anything else". The same page states that when personal data is used to develop an AI system, both the GDPR and the European AI regulation apply.
One more piece of context, often missed: generative AI has hardened the attack itself. According to the same ANSSI summary, phishing messages produced with these tools "get click rates up to 3 times higher than classic emails". A provider that does not frame its own use of AI exposes its customers as much as itself.
The six families of questions
1. What data leaves your premises, and what for?
Ask for the list, not the promise. Which fields, which volumes, which purpose, which retention period. A serious provider can say: "your ticket base, seven fields, never the attachments, thirty-day retention". A provider that answers "everything stays with you" without a data flow diagram is selling an intention, not an architecture.
This question also tests data minimisation. If the provider asks for full access to your CRM for an agent that answers the phone, it collects more than it needs. That is a design flaw, and it is also a legal risk for you.
2. Who can access the data, and under what control?
Leaks rarely come from a break-in at the data centre. They come from a shared administrator account, a former employee whose access was never revoked, a maintenance subcontractor. Make them describe the roles, two-factor authentication, access logs and the removal procedure when someone leaves the project.
A good answer names people and dates. A bad answer is an org chart.
3. Where are the data and the models hosted?
Three questions in one: which country, which cloud provider, which model. An agent built on a US API sends your data outside the European Union on every call, even if the interface is in your language. That is not forbidden, it is a transfer to be framed, and you need to know it before signing.
Also ask who holds the encryption keys, at rest and in transit. If the provider encrypts but keeps the key without a procedure, encryption protects against a stolen disk, not against a bad configuration.
4. Does the provider work with subcontractors?
Almost always yes: a model vendor, a hosting company, sometimes an annotation team. Each one is another link in the chain. Ask for the written list, and ask to be told before a link changes. A provider that adds a vendor without warning breaks your risk analysis without asking your opinion.
5. What happens if there is an incident?
Who detects it, how fast, who tells you, who talks to your customers, who writes the notification. The GDPR leaves you a short window to inform the authority, and that window starts when you become aware of the problem. If your provider discovers the incident and calls you three weeks later, your window is already gone.
ANSSI makes a very concrete recommendation here: "include in supplier contracts an incident notification clause of less than 24 hours and a right to an annual audit". Two lines worth copying as they are into your contract.
6. What happens the day you leave?
It is the question nobody asks, and the one that costs the most. Export format, ownership of the model and the prompts, technical dependencies, deletion of your data at the provider, support period after the contract ends. An AI project you cannot leave is not a project, it is a dependency. exit planning has to be prepared before signature, not after.
The contract carries security better than the brochure
A sales brochure commits nobody. What counts is what the contract states in black and white. Four clauses deserve a careful read before signature.
| Clause | What it should say | Warning sign |
|---|---|---|
| Data | Purpose, fields, retention, fate of backups | Pointer to the model vendor's terms of service |
| Subcontracting | Named list, prior notice of any change | List refused "for security reasons" |
| Incident | Numeric notification deadline, named contact, content of the report | "We will inform you as soon as possible" |
| Exit | Export format, return deadline, certified deletion | No exit clause at all |
The data processing agreement is mandatory as soon as personal data is involved. If it does not exist at signature, it will not exist afterwards.
Answers that should make you walk away
Some phrasings come up again and again, and they all signal the same thing: the provider has a pitch, not a setup.
- "Our servers are secure." Without a hosting company name, a region or a certification, that sentence means nothing.
- "Legal handles GDPR." So does yours, and it cannot work without knowing where the data goes.
- "We cannot show you the architecture, it is confidential." A data flow diagram with no trade secret in it takes ten minutes to draw.
- "We have never had an incident, so we have no procedure." The absence of an incident is not a procedure.
- "Encryption is included." Ask which one, where, and who holds the keys.
These signals are not disqualifying on their own. They point to a company that has not yet met a demanding customer, and you do not want to be the first. To go further on that ground, the list of warning signs of an unreliable AI provider covers the commercial side of the same conversation.
How to check the answers without being an expert
You do not need to understand how a neural network works to audit a provider. You need three things.
Documents. The processing agreement, the data flow diagram, the incident procedure, the subcontractor list. The French authority publishes a checklist for AI project owners: it also works as a reading grid for a customer asking the right questions.
A limited test. Run the setup on a restricted dataset, with degraded cases: an empty file, a scanned document, a request outside scope. You will learn more about robustness in one hour than in three meetings.
Consistency. What sales says, what the contract says and what the architecture shows must tell the same story. When the three diverge, the contract is the only version that protects you, and the architecture is the only one that tells the truth.
Turning these questions into a decision grid
These six families of questions do more than reduce risk: they sort providers. The one who answers with documents is ready to work with a demanding small business. The one who answers with adjectives is not.
At NexeAI, audit and consulting missions rely on this kind of grid, applied to your existing suppliers as well as to projects in progress. The point is not to add another document to your pile, but to know in one meeting whether the setup you pay for holds up. AI agent and automation projects are then built on that base, with limited access and audit trails.
If you first want to compare several approaches before signing, the method to choose an AI agency completes this framework on the project side.
FAQ
Should I require ISO 27001 from an AI provider?
It is a good signal, never sufficient proof. A certification attests to a management system, not to how your data flows through your specific project. A certified provider that cannot show the diagram of your data flows is still a provider that cannot answer. Ask for the certification as a complement, not as a substitute for the six questions.
Can a provider use our data to train its model?
Technically yes, contractually it has to be stated. The French authority reminds that a processor must not use personal data "for anything else" than what you defined. If the provider wants to use it to improve its model, that has to be written, justified and can be refused. By default, assume the answer is no.
What if the provider refuses to answer?
Ask them to specify what is confidential. A blanket refusal on hosting, subcontractors or the incident procedure is an answer in itself: you will not be able to document your own compliance. In that case the question is no longer whether they are competent, but whether you can justify the choice to a customer or an inspector.
Is our data better protected with a model installed on site?
Not automatically. A local model removes the transfer of data to a third party, which solves part of the problem. It moves the rest to your side: updates, logging, backups, access management. A poorly administered local install is often less safe than a hosted service properly run by a provider that documents its practices.
Conclusion
A competent AI provider answers with documents, not adjectives. The six families of questions in this article, applied in one meeting, are enough to tell the two apart. And if the answer is vague, the right decision is to postpone the signature, not to negotiate one more clause.
To move from the grid to a decision on a specific project, get in touch: we start from your real data flows and the contracts you have already signed.


