AI procurement & third-party risk

Vendor AI risk

Vendor AI risk is the risk introduced when an organization relies on an external AI product, model, data source, or service. The review should answer one practical question: can this product be used for this purpose, with these data, permissions, and dependencies, at an acceptable level of risk?

A supplier questionnaire is a starting point, not a decision. Compare material claims with evidence for the service being purchased, test what you can in the intended workflow, assign control ownership, and define what would trigger a new review.

Start before the questionnaire

Scope the intended AI use.

AI vendor due diligence should be proportional to the use. A drafting assistant working from public material and an agent able to change customer records should not receive the same review.

Task

What will the product draft, recommend, rank, predict, retrieve, or execute?

People

Who relies on the output, and who is affected when it is wrong?

Information

Which records, prompts, files, logs, or derived stores may contain sensitive or regulated data?

Access

Which systems, connectors, tools, and permissions can the product reach?

Failure

Which outcomes are unacceptable, hard to reverse, or time-critical?

Owner

Who can approve the use, restrict it, suspend it, and authorize restart?

Choose a proportionate review

Let the consequences of failure determine the evidence you need.

These are illustrative review approaches, not regulatory risk tiers. Assess data sensitivity, decision impact, the product’s ability to act, reversibility, and operational dependence together. A low-cost product can still create high-impact risk.

Limited, reversible assistance

For a non-sensitive drafting aid, confirm permitted data, basic security, output-review controls, and how to disable access. Keep production data out until its use is authorized.

Sensitive or relied-upon output

For private records, customer-facing advice, or analyst recommendations, add representative error testing, data-flow verification, access checks, and a clear reviewer or escalation owner.

Consequential decisions or actions

For decisions affecting people, automated changes, or critical operations, require stronger validation, testing of adverse cases, human authority to intervene, incident response, and a credible fallback.

Third-party does not mean one party

Map the AI supplier’s dependencies.

The company on the contract may depend on a model provider, hosting platform, external data, APIs, or tools. Shared upstream dependencies matter for security, continuity, change control, and fallback planning. [1] Research on foundation-model ecosystems provides a structured basis for tracing the assets and relationships in these chains. [12]

  1. 01Your organizationUsers, data, permissions, workflowWhat you configure and permit
  2. 02Application supplierInterface, retrieval, integrationsWhat wraps the AI capability
  3. 03Model / service providerModel, hosted API, inference serviceWhat generates, scores, or predicts
  4. 04Supporting dependenciesHosting, data, libraries, external toolsWhat the service relies on

Financial-services application. The Financial Stability Board’s October 2026 sound practices explicitly address third-party AI dependencies, including performance, transparency, data quality, supply-chain concentration, and continuity (Sound Practice 12). This is guidance for institutions to consider, not an automatic legal requirement for every purchaser. [14] In diligence, ask whether two nominally separate vendors depend on the same model, cloud service, data feed, or operating region.

Due diligence that changes the decision

Request evidence that answers the decision.

Request evidence for the actual product, version, configuration, and intended use. Separate what the supplier says from what documentation, independent reports, and local testing can establish. Prioritize questions that could change approval.

Supplier requests and decision-useful checks for AI vendor due diligence
Decision question Request from the supplier Examine before approval
What exactly are we buying? Product and model identifiers, enabled AI features, intended uses, known limitations, architecture, and material upstream dependencies. Do the documents describe the service and settings you will actually receive? Record differences between the demonstration, trial, and proposed deployment.
Does it perform the task well enough? Evaluation methods, task definitions, datasets, sample sizes, error breakdowns, test dates, and access to a relevant trial. Do the evaluation conditions resemble your users, inputs, languages, and consequences? Look beyond a headline accuracy number to the errors that matter.
Where does our information go? Data flows covering prompts, files, outputs, logs, support access, retention, deletion, training, and product improvement. Do contractual commitments and technical configuration agree? Clarify backups and derived stores such as indexes or embeddings.
What can the product access or do? Connector permissions, tenant-isolation design, authorization controls, security testing, and mechanisms to restrict or disable actions. Can ordinary users retrieve restricted material or trigger unintended actions? Separate supplier controls from controls that depend on customer configuration.
How are changes and incidents handled? Version policy, change notices, incident contacts, support commitments, diagnostics, and treatment of upstream-provider changes. Can you detect a material change, assess its effect, and restrict use while it is reviewed? Confirm who receives notices and who can act on them.
Can we recover or leave? Export formats, deletion procedures, continuity arrangements, termination terms, and migration support. Can the business continue without the product? Test what can be exported, what must be rebuilt, and whether a fallback shares the same critical dependencies.

Documentation

Use each artifact for the claim it can support.

Model cards can provide intended-use and evaluation context; dataset documentation can describe data creation, composition, use, and maintenance. [2] [3] Neither automatically establishes that the surrounding application’s permissions, retrieval, integrations, or customer configuration work correctly. Research on transparency reporting also documents why disclosures must be assessed within their stated scope. [10]

Independent reports

Read the scope before relying on the badge.

For an audit report, certificate, or test summary, check the assessed entity, service, period, exclusions, exceptions, and customer responsibilities. Record which specific vendor claims the report supports and which remain open.

Need a record structure for the evidence itself? See Vendor AI Risk Evidence. For evaluating claim strength, see AI assurance evidence.

Evidence sufficiency

Four checks before relying on a supplier’s assurance

Complete documentation is not the same as evidence that supports your planned use. These questions help the reviewer decide what can reasonably be concluded.

Scope
Does it cover this product, version, feature, region, configuration, and assessment period?
Relevance
Did the test or review evaluate the outcome and failure modes that matter in our workflow?
Corroboration
Is the claim supported only by supplier self-report, or by test records, independent assessment, or customer observation?
Decision
Which claim is supported, what remains untested, who accepts the limitation, and what would change the decision?

A report can be genuine, current, and still irrelevant to a particular AI feature. For a record-by-record approach, use the Vendor AI Risk Evidence resource.

Decision point · incomplete information

What if the supplier cannot substantiate a claim?

Missing evidence is not proof that a product is unsafe. It does limit what a buyer can conclude. Identify the specific unanswered claim, its consequence for the proposed use, and whether independent or customer-side evidence can close the gap. [9]

  1. 01

    Clarify exactly what is covered.

    Ask which purchased plan, AI feature, provider, region, retention setting, and time period each assurance covers. A company-wide report may not address the service you intend to deploy.

  2. 02

    Test what you can observe.

    Run representative tasks, permission checks, negative cases, and failure tests. Retain the setup, expected outcomes, observed results, dates, and reviewer. Customer tests cannot prove how a supplier uses data internally.

  3. 03

    Close material gaps in the terms.

    Confirm data handling, evidence access, incident cooperation, change notice, and exit support in the applicable agreement. Assign an owner to verify that the obligations can be enforced. [7]

  4. 04

    Change the permitted use.

    If an important claim is still unsupported, limit the pilot to safe inputs and actions, defer the proposed deployment, or select another supplier. Record the missing evidence and the person authorized to decide.

Example · employee policy assistant

Decision: restricted pilot. The supplier says prompts are not used for training, but its answer does not confirm the terms for our purchased plan or upstream providers. The team may test with synthetic policy documents; customer and employee records remain excluded.

To expand: obtain applicable contract terms, verify the data flow and subprocessors, and record privacy and business-owner approval.

Worked review

Turn supplier claims into checks.

Example: an internal assistant answers employee questions from company policies. It has read-only access and returns source references. The decision is whether to permit a limited deployment and under what conditions.

Claim 01

“Answers are grounded in your documents.”

Check

Use clear questions, conflicting policy versions, missing answers, and misleading retrieved passages. Verify that citations actually support the answer and that uncertainty reaches a useful fallback.

Record

Test questions, expected outcomes, cited passages, errors, configuration, and policy-owner adjudication.

Claim 02

“Users only see content they may access.”

Check

Use test accounts with different permissions. Attempt restricted queries, change access, and repeat. Check answers, citations, previews, and conversation history for unintended disclosure.

Record

Access settings, observed results, permission-change behavior, and ownership of any mismatch.

Claim 03

“Your data is not used for training.”

Check

Trace the commitment through the purchased plan, enabled features, upstream providers, support access, retained logs, and relevant contract terms. Functional testing alone cannot establish internal data-use practices.

Record

Applicable commitments, settings, data flows, corroborating assurance, and unresolved retention or secondary-use questions.

Controls that survive procurement

Assign control owners before deployment.

A supplier may operate the model while your organization controls users, data, integrations, and decisions made from its output. Write down the boundary so necessary controls do not fall between teams.

Illustrative allocation of supplier, customer and shared AI controls
Control area Supplier Customer Shared / coordinated
Service & evidence Specified behavior, relevant documentation, supported configurations Approved use, acceptance criteria, local validation Resolving evidence gaps and material limitations
Data & access Service-side handling, isolation, support access Permitted data, users, connectors, permissions Investigating leakage, misuse, or access-control failures
Change & incidents Notices, diagnostics, supplier response Impact assessment, restriction, local incident response Coordinated investigation and corrective action
Continuity & exit Export, deletion, termination support Fallback process, migration readiness, access removal Transition, retained records, unresolved exceptions

Contract essentials: permitted data use and retention; evidence access; change notice; incident cooperation; export and deletion; termination and transition. Match these to the actual service and accountable owners. [5]

Regulated use: applicable requirements may add deployer or procurement duties. For example, EU AI Act Article 26 applies to relevant high-risk AI deployers, not every AI purchase. [11]

The assessment has to end somewhere

Decide, monitor, and preserve the record.

Separate evidence received from evidence sufficient. An unanswered question about confidential-data access can matter more than a long list of satisfactory questionnaire responses.

Decision options

Choose a disposition for the specific use under review. Include conditions and outstanding evidence in the record.

Proceed

Evidence is sufficient for the defined use, essential controls are in place, and an authorized owner accepts the remaining risk.

Record: Approved scope, supporting evidence, remaining risk, decision-maker, and next review.

Restrict

A narrower deployment or controlled pilot can be supported, but the broader use cannot yet be justified.

Record: Excluded users, data, actions, and integrations; restrictions and the evidence required for expansion.

Defer

One or more material questions remain open, and the organization cannot yet justify approval—even with proposed restrictions.

Record: The unresolved claim, evidence requested, accountable owner, deadline, and the decision it blocks.

Decline / exit

The product does not meet the business need, essential controls are unavailable, or the remaining risk is unacceptable.

Record: Decision rationale, alternative or fallback, data-handling obligations, and termination or transition responsibilities.

After approval

Reassess when the original evidence may no longer apply.

Track model and retrieval changes, new integrations or datasets, material incidents, revised supplier terms, and changes to the intended use or affected population. For each trigger, identify which approval assumptions and controls need to be checked again.

Keep the review date, responsible owner, evidence received, decision, and any temporary restrictions with the original approval record.

Continuity and exit

Verify that the organization can actually leave.

Test exports, access removal, fallback operation, and realistic transition effort before relying on an exit plan. A second supplier may share the same upstream model, cloud service, or data source.

Record the trigger for suspension or termination, the person authorized to act, and the evidence needed to confirm data return or deletion. Supply-chain risk management continues after onboarding. [6]

Continue the work

Move from supplier claims to traceable assurance evidence.

Frequently asked questions

Vendor AI risk: practical answers.

These answers summarize the review approach used on this page. The appropriate checks still depend on the intended use, the information involved, the supplier’s actual service, and applicable requirements.

What is vendor AI risk?

Vendor AI risk arises when an organization relies on an external AI product, model, dataset, or service. Key exposures include performance failures, inappropriate data use, unauthorized access or actions, changes beyond the buyer’s control, and interrupted service.

What evidence should an AI vendor provide?

Start with product and feature scope, intended-use limitations, evaluation results, data flows, security and access controls, contractual commitments, material-change procedures, and exit arrangements. Require stronger corroboration for claims that determine whether a consequential use can proceed. [5]

Does SOC 2 certify an AI system is accurate or safe?

No. A SOC 2 report addresses specified controls within its stated scope, period, and criteria. It does not by itself establish the AI system’s task accuracy, fairness, factuality, or suitability for your deployment. Examine the report boundary and test relevant functionality separately. [13]

Can we approve an AI supplier with incomplete documentation?

Possibly for a bounded, lower-risk use if the material risks can be managed. Preserve unresolved questions and the reasons for approval. Where the gap concerns an essential claim—such as permitted sensitive-data use or autonomous permissions—restrict or defer the affected deployment until the basis for reliance is adequate.

When should a vendor AI risk review be repeated?

Reassess after material changes to the AI model, features, integrations, data handling, permissions, incident history, contractual terms, or business use. Set triggers and ownership in advance so supplier updates translate into review and, when necessary, suspension or retesting. [7]

Sources & evidence notesPrimary guidance and peer-reviewed research used on this page

U.S. banking context · October 2026: The 2023 interagency third-party risk management guidance remains the reference cited here. The agencies proposed replacement guidance in September 2026; it was still proposed as of this review, with comments due November 16, 2026. The proposed text must not be described as a final supervisory requirement. [7] [8]

Evidence basis. The review questions, example, responsibility table, and outcomes synthesize the cited primary guidance and peer-reviewed research. They are tools for evaluating a specific supplier, not substitutes for verifying that supplier’s service.

  1. UK National Cyber Security Centre and international partners (2023). Guidelines for secure AI system development: Secure development.
  2. Mitchell, M., et al. (2019). Model Cards for Model Reporting. FAT* ’19, 220–229.
  3. Gebru, T., et al. (2021). Datasheets for Datasets. Communications of the ACM, 64(12), 86–92.
  4. Raji, I. D., Kumar, I. E., Horowitz, A., & Selbst, A. D. (2022). The Fallacy of AI Functionality. FAccT ’22, 959–972.
  5. NIST (2024). Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile, especially GOVERN 6.1–6.2.
  6. Boyens, J., et al. (2024). Cybersecurity Supply Chain Risk Management Practices for Systems and Organizations. NIST SP 800-161 Rev. 1, Update 1.
  7. Board of Governors of the Federal Reserve System, FDIC & OCC (2023). Interagency Guidance on Third-Party Relationships: Risk Management. Current U.S. banking guidance covering planning, due diligence, contracts, monitoring, termination, and risk-based tailoring.
  8. OCC, Federal Reserve, FDIC & NCUA (2026). Proposed Third-Party Risk Management Guidance, September 11, 2026. Proposed, not final. Would replace existing guidance and emphasize tailoring to reasonably assessed third-party risk.
  9. Raji, I. D., et al. (2020). Closing the AI Accountability Gap: Defining an End-to-End Framework for Internal Algorithmic Auditing. FAT* ’20, 33–44. Peer-reviewed research on lifecycle audit documentation and traceability.
  10. Bommasani, R., et al. (2024). Foundation Model Transparency Reports. AAAI/ACM AIES. Peer-reviewed work proposing systematic transparency reporting; it does not independently verify any specific supplier’s product claims.
  11. European Union (consolidated text viewed 2026). EU AI Act, Article 26: Obligations of deployers of high-risk AI systems. Applicable only where the law, system category, role, and relevant effective provisions require it.
  12. Bommasani, R., Soylu, D., Liao, T. I., Creel, K. A., & Liang, P. (2024). Ecosystem Graphs: Documenting the Foundation Model Supply Chain. Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society, 7(1), 196–209. Peer-reviewed research on assets and supply-chain dependencies.
  13. American Institute of Certified Public Accountants (AICPA). Trust Services Criteria for Security, Availability, Processing Integrity, Confidentiality, and Privacy (revised points of focus, 2022). Primary criteria underpinning SOC 2 examination scopes; not a functional certification of AI models.
  14. Financial Stability Board (15 October 2026). Sound Practices for Responsible Adoption of Artificial Intelligence (AI): Final report. Sound Practice 12 addresses third-party AI risk, supply-chain and concentration exposure, performance, and continuity. The FSB presents sound practices for financial institutions, not a universal statute.

Published by InfoSecured.ai. Editorial standards · Suggest a correction.