Tools · Procurement · Third-party risk

Vendor risk assessment for AI suppliers

A reference for the teams who vet AI vendors. It sorts the AI supplier landscape into seven categories, maps each to the SRF layers and accountability boundaries it touches, and lists the attestations and evidence to demand before you sign.

Audience Vendor risk · Third-party risk management · Procurement · Security GRC · Legal
Practitioner aid. This page operationalizes the CoSAI Shared Responsibility Framework for vendor risk work. It is a proposed extension developed to demonstrate the approach, not part of CoSAI SRF v1.0 and not an endorsement of any named vendor. Validate every control and contract term with your security, legal, and compliance functions before use.

Vendor risk teams protect the company from supply-chain exposure. With AI suppliers that job gets harder, because risk no longer sits in one place: it spreads across the data a vendor touches, the model behind a feature, the autonomy you grant an agent, and the sub-processors none of those vendors named. A SOC 2 report alone will not tell you who is accountable when an agent takes the wrong action or a model leaks a prompt.

The fix is to know, per vendor type, which SRF responsibilities the vendor owns and which stay with you, then ask for evidence that proves the vendor holds up its side. Each category below gives you that split plus a concrete evidence list. Start with the risk tier to decide how deep to go.

Read the how-to guide Download VRA workpaper (.xlsx) SRF layer reference

Step 1

Tier the vendor before you assess it

Depth of assessment should track the risk a vendor introduces, not the size of the contract. Score each vendor on three axes: the sensitivity of data it touches, the autonomy it exercises, and how deep into the SRF stack it reaches. The highest of the three sets the tier.

Tier Triggers (any one) Assessment depth
Tier 1 · Critical Regulated or highly sensitive data; autonomous actions on your systems; deep platform or model access (L4 or L5). Full assessment, all evidence asks, contractual SRF responsibility matrix, continuous evidence rather than point-in-time, named incident path.
Tier 2 · Elevated Internal business data; human-in-the-loop; application-layer AI features (L3). Core attestations plus the AI-specific asks for the category, data-use and sub-processor disclosure, admin controls review.
Tier 3 · Standard Low-sensitivity data; bounded scope; no autonomy; no training on your data. Baseline attestation (SOC 2 or ISO 27001), acceptable use terms, data-retention and deletion commitment.

Step 2

Assess by vendor category

Seven categories cover the AI supplier market. Filter by the SRF layer you care about to narrow the list. Asks flagged AI-specific go beyond a standard security questionnaire and are the ones most VRA checklists miss.

SRF layer
Category 1

Core AI platform & infrastructure

Cloud AI services, MLOps pipelines, model hosting, GPU infrastructure. For example AWS AI, Azure AI, Google Cloud AI.

L4 AI Platform L5 (partial) IaaS · AI-PaaS AI Platform Provider
Vendor is accountable for

Tenant isolation, platform and infrastructure security, encryption services, availability, and the platform telemetry it exposes.

You stay accountable for

Workload and key configuration, data classification, identity and access, app-layer controls, and deciding which AI events and logs you collect.

Risk areaAttestation baselineEvidence to ask for
Multi-tenant data isolationSOC 2 / ISO 27001Architecture diagrams showing how one customer is prevented from reaching another's data, and where data resides.
Encryption at rest and in transitSOC 2 / ISO 27001Key-management description and bring-your-own-key options.
Security certificationsSOC 2 / ISO 27001 / FedRAMPCurrent attestation reports under NDA, with scope and exceptions.
Incident response and monitoringSOC 2 +Which AI events the platform generates, how models are monitored, and which logs you are expected to collect.
Shared responsibility boundaryDocumented matrixA responsibility matrix that states who owns what, plus the adversarial or bias testing the platform expects you to run.
Category 2

Pre-trained model & algorithm providers

Foundation and domain-specific models reached by API: LLMs, vision, and speech models. For example OpenAI, Anthropic, Mistral, Meta, Google.

L5 AI Model Provider AI-PaaS Model Provider
Vendor is accountable for

Training and model integrity, safety tuning, model and system cards, API security, and not training on your prompts where contracted.

You stay accountable for

Use-case risk classification, prompt and output handling, application guardrails, and human oversight of model output.

Risk areaAttestation baselineEvidence to ask for
Model integrity and benchmarksSOC 2 +Third-party benchmark certification and robustness metrics against adversarial attacks and jailbreaks.
Training-data provenance and biasISO 42001Proof of data opt-out (GDPR, CCPA) and a list of third-party datasets used.
Responsible AI and explainabilityISO 42001System and model cards covering intended use, limits, and evaluation.
Licensing and IP indemnificationContractLicense and acceptable-use terms, plus an IP indemnity clause for model output.
Regulatory alignmentEU AI ActMapping of the model to EU AI Act risk categories and proof that prompts are not used for training.
Model supply-chain integrity AI-specificISO 42001Model provenance and signing, weight-integrity controls, and a vulnerability disclosure process.
Category 3

AI-enabled SaaS applications Added

SaaS products with embedded AI features: copilots, assistants, and AI inside tools you already buy. For example CRM copilots, support bots, document AI. The category most companies actually procure.

L3 AI Application AI-SaaS Application Developer
Vendor is accountable for

Application logic, in-app guardrails, output filtering, feature-level access control, and the foundation models it selects underneath.

You stay accountable for

What data you feed the app, user permissions, acceptable use, and monitoring AI output for your own context and accuracy.

Risk areaAttestation baselineEvidence to ask for
Underlying model and sub-processorsDisclosureWhich foundation models and sub-processors power the feature, and the data flow to each.
Prompt injection and output safety AI-specificISO 42001Input and output filtering, indirect prompt-injection defenses, and jailbreak test results.
Data use for trainingContractContractual proof that tenant data is not used to train shared models.
Feature governanceSOC 2 +Admin controls to disable AI features and audit logs of AI actions taken in your tenant.
Tenant isolationSOC 2 / ISO 27001Isolation evidence and data-residency options.
Category 4

Agentic AI & autonomous systems Added

Autonomous agents, tool-using copilots, workflow automation, and the tool and connector ecosystems (for example MCP servers) that agents act through.

L3 AI Application L4 AI Platform Agent-PaaS Agentic Platform Provider
Vendor is accountable for

The agent runtime, tool and permission scoping, guardrail enforcement, action logging, and the autonomy and override controls it exposes.

You stay accountable for

Which tools and scopes you grant, how you configure human override and approval gates, and the blast-radius limits you set.

Risk areaAttestation baselineEvidence to ask for
Autonomy and human override AI-specificDisclosureThe autonomy levels (L0 to L5) and override tiers (T1 to T5) the product supports, and its default settings.
Tool and action authorization AI-specificISO 42001The permission model, scope limits, and the ability to require human approval per action class.
Guardrail bypass monitoringSOC 2 +Detection of guardrail bypass, runtime integrity verification, and a complete action audit trail.
Blast-radius containmentContractRate limits, spend caps, rollback, and a documented kill-switch.
Indirect prompt injection AI-specificISO 42001Defenses for the case where an agent reads untrusted content or tool output and acts on it.
Category 5

Data, labeling & retrieval services

Managed labeling, RLHF, and crowdsourcing platforms, plus retrieval and vector-data providers that feed RAG systems. For example Scale AI, LXT, Toloka.

L2 AI Information Data Provider
Vendor is accountable for

Workforce vetting, confidential-data controls, annotation quality, retrieval-index security, and secure deletion on request.

You stay accountable for

Classifying data before you share it, defining the retention policy, and accepting the quality thresholds you contract to.

Risk areaAttestation baselineEvidence to ask for
Workforce vetting and background checksSOC 2 +Which roles or regions are exempt from screening, and the re-screening cadence for gig workers and contractors.
Confidential-data controlsSOC 2 +DLP evidence, GeoIP and embargo restrictions, and penetration-test results for any contractor-provided VDI.
Annotation QA and accuracyISO 9001 +QA methodology documentation and the prior-quarter metrics dashboard.
Sub-processors and cross-border flowDisclosureSub-processor list with change control, a data-flow diagram with geographies, audit results, and opt-out options.
Retention and secure deletionSOC 2 +Whether you can define the retention policy, plus certificates of data deletion or destruction.
Retrieval and RAG integrity AI-specificISO 27001Access controls on the vector store, data-poisoning defenses, and tenant isolation of embeddings.
Category 6

AI consulting & professional services

Strategy, custom model development, and system integration. For example Big-4 consultancies and boutique AI shops.

L1 AI Business & Usage Cross-layer AI System Governance Application Developer
Vendor is accountable for

Secure development practice, deliverable quality, the firm's own security posture, and clean exit from your environment.

You stay accountable for

Scope definition, environment-access boundaries, IP terms, and acceptance criteria for what the engagement produces.

Risk areaAttestation baselineEvidence to ask for
Scope, SLAs, and success KPIsContractExamples of AI KPI dashboards and prior-project scorecards.
Access controls in client environmentsSOC 2 +Restricted data and system-access requirements such as VPN, VDI, Jupyter Hub, and IDE.
Ownership of code and models producedContractContracts addressing IP or joint ownership, plus SBOM and AIBOM breakdowns of what was delivered.
Security posture of the firmSOC 2 / ISO 27001Third-party attestation results for the consulting firm itself.
Post-engagement data destructionSOC 2 +Destruction certificates, the process for sanitizing IDE and repos (Jira, Confluence, Git), and wiping developer workstations and VDI caches.
Category 7

Software, VARs & distributors

Bundled AI software, marketplace sellers, and OEM integrators that resell or repackage AI built by someone else.

L3 AI Application L4 AI Platform Reseller of OEM
Vendor is accountable for

Passing through the OEM's controls without weakening them, delivering patches, and providing support and escalation.

You stay accountable for

Validating the OEM behind the reseller and holding both to the contract terms and the downstream supply chain.

Risk areaAttestation baselineEvidence to ask for
Patch and update responsibilityContractModel, dataset, and tool patch schedule, plus an end-of-life matrix by product.
Reseller controls aligned to OEMDisclosureA controls-hardening benchmark and a contract clause ensuring the vendor cannot weaken security settings over time.
Support and escalation pathContractNamed contacts and numbers, and the incident-response process.
Downstream supply-chain vettingDisclosureCurrent sub-provider inventory with geographic locations and data roles, audit results, and opt-out options.
Licensing terms and warrantyLegal reviewLicensing and warranty coverage, with legal sign-off.

Step 3

Cross-cutting AI risks to check on every vendor

These risks do not belong to one category. They apply to any vendor that builds, hosts, resells, or services AI, and they are where a traditional third-party questionnaire goes quiet.

Prompt injection, direct and indirect

Whether the vendor defends against malicious instructions in user input and in content the system retrieves or a tool returns.

Model and weight supply chain

Provenance, signing, and poisoning or backdoor defenses for models and weights. Ask for an AIBOM alongside the SBOM.

Training-data provenance and IP

What the model was trained on, opt-out handling, and indemnity for IP claims arising from output.

Data residency and sub-processors

Where data and inference run, the full sub-processor chain, and notice and opt-out on changes.

Continuous vs point-in-time evidence

Whether the vendor can show controls operating over time, not just a once-a-year attestation snapshot.

Versioning and decommissioning

How model and feature versions are tracked, notified, deprecated, and rolled back without breaking your controls.

Liability for AI-generated harm

Contractual allocation of liability when model output causes financial, legal, or safety harm.

AI management system maturity

Whether the vendor runs a governed AI program. ISO 42001 certification and a NIST AI RMF profile are the signals.

Reference

What each attestation actually proves

Attestations are not interchangeable. Security certifications say nothing about AI governance, and an AI-management certification says nothing about quality control on a labeling floor. Match the attestation to the risk.

AttestationWhat it coversWhat it does not cover
SOC 2 Type IIOperating effectiveness of security, availability, and confidentiality controls over a period.AI-specific risk, model behavior, bias, or autonomy.
ISO/IEC 27001A certified information security management system.AI governance and model-lifecycle controls.
ISO/IEC 42001A certified AI management system: governance, risk, and lifecycle controls for AI. The AI-specific signal.Point-in-time security control testing of infrastructure.
ISO 9001Quality management. Relevant to labeling and annotation accuracy.Security or AI-specific risk.
FedRAMPUS federal authorization for cloud services at a defined impact level.AI governance beyond the underlying cloud platform.
NIST AI RMFA voluntary AI risk-management profile. Shows the vendor has a structured AI risk process.Independent certification; it is self-applied unless audited.
EU AI ActRisk categorization and the legal obligations that follow for in-scope systems.A security or quality attestation on its own.

Templates

Contract clauses and RFP language

Copy-ready starting points that move the assessment from a spreadsheet into the agreement and the RFP. Edit the bracketed terms to fit, and have counsel review before use. These are drafting aids, not legal advice.

Contract clauses

Shared-responsibility matrix
Vendor shall maintain and attach as an exhibit a responsibility matrix that assigns one accountable party to each control across the applicable SRF layers (L1 to L5). Where a control is jointly operated, the matrix shall separate the portion Vendor controls from the portion [Customer] controls and assign each to a single owner. Vendor shall not characterize any control as "shared" without this split.
Data use and no training on customer data
Vendor shall not use [Customer] data, prompts, inputs, or outputs to train, fine-tune, or evaluate any model that is made available to other customers, and shall not retain such data beyond the period required to deliver the service. Vendor shall provide written confirmation of this control and evidence of technical enforcement on request.
Sub-processors, residency, and change notice
Vendor shall maintain a current list of all sub-processors, including any model, infrastructure, and labeling providers, with each provider's geographic location and data role. Vendor shall process [Customer] data only in [permitted regions] and shall give [Customer] at least [30] days' written notice before adding or replacing a sub-processor, changing a processing location, or changing the model that powers the service, with a right to object.
AI liability and IP indemnification
Vendor shall indemnify [Customer] against third-party claims that the service's AI output infringes intellectual property rights. The parties shall allocate liability for losses arising from AI-generated output, including erroneous, harmful, or non-compliant output, as set out in [Schedule X], and such liability shall not be excluded by general limitations on consequential damages to the extent it arises from Vendor's failure to meet the controls in this agreement.
Agentic controls and kill-switch
For any autonomous or agentic capability, Vendor shall provide [Customer] with configurable controls over autonomy level and human override, scope and rate limits on tool and action use, the ability to require human approval for [defined high-impact actions], a complete audit trail of agent actions, and a documented mechanism to immediately suspend agent activity. Vendor shall not change default autonomy or override settings without prior written notice.
Continuous evidence, re-assessment, and secure deletion
Vendor shall provide evidence that the controls in this agreement are operating, on a [continuous / quarterly] basis rather than only at point-in-time, and shall permit [Customer] to re-assess on any material change to the model, sub-processors, or autonomy configuration. On termination, Vendor shall securely delete all [Customer] data and provide a certificate of deletion within [30] days.

RFP question language

AI-specific RFP questions
1. Which SRF layers does your offering touch (L1 to L5), and what is your operating model (AI-SaaS, AI-PaaS, Agent-PaaS, IaaS)? 2. Provide a responsibility matrix stating which controls you own and which remain with us. 3. List the foundation models and sub-processors behind the service, with their geographic locations and data roles. 4. Confirm in writing that our data, prompts, and outputs are not used to train shared models, and describe how this is enforced. 5. Provide your current attestations (SOC 2 Type II, ISO/IEC 27001, ISO/IEC 42001) and, for AI governance, your ISO 42001 certification or NIST AI RMF profile. 6. Describe your defenses against prompt injection, including indirect injection from retrieved content or tool output. 7. Provide model and weight provenance, signing, and an AI bill of materials (AIBOM). 8. For agentic features, state the autonomy levels and human-override tiers supported, the permission model, and the kill-switch mechanism. 9. Describe how you notify customers of model, sub-processor, and autonomy changes, and the notice period. 10. Provide your data-retention and secure-deletion policy, including certificates of destruction on exit.

The downloadable VRA workpaper includes these clauses, the RFP questions, and an attestation-to-evidence mapping as separate tabs.

Related on this site