Practical guide

Vendor risk policy for AI-enabled work

An AI vendor policy should convert risk intent into mandatory decisions: which uses enter review, what evidence is required, who can accept residual risk, what must be monitored, and when access ends. A questionnaire alone is not a policy, and a favourable onboarding review does not show that controls continue to operate after a model, subprocessor, data flow, or product feature changes. The.

By Two Prune

An AI vendor policy should convert risk intent into mandatory decisions: which uses enter review, what evidence is required, who can accept residual risk, what must be monitored, and when access ends. A questionnaire alone is not a policy, and a favourable onboarding review does not show that controls continue to operate after a model, subprocessor, data flow, or product feature changes. The policy below uses GUARD: govern scope, use-case tiering, assess diligence, require controls, and detect change or retirement. Adapt it to the organisation's jurisdiction, contracts, architecture, information classification, and risk appetite. Current legal, privacy, security, procurement, and sector review may add requirements.

Policy objective and scope

Objective. Manage risks created when a third-party AI service receives organisational data, influences a material workflow, invokes tools, or produces output used in a decision or deliverable.

Scope. This policy applies before purchase, trial, integration, material expansion, renewal, or reactivation of an external AI-enabled service. It includes embedded AI features in existing software, APIs, hosted models, agents, plugins, evaluation services, and vendor-operated support involving organisational data.

NIST SP 800-161r1-upd1 treats supply-chain risk management as a lifecycle activity involving organisations, systems, and suppliers. NIST SP 1326 adds an ICT supplier due-diligence assessment approach. These sources inform the operating structure; they do not decide an organisation's contractual or legal obligations.

G — Govern roles and records

The service owner is accountable for the use case, business need, data, user population, and continued operation. Procurement owns the commercial process. Security, privacy, legal, records, accessibility, and risk functions review matters within their remit. An approval authority accepts residual risk or declines the use at the level defined by the organisation.

Each service must have one vendor-risk record containing:

The record is the decision trail. Vendor assertions should be attributed and dated rather than rewritten as organisational facts.

  • service, supplier, owner, purpose, users, integration, and renewal date;
  • data categories, sources, destinations, retention, and deletion route;
  • model or AI functions used and whether the vendor may train on submitted data;
  • tools, permissions, network paths, and downstream systems;
  • risk tier, evidence requests, findings, decisions, exceptions, and review dates;
  • material changes, incidents, access reviews, and retirement evidence.

U — Use-case tiering and triggers

Tier the actual use, not the vendor's brand. A low-impact public-information drafting aid is different from the same platform connected to confidential records or authorised to take actions.

| Trigger | Minimum disposition | |---|---| | Public or synthetic input; no system action | Standard review | | Internal non-public data or persistent memory | Enhanced data and access review | | Confidential, personal, regulated, or client data | Specialist review and explicit approval before use | | Tool execution, write access, automated action, or external communication | Architecture review, least privilege, test evidence, and accountable approval | | Material assessment, safety, financial, legal, or rights impact | Highest applicable governance route; do not rely on vendor review alone |

A new data category, integration, model family, autonomous action, region, subprocessor, or purpose is a reassessment trigger. Users may not expand the use while reassessment is pending unless an authorised exception explicitly permits it.

A — Assess supplier and service diligence

The review owner requests evidence proportionate to the tier. Evidence may include:

NIST's Generative AI Profile highlights risks including confabulation, data privacy, information security, harmful bias, human-AI configuration, and value-chain integration. The reviewer should select relevant risks for the use case rather than copying the entire catalogue into every record.

Unanswered questions are findings, not automatic passes. Record whether a gap causes non-approval, a compensating control, a time-bounded exception, or acceptance by the authorised owner.

  • corporate identity, ownership, service description, financial or continuity information;
  • security architecture, independent assurance reports, vulnerability handling, incident notification, and recovery arrangements;
  • data locations, subprocessors, retention, deletion, training use, encryption, and access controls;
  • model documentation, known limitations, evaluation scope, change notices, human-oversight features, and logging;
  • accessibility, support, export, termination assistance, and portability;
  • contract terms governing confidentiality, data use, audit or evidence rights, material change, incident duties, deletion, and exit.

R — Require controls before operation

Approved services must implement the controls named in the decision record. Depending on the tier, these can include:

An approval is conditional on the described configuration. A control described in a policy or contract is not treated as operating until the organisation has suitable implementation and test evidence.

  • single sign-on, managed accounts, least privilege, periodic access review, and prompt removal after role change;
  • approved data classes, input minimisation, redaction, retention settings, and blocked training use;
  • restricted tools and destinations, per-action authorisation, output validation, rate limits, and separation between retrieval and execution;
  • human review before consequential use or external release;
  • logging sufficient to investigate access, actions, changes, and incidents;
  • representative tests for expected use, foreseeable misuse, failure handling, and vendor change;
  • user instructions covering permitted use, source checking, escalation, and prohibited data or actions.

D — Detect change, handle exceptions, and retire

The service owner reviews the record at least on the scheduled date and when a trigger occurs. Monitoring should reconcile vendor notices, configuration, accounts, data flows, incidents, test results, exceptions, and actual use. Material findings go to the named approval authority.

An exception must state the requirement, rationale, affected use, owner, compensating measures, expiry, monitoring, and approval. It expires automatically unless renewed with current evidence. Repeated exceptions should prompt a policy or architecture decision rather than silent continuation.

At retirement, disable accounts and integrations, revoke keys and tokens, export required records, request deletion where applicable, verify the disposition available under the agreement, update inventories, and record unresolved dependencies. Renewal is a new decision when the service or use has materially changed.

Worked policy record: document assistant

A synthetic operations team proposes a hosted AI assistant that can search supplier contracts and draft renewal briefs. The first request describes it as “read-only,” but architecture review finds an optional email connector and persistent conversation memory.

The GUARD record produces this disposition:

Representative testing finds that a user outside the procurement group can see a document title but not its contents. The metadata exposure fails the acceptance criterion. Deployment pauses until access filtering is repaired and retested. The policy record does not say the service is broadly safe or that all controls are effective; it records a bounded decision about a defined configuration and use.

  • Govern: operations is service owner; security and privacy review; the commercial director is residual-risk approver.
  • Use-case tier: confidential contract text and persistent memory require enhanced review. The email connector would introduce external communication and write capability, so it is excluded from initial scope.
  • Assess: the supplier provides current security evidence and subprocessor information. Its deletion description does not explain backup timing, so the gap remains open and the contract owner requests a defined commitment.
  • Require: managed accounts, contract-level access groups, blocked training use, a 30-day conversation-retention setting, source links in drafts, human review before a brief is shared, and quarterly access reconciliation.
  • Detect: the owner records a six-month review, immediate reassessment for any connector or model change, an incident route, and an exit procedure.

Verification and claims boundary

Policy verification asks whether required decisions and evidence exist, whether configured controls match the approval, whether tests meet preset criteria, whether exceptions are current, and whether monitoring and retirement actions are recorded. Report failed, missing, and stale evidence separately.

This policy is a governance template, not legal advice, a jurisdiction-specific compliance determination, or an assurance report. Supplier documents and certifications can contribute evidence but do not by themselves establish the operation of the customer's configured controls.

Two Prune is an assessment product. It can simulate vendor-risk work and capture evidence of a participant's research, risk judgment, verification, and communication. It does not approve vendors, operate customer controls, certify suppliers, forecast later workplace performance, or make employment decisions.

Primary sources

NIST SP 1326, Cybersecurity Supply Chain Risk Management: Due Diligence Assessment Quick-Start Guide

  • NIST SP 800-161r1-upd1, Cybersecurity Supply Chain Risk Management Practices for Systems and Organizations
  • NIST AI 600-1, Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile

Sources: [1] [2] [3]