What should you ask an AI provider before signing?
Before choosing an AI provider, ask about data, model training, security, quality, support, costs and regulation. Get the answers in writing so you can review the commitments and assess the risks.
Before signing with an AI provider, ask what happens to your data, where the model runs and who can access it. Discuss whether your data is used for training, how errors and incidents are handled, how you can cancel and take your data with you, and how the provider approaches regulation. Ask for clear written answers and record important commitments in the agreement or supporting documentation.
A demonstration or verbal assurance is not enough to establish that the service is suitable. You need to be able to assess it against your data, processes and responsibilities. The requirements depend on the use case, the data involved and your organisation’s role. The questions below can make the conversation more specific. They are general considerations, not legal advice.
Data, privacy and access
Start by identifying the data you provide and the data the service creates, such as prompts, files, responses and usage logs. Ask where it is processed and stored. Have the provider explain whether data stays in a particular region, can be transferred between countries and which other parties are involved. Ask how long data is retained and how you can request its deletion.
Next, find out who can access it. Is access limited to your own users, or might the provider’s staff or other parties view your data? Under what circumstances, such as support or security investigations? Ask how access is restricted, logged and revoked. Also check whether you can manage access within your organisation, for example by employee or department.
Ask what role the provider has when personal data is processed and what agreements are available. A data processing agreement may describe relevant commitments, but check what it actually says and whether it covers the service you will use. Ask which subprocessors are involved, what they do and how changes to the list are communicated. For more background, see AI and the GDPR.
Training, security and incidents
Ask the training question directly: does the provider use your inputs, files or responses to improve its models or services? Does this apply to every type of data and every part of the service? Can you turn this use off, how will the provider confirm your choice, and does it apply to subprocessors as well? Check whether training settings differ by account, user or application.
Ask how the service is secured. Have the provider explain how accounts, administrator privileges, data storage and connections are protected. Find out how security issues are reported and who in your organisation should be contacted. Also discuss the provider’s response to an incident involving your data: when will you be notified, what information will you receive, and how will you work together on investigation and recovery? A contractual commitment to notify you is more useful than a general promise to handle incidents carefully.
Quality, limitations and human review
Ask how the provider tested the service and which tasks it is designed to handle. Request examples of known limitations, such as convincing but incorrect answers, difficulty with particular documents or inconsistent results on similar questions. Ask what the provider’s claims are based on and whether they apply to your language, data and workflow. A general demonstration does not by itself show that the service is reliable enough for your situation.
Discuss how you can check outputs before anyone acts on them. Can you see which sources were used, have responses reviewed and report errors? Can an employee edit or reject a result? Ask what happens when a source is missing, out of date or contradicted by another source. Decide which tasks need human review and who is responsible for it. The provider can give you information about the system, but you must assess how its outputs will be used in your own process.
Documentation and responsibilities
Ask what information you can obtain for your own records and risk assessment. This may include a description of the service, data flows, security measures, known limitations, subprocessors and how changes are announced. Ask whether the provider will tell you about versions or changes that could materially affect how the service works.
Ask how the provider approaches the GDPR and the AI Act. Can it explain its own role and what information it gives customers to help them assess their responsibilities? Can it provide relevant documentation and notify you when the service changes? A provider’s general statement does not automatically establish that your use meets every applicable requirement. That depends on the specific use and division of responsibilities. See also how to maintain an AI register and what the AI Act means for small and medium-sized businesses.
Availability, support and costs
Ask what to expect if the service becomes unavailable or slow. Is there a status page, how do you report an outage and who will help? Discuss how support can be reached and what information you need to provide. Ask what happens if an integration, feature or model changes, and whether you will be notified in advance. If an important process depends on the service, find out how you can keep working temporarily when it is down.
Make commercial dependency part of the discussion. How can you cancel, what notice period applies and what happens to your data afterwards? Can you export data in a usable format, including files, settings or relevant logs? Ask whether there are costs for doing so and how price changes are announced. Check whether you can move your prompts, integrations or workflows to another provider without undue difficulty. The more your process depends on one service, the more important it is to have a workable exit plan.
Questions to use as a checklist
Use this checklist to collect answers. Where possible, ask for a reference to contract language, documentation or a specific procedure.
- Where are my inputs, files, responses and logs processed and stored?
- Who can access my data, and which subprocessors handle it?
- Is my data used for training or improvement, and can I turn that off?
- How are access and security managed, and how and when will I hear about an incident?
- How was the service tested, what are its known limitations, and how can I have outputs reviewed?
- What documentation can I use for my register and risk assessment?
- How do I report outages, and what support is available when the service fails or produces errors?
- How do I cancel, export my data and understand what happens if prices change?
- How do you explain your role under the GDPR and the AI Act?
Record the answers and commitments before rolling out the service more widely. An unresolved issue does not always mean you should walk away, but you should know who carries the risk and what measure addresses it. If an answer is vague, ask for clarification or for a commitment that you can verify later.
Have it reviewed legally
This article provides general information and is not legal advice. Ask a lawyer or data protection officer to assess what the agreements and obligations mean for your specific situation.