When a customer, auditor or procurement team reviews a high-risk AI system, the questions tend to map onto the same provisions. Here are 14 of them, with the evidence worth keeping for each.
Use it as a supplier questionnaire in either direction: as a provider, to check what you can show; as a deployer, to check what your vendor can show.
| Question | Provision | Evidence to keep |
|---|---|---|
| Is this system high-risk, and who decided? | Art. 6, Annex III | Dated classification memo: intended purpose, Annex III area, rationale, named owner, review date. |
| Are you the provider or the deployer of this system? | Art. 3(3)–(4), Art. 25 | Role map per system and per customer, with the matching contract clauses. Check the Art. 25 triggers: your own name or trademark on the system, a substantial modification, or a change of intended purpose. |
| Is risk managed as a continuous process? | Art. 9 | Risk register covering intended use and reasonably foreseeable misuse; named risk owners; test results against defined thresholds; residual risks; scheduled reviews. |
| What data was the system built and tested on? | Art. 10 | Dataset documentation: provenance, selection criteria, bias examination, known gaps. |
| Is there technical documentation? | Art. 11, Annex IV | A documentation package kept up to date and available to authorities. |
| Does the system keep logs? | Art. 12 | Logging specification: events recorded, retention period, access controls. |
| What do deployers receive to use it correctly? | Art. 13 | Instructions for use: intended purpose, performance levels and limitations, human-oversight measures. |
| Can a person meaningfully oversee and override it? | Art. 14 | Oversight design: who can intervene, how, and with what information; measures against over-reliance on the output (automation bias). |
| How accurate, robust and secure is it? | Art. 15 | Test reports, declared accuracy metrics, resilience and cybersecurity testing. |
| Is there a quality management system? | Art. 17 (see Art. 40) | QMS documentation. A presumption of conformity can only come from harmonised standards whose references are published in the Official Journal; check what is currently cited before relying on one. |
| Has conformity been assessed and declared? | Arts. 43, 47–49 | Conformity assessment record, EU declaration of conformity, CE marking and registration where required. |
| What happens after launch? | Arts. 72–73 | Post-market monitoring plan; serious-incident procedure and contacts. |
| Do the people operating it know what they are doing? | Art. 4 | AI literacy training records by role. |
| Do users know they are dealing with AI? | Art. 50 | Disclosure design; content-marking approach for generative systems. |
Every other row depends on the second one. A deployer can become a provider by putting its own name or trademark on a high-risk system, by making a substantial modification, or by changing the intended purpose of a system so that it becomes high-risk. We cover those triggers in You're Probably Both: Provider and Deployer Under the EU AI Act, and the terms are defined in the glossary.
If you want this as a structured vendor review rather than a list, the EU AI Act Vendor Risk Assessment & Due Diligence Toolkit is available on Gumroad. For questions about a specific use case, write to venactis@proton.me.
Dates as published in the Official Journal. Not legal advice.