Assessing an AI vendor: twenty questions for the CISO, DPO and procurement
Short answer
Assess an AI vendor in five areas: what happens to your data (training, retention, location), who can access it (subprocessors, foreign jurisdiction), how it is secured (certifications, logging), what the contract covers (data processing agreement, audit, exit) and which role the vendor has under the AI Act. Have the answers confirmed in writing: marketing copy is not a contract term.
Data: what happens to what you enter?
- Are our inputs and outputs used to train or improve models, and is that excluded in the contract?
- How long are prompts, outputs, uploaded files and logs retained, and can we configure that?
- Where does processing physically take place, and where are backups stored?
- Can we have data fully deleted on request, and how is that confirmed?
Access: who can get to it?
- Which subprocessors process our data, and where are they established?
- Is the vendor or a subprocessor subject to non-European law that can compel disclosure, such as the CLOUD Act? See The CLOUD Act and your European data.
- Which of the vendor’s staff can view our content, in which situations, and is that logged?
- Do we run on shared or on dedicated infrastructure?
Security: how is it handled?
- Which certifications does the vendor hold (ISO 27001, SOC 2, sector standards such as NEN 7510, HDS or C5), and what is in scope?
- How is the service protected against prompt injection and misuse? See Prompt injection.
- Do we get an audit log of usage, and can we export it to our SIEM?
- How and within what timeframe does the vendor report a security incident?
- Can we connect access to our identity provider, with roles?
Contract, exit and the AI Act
- Is there a data processing agreement that meets Article 28 GDPR?
- Do we have audit and inspection rights? For financial entities, DORA requires this explicitly.
- What happens on termination: how do we get our data, automations and configuration back?
- Can we switch models without rebuilding everything? See Choosing open-weight models.
- Which role does the vendor have under the AI Act (provider of a system or of a model), and which documentation does it supply?
- Which models are used, from which developer, under which licence?
- How are we informed of changes to models, terms and subprocessors?
Record the answers in your AI register and in the DPIA.
Frequently asked questions
Questions on this topic
Which questions matter most for an AI vendor?
Whether your data is used for training, where processing takes place, which subprocessors and which foreign law are involved, whether there is a data processing agreement and an audit log, and how you get out on termination.
Is an ISO 27001 certificate enough?
It is a good sign, but check the scope: is the AI service you buy within it? And a certificate says nothing about training on your data, foreign jurisdiction or exit. You have to ask those questions separately.
How do I assess an AI feature in existing software?
With the same questions. Ask the software vendor which model the feature uses, with which party it runs, whether your data is used for it and whether you can switch the feature off.
What changes if the AI runs on your own infrastructure?
Many questions about data and access fall away: no training at a third party, no foreign jurisdiction over the content, no subprocessors in the processing. The questions about security, logging, updates and exit remain.
Sources
- GDPR (Regulation (EU) 2016/679), Article 28
- DORA — Regulation (EU) 2022/2554, Article 30
- NIS2 Directive — Directive (EU) 2022/2555, Article 21
This article is general information, not legal advice. Rules and guidance change; check the sources listed or ask your lawyer or DPO if in doubt.