Practical guide

What evidence shows vendor risk management is operating?

A vendor-risk policy or completed questionnaire shows design intent, not ongoing operation. Operating evidence should demonstrate that the organisation knows which suppliers and AI components matter, performs proportionate due diligence, translates findings into contractual and technical controls, monitors change, manages exceptions, and can respond when a dependency fails.

By Two Prune

A vendor-risk policy or completed questionnaire shows design intent, not ongoing operation. Operating evidence should demonstrate that the organisation knows which suppliers and AI components matter, performs proportionate due diligence, translates findings into contractual and technical controls, monitors change, manages exceptions, and can respond when a dependency fails.

Use authoritative guidance without overstating it

NIST Special Publication 1326, finalised in July 2026, describes supplier due diligence across ownership and control, provenance, resilience, foundational cyber practices, and supply-chain tiers. NIST's Generative AI Profile recommends use-case-based supplier assessment, inventories of third parties, contract evaluation rights, ongoing monitoring, provenance records, and contingency processes for third-party failures. NIST SP 800-161 Revision 1 provides broader cybersecurity supply-chain risk-management practices across the system lifecycle. These sources support an evidence model; they do not establish legal compliance or prove that a particular vendor is safe. NIST SP 1326: Due Diligence Assessment Quick-Start Guide NIST AI 600-1: Generative AI Profile NIST SP 800-161 Rev. 1 update 1

Sources: [1] [2] [3]

Test operation with TRACE

Use TRACE: tier the dependency, retain due diligence, align obligations and controls, check operation and change, and evidence exceptions and exit. The review follows one supplier from intake through use and renewal instead of counting completed forms.

### T: Tier the dependency and use case

Retain a current inventory entry with the legal entity, service, product components, subcontractors, data accessed, integrations, users, business process, regions, deployment model, and accountable owner. Record whether AI functions are embedded, optional, or introduced through another service.

Assign a review tier using explicit consequence factors: data sensitivity, operational criticality, autonomy, decision influence, substitutability, concentration, regulatory context, and recovery time. The tier should determine the evidence depth, approval route, monitoring cadence, and contingency requirement.

Evidence of operation includes a dated tier decision tied to the actual use case. A vendor's general security page cannot substitute for the organisation's dependency analysis.

### R: Retain due diligence and provenance

Keep the questions asked, sources consulted, evidence received, reviewer, date, findings, limitations, and unresolved claims. For significant dependencies, cover ownership or control, product and component provenance, resilience, foundational security practices, supply-chain tiers, incident history, and financial or operational continuity as appropriate.

Distinguish vendor attestations, independent assurance, public records, technical tests, and internal observations. Record assurance scope and period. An expired report or certificate for a different product is not current evidence for the service in use.

For AI services, record model and provider dependencies, data handling, retention, training use, content provenance, evaluation information, change notices, and known limitations where applicable. Missing information should remain visible as missing rather than being scored as satisfactory.

### A: Align obligations and controls

Map material findings to contractual, technical, procedural, and governance controls. Retain the relevant clause or configuration, accountable owner, test method, and consequence of failure.

Examples include incident notification, audit or evaluation rights, data-use restrictions, deletion, access control, logging, vulnerability handling, subcontractor notice, service levels, model-change notice, continuity, export, termination, and transition assistance. Verify that promised controls are implemented in the configured service, not merely present in a standard contract.

Risk acceptance must identify the decision-maker, rationale, compensating measures, expiry date, and trigger for reconsideration.

### C: Check operation and change

Sample evidence from the review period: access reviews, configuration checks, incident notices, service reports, vulnerability responses, model or subcontractor changes, data-flow reconciliations, renewal checks, and contingency exercises. Compare them with the contract and control map.

Test whether alerts reach an owner, findings are triaged, and remediation closes with retained evidence. Monitor according to consequence and change, not an arbitrary questionnaire anniversary alone.

The operating test should include a negative or changed case. A control that has never encountered a relevant event may be implemented, but its response path remains less evidenced.

### E: Evidence exceptions, contingency, and exit

Maintain an exception register with the affected use, risk, interim protection, owner, approval, due date, status, and expiry. Reopen the review when a vendor, model, subprocesser, use case, data type, incident, or legal context changes materially.

Retain a practicable contingency: service degradation steps, manual fallback, alternative supplier or internal route, data export, credentials and integration removal, communications, and recovery ownership. Test critical parts such as export or failover rather than assuming contract language is executable.

At renewal or exit, reconcile data deletion, access revocation, retained records, unresolved incidents, and transition completion.

Operating-evidence register

| Evidence ID | Supplier/use | Risk tier | Control objective | Period sampled | Evidence retained | Exception | Owner | Result | Next review | | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | | | | | | | | | | | |

Use four results: operating as scoped, operating with limitation, not evidenced, or failed. Do not convert “not evidenced” to “failed,” and do not average a critical failed control into a reassuring percentage.

Worked example: synthetic customer-support AI vendor

A fictional company uses an AI service to draft customer-support responses. The service receives ticket text containing account identifiers, connects through a helpdesk integration, and relies on a separately named model provider. The initial inventory lists only the reseller and marks the service “low risk” because a human sends every reply.

TRACE retiering finds that the service processes confidential support data, affects customer communications, and creates concentration risk because there is no tested alternative. It is moved to a higher review tier. The due-diligence file retains the vendor questionnaire, a scoped assurance report, subprocesser list, data-flow diagram, incident page, public ownership research, and an internal configuration test. The assurance report is current but excludes the newly introduced model feature, so that limitation remains open.

The contract requires 48-hour incident notification and advance subprocesser notice. The control map identifies ticket-field minimisation, single sign-on, restricted connector permissions, logging, monthly access review, model-feature approval, and quarterly vendor review. Sampling finds the access reviews and permissions operating, but the data-minimisation rule is only documented; five sampled tickets contain a field the approved configuration was supposed to remove.

The vendor also added a model provider 12 days earlier. The notice reached a dormant procurement mailbox and was not reviewed. This is evidence that the notification clause exists but the internal response path failed. The owner opens two exceptions: repair the redaction configuration within seven days and route notices to the vendor-risk queue with a named deputy. Until then, the affected ticket category is disabled.

The contingency exercise exports prompt templates and integration settings but discovers that conversation history cannot be exported in the assumed format. The exit plan is updated, the recovery objective is qualified, and a manual-response fallback is tested for one queue.

The final disposition is operating with limitations. Inventory, access, review, and containment evidence exists, but change-notice handling, redaction, and export remain under dated remediation. The organisation schedules a targeted retest rather than claiming that the vendor is safe or that the overall control is effective without qualification.

Two Prune boundary

Two Prune can assess whether a participant identifies supplier dependencies, navigates due-diligence evidence, tests controls, handles exceptions, and communicates a bounded disposition in a realistic simulation. It reports observed assessment evidence. It does not certify vendors, perform continuous monitoring, provide legal or compliance assurance, establish control effectiveness, forecast workplace outcomes, or make employment decisions.

Sources

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

  • NIST AI 600-1: Generative Artificial Intelligence Profile
  • NIST SP 800-161 Revision 1 Update 1

Sources: [1] [2] [3]

Sources

  1. 1.NIST SP 1326: Due Diligence Assessment Quick-Start Guide · National Institute of Standards and Technology
  2. 2.NIST AI 600-1: Generative AI Profile · National Institute of Standards and Technology
  3. 3.NIST SP 800-161 Rev. 1 update 1 · National Institute of Standards and Technology