Home / Resources

Guides

20 Questions to Ask an AI Provider Before You Sign

17 November 2026 · 3 min read

Before signing with an AI integration provider, ask specific questions in four areas: their experience in production, your data and the architecture, security and compliance, and costs and exit. Good answers are concrete, quantified and verifiable. A vague answer on security or maintenance should give you pause. Here are the 20 questions to ask an AI provider, and the points to carry over into the contract. To score and compare several providers’ answers, use our weighted comparison scorecard.

Experience: 5 questions to ask an AI provider

  1. Which comparable project do you have in production today, for how long, and with what results? A pilot or a demo is not enough.
  2. What problem did you run into on that project, and how did you solve it? An experienced provider will answer without hesitation.
  3. Have you worked in our industry or on a similar process before?
  4. Have you connected AI to our ERP or to legacy systems before?
  5. Who exactly will work on the project? And will it be the same team from scoping to go-live?

Data and architecture: 5 questions

  1. How will you assess our data before starting?
  2. How will the AI respect each user’s access rights? The right answer: it inherits the user’s permissions, never more.
  3. How do you limit wrong answers? Cited sources, checks, human approval.
  4. How do you test a system whose answers can vary? A set of real cases, quality thresholds, repeated tests.
  5. Can we switch AI models without redoing everything?

Security and compliance: 5 questions

  1. Where does our data flow and where is it stored?
  2. Will our data be used to train a model, by you or by the model provider?
  3. How is each AI action logged?
  4. How do you protect against hidden instructions in the documents or emails the AI reads?
  5. What is your role under the GDPR and the AI Act, and what is ours?

Costs, maintenance and exit: 5 questions

  1. What does the price include, line by line? Scoping, build, licenses, model usage, hosting, maintenance.
  2. How do you estimate monthly model usage, and how can we track it?
  3. Which KPIs will measure success, and when?
  4. Who maintains the solution after go-live, and with what response times?
  5. If we stop, what do we get back? Code, documentation, configurations, data.

Answers that should raise a red flag

Topic Red flag Good practice
Experience Only demos or pilots Solutions in production, with a customer contact
Data “Your data is ready” without having seen it A short data audit before any commitment
Security Generic answers, no diagram An architecture diagram and a logging policy
Quality “The AI doesn’t make mistakes” Quality thresholds measured on your cases
After the project Nothing planned after delivery Monitoring, maintenance and a handover plan

What to put in the contract

  • Each party’s role under the AI Act and the GDPR (a data processing agreement under GDPR Article 28 if the provider processes personal data on your behalf).
  • Data location and a ban on using the data to train a model.
  • Ownership of the code, configurations and generated content.
  • Acceptance criteria: test set, quality thresholds, response times, target costs.
  • Maintenance: scope, response times, enhancements.
  • Exit: documentation, knowledge transfer, return of data.

This article offers practical guidance, not legal advice. Have your contracts reviewed by a lawyer.

Our answers at Tocndix

Our Forward Deployed Engineers work embedded in your teams, from scoping to go-live: you know them by name. AI-generated code is reviewed by our experts. We work on a fixed-price basis after scoping, so you know the price before we start.

AI and secure applications · AI integration · Ask us these 20 questions

FAQ

Should you ask every provider all of these questions?

Yes, and ask each of them the same ones: that is what makes the answers comparable. Send them in writing along with your use case.

What if a provider refuses to answer on security?

That rules them out. A serious provider shows how it will protect your data before touching it.

Do these questions also apply to an AI software vendor?

Most of them, yes, especially those on data, security, costs and exit.

All articles

An application to put into production?

30 minutes with an expert to assess your needs.

Book a call