AI Vendor Due Diligence Questionnaire
Procurement and security teams across the UK, EU, Türkiye and the UAE are converging on the same AI vendor risk questions: how suppliers handle data governance, model drift, sub-processors and autonomous agent actions.
SOC 2, ISO 27001 and Cyber Essentials remain foundational compliance frameworks, designed before generative AI moved into production. They cover infrastructure, access control, incident response and operational security. They leave model behaviour, training data provenance, inference data routing, embedded AI sub-processors and agentic permissions unaddressed.
This AI vendor due diligence resource addresses that gap.
How it works
- 01Answer five questions. It takes about 90 seconds.
- 02See how deep your checks should go, and the exact questions to send.
- 03Download the free scoring tool to rate the replies.
Set the depth of your vendor checks
Five questions set the depth of your assessment: Minimal, Light, Standard or Deep. For each, choose the highest answer that applies. Your answers stay in your browser.
Rock 'n Roll CEO · rocknrollceo.ai
AI vendor due diligence questions
Your 35 questions
Copy them into an email or document for the vendor. They sit under the five areas the scoring tool rates, with the same numbers, so each reply is easy to score.
1. Model behaviour and drift
- 1.1What is the documented baseline accuracy or performance rate of the AI system as deployed for our use case?
- 1.2How is model drift detected, and on what frequency?
- 1.3What notification will we receive before a material model change, retraining, or fine-tuning event affects our workflow?
- 1.4Can we opt out of receiving model updates that we have not approved, and for how long?
- 1.5What error or hallucination rates does the vendor publish for the system, and on what benchmarks are they measured?
- 1.6What is the vendor's process for handling adverse outputs reported by customers, and what is the disclosure timeline?
- 1.7What is the vendor's published policy on AI literacy obligations under EU AI Act Article 4, and does the vendor support customer training requirements?
2. Training data and IP exposure
- 2.1What documentation of training data composition, sources, and licensing can the vendor provide?
- 2.2Is our data used to train, fine-tune, or improve your models in any form, and is this opt-in or opt-out by default?
- 2.3If our data contributes to model improvement, how is it isolated, anonymised, and retained, and for how long?
- 2.4What is the vendor's policy on customer data appearing in outputs delivered to other customers, and what controls prevent this?
- 2.5What IP indemnity is offered on outputs the vendor generates for us, and what is the financial cap?
- 2.6What contractual audit rights do we have over training data composition or governance practices?
- 2.7What is the vendor's policy on training with publicly available content under restrictive licences (for example CC BY-NC, GPL or copyrighted images)?
3. Prompt logging and data residency
- 3.1Where are prompts and completions stored, in which jurisdiction, and for how long?
- 3.2Who, within the vendor and any sub-processor, has access to prompt and completion data, and under what controls?
- 3.3Is prompt and completion data encrypted at rest, and are the encryption keys held by the vendor or by the customer?
- 3.4What happens to prompt and completion data at the end of the contract, and what is the documented deletion timeline?
- 3.5Does the vendor route inference data through any third-party model provider, and where is that provider located?
- 3.6Can sensitive prompt data be flagged, masked, or redacted before transmission to the model?
- 3.7Is prompt and completion data ever used for vendor product analytics, model improvement, or commercial reporting, and is this contractually restricted?
4. Embedded AI and sub-processors
- 4.1Provide a current inventory listing every AI feature, model, and sub-processor used to deliver this product, including any that ship as part of default functionality.
- 4.2Which foundation model providers are used in this product, and is this disclosed in the contract?
- 4.3What is the vendor's process for vetting new AI sub-processors before they are introduced into the product?
- 4.4How is notification of new sub-processors delivered, and what objection period is provided?
- 4.5What is the vendor's incident response process when a sub-processor is compromised, and what is the customer notification timeline?
- 4.6What contractual liability flows through to the customer if a sub-processor breach affects the vendor's service?
- 4.7Does the vendor maintain a documented AI risk management framework (NIST AI RMF, ISO/IEC 42001, or equivalent), and can certification or alignment evidence be provided?
5. Agentic permissions
- 5.1How is agent identity provisioned, and is each agent assigned a unique credential?
- 5.2What scope, permissions, and tool access does each agent hold, and how is this documented?
- 5.3Is every agent mapped to an accountable human owner, and is this mapping auditable?
- 5.4Can agents be limited to read-only actions, with a human approval step required before any write or send action?
- 5.5What is the documented rollback procedure for unintended agent actions, and is rollback supported by the vendor or by the customer?
- 5.6How are agent actions logged and monitored, what level of detail is captured, and how long are logs retained?
- 5.7What is the vendor's recommended approval process before deploying a new agent into a production workflow?
Score the replies with the free toolkit
When the vendor replies, mark each answer Green, Amber or Red in the scoring tool. It scores all five areas and tells you whether to go ahead, negotiate or look elsewhere, for up to three vendors at once.
A five-page guide explains each area, and the Audit Trail tab keeps a record of your decision for any buyer or regulator who asks how it was made.
what's inside
AI Vendor Due Diligence Questionnaire (PDF)
Questionnaire & Scoring Tool (XLSX)
Get instant access
If the button stays grey, please email us at [email protected] and we will reply directly.