Login Get a Demo
01 Free template

Vendor AI Risk Questionnaire

Thirty-two questions to send any vendor that runs AI on your data, each with the evidence a good answer includes. Free, ungated and mapped to the EU AI Act, NIST AI RMF and ISO/IEC 42001.

The short answer

What is a vendor AI risk questionnaire?

Last updated
Read time
10 min read
From
Cranium AI
Share on LinkedIn

A vendor AI risk questionnaire is a structured set of questions a buyer sends to a supplier to learn which AI models touch its data, how they are tested and controlled, and what evidence backs each answer.

It covers models, data use, subprocessors, change notice, security testing, runtime controls, compliance and human oversight.

Key takeaways
  • This template has 32 questions in eight groups, free to copy with no form to fill in.
  • Each question lists the evidence a good answer includes, so you can tell proof from assurance.
  • The groups map to EU AI Act Articles 25 and 26, NIST AI RMF GOVERN 6 and ISO/IEC 42001.
  • A questionnaire is self-reported and point-in-time. Pair it with evidence requests and notice of change.
By the numbers Each figure links to its sourceAll sources 
02 How to use it

How do you use this questionnaire?

Send it before you sign, and again when a vendor’s AI changes. Four rules keep it from turning into one more form that is read and filed.

01

Tier first

Send all eight groups to vendors whose AI touches sensitive data or acts on its own. For low-exposure tools, groups 1 to 4 are the floor.

02

Ask for the artifact

Each question lists what a good answer includes. If a reply has no document, date or named control behind it, treat it as unanswered.

03

Set a date and a trigger

Give a due date, then re-send the affected groups when the vendor changes a model, provider, subprocessor or data use.

04

Write answers into the contract

Notice periods, training restrictions and audit rights only hold when they are contract terms. Hand groups 2, 4 and 6 to legal.

See the due diligence checklist
03 The questions

What questions should you ask an AI vendor?

Thirty-two questions in eight groups. The right-hand column is the test: a good answer comes with evidence, not assurances.

Free to use. All 32 questions are on this page, with no form to fill in. Copy them into your own assessment tool and edit them to fit your risk tiers. This is a starting template, not legal advice: have counsel review anything you turn into contract terms.

Models and providers

Which AI touches your data, and who built it.

4 questions
QuestionWhat a good answer includes
Q1Which AI models and model providers process our data in this product, and for which features? A list of each AI feature with the model name and version, the provider and where it is hosted. “A third-party LLM” is not an answer.
Q2Do you host the models yourself or call them through a provider’s API, and in which regions? The hosting model for each feature, the regions used, and whether a request can leave the region we contracted for.
Q3Which open-source or pre-trained models do you use, and how do you vet them before use? Each model’s source and license, the checks run before adoption such as file scanning and pinned versions, and who approves a new model.
Q4Can we turn AI features off, or opt out of specific ones, without losing the core service? A named admin control or contract option per feature, and whether new AI features ship on or off by default.

Maps to: NIST AI RMF GOVERN 6.1 and MANAGE 3.2 · EU AI Act Article 25(4)

Data use and training

What happens to what you send, and whether it trains anything.

4 questions
QuestionWhat a good answer includes
Q5Is our data, including prompts and outputs, used to train, fine-tune or evaluate any model, yours or a provider’s? A yes or no for each data type, the contract clause that makes it binding, and the same commitment from each upstream model provider.
Q6What do you and your model providers retain from our prompts, outputs and files, and for how long? Retention periods for each data type, the provider settings that apply, and how deletion is evidenced.
Q7How is our data separated from other customers’ data in prompts, caches, indexes and fine-tuned models? The isolation design for each store, whether any model is tuned on pooled customer data, and test results showing one tenant cannot retrieve another’s content.
Q8Which categories of personal, regulated or confidential data may the AI features process, and what stops the rest? The data categories in scope for each feature, the controls that filter or mask the others, and the agreement that covers each.

Maps to: NIST AI RMF GOVERN 6.1

AI subprocessors and supply chain

Who else is in the chain, and what the system is built from.

4 questions
QuestionWhat a good answer includes
Q9Which subprocessors are involved in delivering AI features, and what does each receive? A current list naming each AI subprocessor, its role, the data it receives and its location, plus how we are told when the list changes.
Q10Can you provide an AI bill of materials (AI‑BOM) for the system, in CycloneDX or SPDX? A machine-readable file listing models, datasets, frameworks and key dependencies with versions, and the date it was generated.
Q11Which agent frameworks, tools, plug-ins or MCP servers can your AI call, and what can they reach? An inventory of each tool or connector, the systems and data it can reach, and the permissions it runs with.
Q12How do you track and fix vulnerabilities in AI libraries, models and dependencies? The scanning method and frequency, remediation targets by severity, and a recent AI-component fix with dates.

Maps to: EU AI Act Article 25(4) · NIST AI RMF GOVERN 6.1 · ISO/IEC 42001 · What is an AI‑BOM?

Change notification

How you find out before the AI you approved becomes something else.

4 questions
QuestionWhat a good answer includes
Q13How and when will you tell us before you add an AI feature or change a model, provider or subprocessor? A notice period written into the contract, the channel used and a named contact. “See our release notes” is not notice.
Q14What counts as a material change to your AI, and who decides? A written definition that covers model swaps, new data uses, new subprocessors and new autonomy, and the role that classifies each change.
Q15Do you re-test and re-assess after a model or provider change, before it reaches customers? The regression, safety and security tests run on change, the release criteria, and the record from the last change.
Q16How can we see which model version produced a given output? Version identifiers in logs, release notes or an API field, and how long that history is kept.

Maps to: NIST AI RMF MANAGE 3.1

Security testing and red teaming

Whether anyone has attacked it on purpose, and what they found.

4 questions
QuestionWhat a good answer includes
Q17What adversarial testing has the AI system passed, and when was the last round? Test dates, who ran them, the attack categories covered, and a summary of findings and fixes.
Q18Which attack classes do you test for: prompt injection, jailbreaks, data leakage, model theft? Coverage mapped to a public list such as the OWASP Top 10 for LLM Applications, with the gaps stated.
Q19Do you test again after material changes, and how often otherwise? A cadence, the change triggers that force a retest, and the last two test dates as proof.
Q20Will you share test evidence, or let us test the integration ourselves? A report under NDA, a summary letter from the tester, or contract terms that allow customer testing outside production.

Maps to: NIST AI RMF MANAGE 3.1

Runtime controls and incident response

What watches the AI while it runs, and what happens when it goes wrong.

4 questions
QuestionWhat a good answer includes
Q21What controls inspect prompts and responses at runtime, and what do they block or change? The control in place, the categories it detects such as personal data, secrets and jailbreak attempts, and whether it blocks, modifies or only logs.
Q22What do you log about AI interactions, and what can we access? Log fields, retention, export options, and whether we can receive the logs for our own tenant.
Q23What is your process for an AI-specific incident, and how fast will you notify us? A written runbook covering model misuse, data exposure through outputs and provider incidents, plus a notification window in the contract.
Q24Have you had an AI-related security incident in the last 24 months, including at a model provider? A direct answer, a summary of impact and fixes, and how affected customers were told.

Maps to: EU AI Act Article 26(5) and (6) · NIST AI RMF GOVERN 6.2

Compliance frameworks and evidence

Which rules the system is mapped to, and the paper that shows it.

4 questions
QuestionWhat a good answer includes
Q25Which AI frameworks is the system mapped to: EU AI Act, NIST AI RMF, ISO/IEC 42001? The mapping document or control matrix itself, with its date and who reviewed it.
Q26Under the EU AI Act, what role do you take for this system, and how do you classify it? Provider or deployer status, the risk classification with reasoning, and for high-risk systems the instructions for use you will supply.
Q27Which independent audits or certifications cover the AI system itself, not only the company? Scope statements that name the AI service, the period covered and the exclusions.
Q28What documentation will you give us for our own records? A list of artifacts such as system cards, test summaries and data sheets, their format, and how often each is refreshed.

Maps to: EU AI Act Article 25(4) · Article 26(1) · ISO/IEC 42001

Human oversight

What the AI can do alone, and how a person steps in.

4 questions
QuestionWhat a good answer includes
Q29Which decisions or actions can the AI take without a person approving them? A list of automated actions, their limits, and which ones need human review before they take effect.
Q30How can our staff review, override or stop an AI output or action? The control in the product, who can use it, and how quickly a stop takes effect.
Q31What information do you give users so they can interpret AI outputs? Instructions for use, known limitations and failure modes, and how users know they are dealing with AI.
Q32How do you measure and report accuracy, error rates or drift once the system is live? The metrics tracked, the thresholds that trigger action, and a recent report.

Maps to: EU AI Act Article 26(1) and (2)

04 Framework map

Which frameworks back these questions?

You are not asking out of curiosity. Three widely used frameworks expect buyers to manage third-party AI, and each group traces to at least one of them. For the frameworks themselves, see AI compliance frameworks, compared. This is a practitioner’s summary, not legal advice.

ProvisionWhat the text saysGroups
EU AI Act, Article 25(4)For high-risk AI systems, the provider and any third party that supplies an AI system, model, tools, services, components or processes must set out in a written agreement the information, capabilities, technical access and assistance the provider needs to comply.1, 3, 7
EU AI Act, Article 26(1) and (2)Deployers of high-risk AI systems must use them in accordance with the instructions for use, and assign human oversight to people with the necessary competence, training and authority.7, 8
EU AI Act, Article 26(5) and (6)Deployers must monitor the system in operation, inform the provider and authorities of risks and serious incidents, and keep the logs under their control for at least six months.6
NIST AI RMF, GOVERN 6.1“Policies and procedures are in place that address AI risks associated with third-party entities, including risks of infringement of a third-party’s intellectual property or other rights.”1, 2, 3
NIST AI RMF, GOVERN 6.2“Contingency processes are in place to handle failures or incidents in third-party data or AI systems deemed to be high-risk.”6
NIST AI RMF, MANAGE 3.1 and 3.2Risks and benefits from third-party resources are regularly monitored, with risk controls applied and documented, and pre-trained models are monitored as part of regular maintenance.1, 4, 5
ISO/IEC 42001:2023A management system standard for organizations that provide or use AI. NIST’s crosswalk to the final draft of the standard maps GOVERN 6.1 and 6.2 to its guidance on allocating responsibilities and on suppliers.3, 7
05 The limits

Where does a questionnaire stop?

Use the template knowing what it cannot do. Three limits apply to every questionnaire, including this one, and each has a fix that is not another question.

01

It is self-reported

An answer is what the vendor believes or chooses to say. Without an artifact behind it, such as a bill of materials or a test report, you are filing an assurance.

Why questionnaires fall short
02

It is a point in time

A model, provider or data use can change the week after the form comes back. The answer on file then describes a system that no longer exists.

See the due diligence checklist
03

It stops at the vendor’s door

No questionnaire and no customer-side tool can inspect a model running inside a vendor’s infrastructure. Disclosure is the only way in, so contract for it.

How evidence requests work
06 Where Cranium AI fits

Where Cranium AI fits

The template tells you what to ask. Cranium AI is one way to collect the answers as evidence, through the five moves of the AI Trust Loop. It does not replace the TPRM suite that sends your questionnaires. An AI Card request can repeat monthly, quarterly or on a custom schedule, so the vendor is asked for an updated card each cycle. The third-party AI risk hub walks through the workflow.

  1. Discover

    Decide who gets the questionnaire. The Shadow AI view shows AI-enabled SaaS in your connected SIEM activity, a starting list of vendors to ask.

  2. Observe

    Watch for the changes group 4 asks about. Put the AI Card request on a refresh schedule, and re-request sooner when one of those changes lands.

  3. Govern

    Map what comes back to the EU AI Act, NIST AI RMF and ISO 42001, the same frameworks this template points to.

  4. Secure

    Test the AI applications you build on vendor models with Arena, so group 5 is not answered by the vendor alone.

  5. Prove

    Request an AI Card with the frameworks, bill of materials and documentation you need, and a due date. Accept it, request changes or decline it, then keep it in My Vendors.

07 Keep going

Related reading.

08 Questions

Frequently asked questions.

Straight answers, kept current. Last updated .

Ask us anything →
What is a vendor AI risk questionnaire?

It is a structured set of questions a buyer sends to a supplier to learn which AI models touch its data, how those models are tested and controlled, and what evidence backs each answer. It extends a security questionnaire to cover models, training, AI subprocessors, change notice and human oversight.

What questions should you ask an AI vendor?

Ask which models and providers process your data, whether your data trains any model, which AI subprocessors are involved, how you will be told before the AI changes, what adversarial testing it has passed, which controls run at runtime, which frameworks it maps to and what a person can override. This template turns those eight areas into 32 questions.

Is this AI vendor questionnaire template free?

Yes. All 32 questions are on this page with no form to fill in. Copy them with the button above the first group, paste them into your own assessment tool or spreadsheet, and edit them to fit your risk tiers. It is a starting template, not legal advice, so have counsel review anything you turn into contract terms.

How is an AI vendor questionnaire different from a security questionnaire?

A security questionnaire asks how a vendor protects systems and data in general. An AI questionnaire asks what the vendor’s AI does with your data: which models see it, whether it trains them, how the models are tested, what can change without notice and what a person can override. Most security questionnaires and audit reports were not written to answer those points.

Which frameworks support asking vendors these questions?

The NIST AI Risk Management Framework calls for policies and procedures that address AI risks from third-party entities, in GOVERN 6. The EU AI Act requires written agreements between providers of high-risk AI systems and their suppliers in Article 25, and sets duties for deployers in Article 26. ISO/IEC 42001 is a management system standard for organizations that provide or use AI.

How often should you send a vendor AI questionnaire?

Send it before you sign, then again when something changes: a new model or provider, a new AI subprocessor, a new use of your data or a new AI feature. Treat an annual refresh as the backstop and the change as the trigger, and put the vendor’s duty to tell you about changes into the contract.

Where does a vendor questionnaire stop working?

A questionnaire is self-reported and describes one moment. It cannot verify an answer, notice a change after it is returned or see inside a vendor’s infrastructure. Close the gap with evidence requests such as an AI bill of materials and test reports, contract terms that require notice of change, and one place to keep what comes back. Cranium AI collects that evidence through AI Cards.

Sources

Primary sources for the figures and frameworks cited on this page. Last verified .

  1. Bank of England and FCA, 2026
  2. Ncontracts, 2026 TPRM survey Vendor survey
09 See it for yourself

Ask better questions.
Collect real evidence.

See how Cranium AI Cards turn a vendor’s answers into evidence you can review, keep and prove.