AI vendor due diligence is the assessment carried out before buying a tool or service that uses artificial intelligence. It checks how the vendor handles the data you send, whether it uses that data to train models, how it tests security, what documentation it provides about the model, how it reports incidents and which role it claims under the EU AI Act. The result informs the buying decision and the contract.
How to use
- Enter the vendor, the product, the criticality of the use and the data that will be shared.
- Answer the fifteen questions with yes, partly, no or not known, based on the vendor answers and documents.
- Read the score, the residual risk, the recommendation and the red flags.
- Send the open questions to the vendor and take the suggested clauses into the contract negotiation.
- Export as DOCX, CSV, PDF or Markdown to keep as evidence.
What each field asks for
| Field | What to enter |
|---|---|
| Vendor | Name of the company being assessed. |
| Product or service | E.g. contract review assistant |
| Criticality of the use | How much an error or a vendor failure would affect people or operations. |
| What the tool will be used for | Who uses it, with which data and for which decision. |
| Data shared with the vendor | Tick everything that leaves the organisation. |
| Does the vendor publish a responsible AI policy or principles? | |
| Does it run an AI management system (ISO/IEC 42001 or equivalent)? | Certified or being implemented counts as partly. |
| Is a data processing agreement signed or available? | It defines the vendor as processor and the instructions it follows. |
| Is the list of sub-processors published, with notice of changes? | |
| Do international transfers have safeguards? | E.g. standard contractual clauses or an adequacy decision. |
| Is customer data kept out of model training (or is there an opt-out)? | |
| Does it hold a security certification or an independent report? | E.g. ISO/IEC 27001 or an independent audit report. |
| Does it test for prompt injection, abuse and misuse (red teaming)? | |
| Does it provide model documentation (a model card or equivalent)? | |
| Does it publish accuracy and bias evaluation results and known limitations? | |
| Does the contract commit it to report incidents within a set time? | |
| Are usage logs available to the customer? | |
| Does it state its own role and the system classification under the EU AI Act? | E.g. provider of a limited-risk AI system. |
| Does it support transparency duties (telling people it is AI, labelling generated content)? | |
| Does it return and delete the data at the end of the contract, with confirmation? |
Example
The legal team of a fictional company assesses Lexora, a cloud contract review assistant that will receive contracts with names and contact details of representatives and confidential information. With a data processing agreement, certified security and incident notice, but partial answers on training and transfers and no bias evaluation results, the score comes to 67 out of 100, the residual risk is medium and the recommendation is to approve with conditions, with eight follow-up questions.
When to use
- Before buying a generative AI tool or a service that uses AI.
- When renewing contracts with vendors that have started using AI.
- When the AI inventory shows a tool used by teams without any assessment.
- To standardise assessment across procurement, security, privacy and legal.
Frequently asked questions
What should I ask an AI vendor?
The essentials: how it handles the data you send, whether it uses it to train models, where the data is stored, how it tests security and misuse, what documentation it provides about the model, how it reports incidents, which role it claims under the EU AI Act and how it deletes data at the end of the contract. The questionnaire covers these points in fifteen questions.
How is the score calculated?
Yes is worth 2 points, partly is worth 1 and no or not known are worth 0. Data protection questions weigh more when personal or special category data is shared, and security questions weigh more with confidential information or source code. The criticality of the use and the red flags adjust the residual risk.
What is a red flag?
An answer that calls for attention on its own: personal data without a data processing agreement, no assurance that data stays out of training, no deadline for reporting incidents, international transfers without a safeguard or no deletion of data at the end of the contract.
Does "not known" count as "no"?
Yes, in the score. Missing information is a risk until the vendor answers; that is why every answer other than yes becomes a follow-up question to send to the vendor.
Does the result replace legal and security review?
No. It is an educational tool that organises the vendor answers; it is not legal advice or an audit. The final decision belongs to the security, privacy and legal teams of your organisation.
Which DPCert certification covers this topic?
Managing third-party and vendor AI is a domain of RAIP Responsible AI Practitioner. Supply chain assurance is covered by AUDAI Responsible AI Auditor, whose prerequisite is RAIP or DPAI.