Practical guide

How to operationalize vendor risk in AI workflows

A vendor-risk workflow turns an approved policy into repeatable handoffs, evidence, decisions, and change triggers. It should start before a trial or purchase, follow the configured service through operation, and end only after access, data, integrations, and obligations are reconciled. A questionnaire stored at onboarding is not an operating workflow. Use the FLOWER runbook: front door, level.

By Two Prune

A vendor-risk workflow turns an approved policy into repeatable handoffs, evidence, decisions, and change triggers. It should start before a trial or purchase, follow the configured service through operation, and end only after access, data, integrations, and obligations are reconciled. A questionnaire stored at onboarding is not an operating workflow. Use the FLOWER runbook: front door, level the use, obtain evidence, write the decision, execute controls, and reassess or retire. The procedure complements policy and control-testing records; it does not replace them.

F — Front door every proposed use

Route a new service, embedded AI feature, API, model, agent, plugin, renewal, or material expansion through one intake. The requester records:

The intake owner checks the service inventory for duplicates and existing decisions. A previous approval is reusable only when the use, data, configuration, and material supplier conditions remain within its boundary.

  • supplier and service;
  • business owner and decision required;
  • intended users and workflow;
  • data sources, classifications, destinations, retention, and deletion need;
  • integrations, permissions, tools, and external actions;
  • regions, subprocessors, and contractual route;
  • desired trial or operating date.

L — Level the actual use case

Assign a review route based on the proposed use, not the vendor's marketing category. A drafting tool using public material may follow a standard route. Confidential data, persistent memory, personal information, tool execution, external communication, or a material decision requires additional review and approval.

Record the tier, trigger, rationale, reviewers, and due date. If information is missing, select the more cautious route until evidence supports a change. The tier determines diligence depth; it does not declare the service safe.

NIST SP 1326 describes supplier due diligence as research into pertinent information for informed acquisition and system decisions. Its factors cover ownership or control, provenance, resilience, foundational cyber practices, and supply-chain tiers for ICT suppliers. An organisation should adapt the evidence request to its context.

O — Obtain and reconcile evidence

Create a request register with an owner and disposition for every item. Depending on the tier, request evidence about:

Reconcile vendor claims with contracts, technical configuration, independent reports, public records, and internal tests. Mark evidence as current, stale, partial, contradictory, or missing. An unanswered question remains a finding; it is not silently converted into acceptance.

  • service architecture, security, vulnerability handling, incidents, continuity, and recovery;
  • data locations, subprocessors, training use, retention, deletion, encryption, and access;
  • model or AI functionality, limitations, evaluations, change notices, and logging;
  • accessibility, support, export, portability, and termination;
  • contract terms, confidentiality, audit or evidence rights, incident notice, material change, and exit.

W — Write the decision and conditions

The decision record names:

Procurement, security, privacy, legal, accessibility, records, and operational reviewers decide matters within their mandates. Approval is effective only after required conditions are implemented and verified. NIST SP 800-161r1-upd1 frames supply-chain risk management as a lifecycle discipline; the decision must therefore connect to operation and change, not end at contract signature.

  • approved or declined use and configuration;
  • accountable service owner;
  • required controls and evidence anchors;
  • prohibited data, integrations, or actions;
  • residual risks and accepting authority;
  • exceptions with expiry and compensating measures;
  • review date and material-change triggers;
  • incident, suspension, and exit routes.

E — Execute controls and acceptance tests

The service owner coordinates configuration. Common tasks include managed accounts, least privilege, permitted data classes, retention settings, blocked training use, restricted tools, human review, logging, user instructions, and escalation contacts.

Run representative acceptance tests before operation. Test expected use and failure conditions such as unauthorised access, a prohibited data input, an unapproved external action, missing source attribution, or loss of a required log. For each test, record setup, expected result, observed result, evidence, tester, date, and disposition.

A failed critical test pauses the affected use. Repair and retest must create new evidence rather than overwriting the failure record.

R — Reassess change or retire the service

Feed vendor notices, incidents, access reviews, user reports, monitoring, contract milestones, and inventory reconciliation into the workflow. Reassessment triggers can include:

At retirement, disable accounts and integrations, revoke keys and tokens, export required records, complete the available deletion process, update inventories, and record unresolved dependencies. The owner closes the workflow only when every action has evidence or an authorised residual item.

  • a new model family, AI feature, agent capability, or connector;
  • a new data class, region, subprocessor, purpose, or population;
  • changed retention, training use, terms, ownership, or control;
  • a material incident, failed test, or repeated exception;
  • renewal, suspension, or reactivation.

Workflow evidence artifact

| Stage | Required evidence | Accountable role | Stop condition | |---|---|---|---| | Front door | Complete intake and inventory match | Requester / service owner | Unknown owner or use | | Level | Tier, rationale, reviewers, due date | Risk triage owner | Material trigger not routed | | Obtain | Request register and dispositions | Review coordinator | Critical evidence missing | | Write | Decision, conditions, exceptions, triggers | Approval authority | Residual risk unowned | | Execute | Configuration and acceptance-test record | Service owner | Critical test failure | | Reassess / retire | Change, review, incident, or exit record | Service owner | Access or dependency unresolved |

Worked example: meeting assistant adds external actions

A synthetic organisation has approved a meeting assistant for internal summaries using managed accounts and a 30-day retention setting. The vendor releases a feature that can draft and send follow-up emails.

The workflow handles the change:

The record shows a bounded feature decision and an unresolved condition. It does not extend approval to every function of the platform or establish that the vendor's controls operate in all environments.

  • Front door: the owner submits a material-change intake rather than enabling the feature from the admin console.
  • Level: external communication and write capability move the proposed feature to an enhanced route. The existing summary approval remains valid only for its original configuration.
  • Obtain: reviewers request tool-permission documentation, per-action authorisation behaviour, logs, recipient controls, updated subprocessors, and contract-change details.
  • Write: the authority permits testing in a synthetic tenant, prohibits production email access, and requires an exception to expire after 30 days.
  • Execute: a test shows the assistant can prepare a draft but also retains an unapproved recipient from an earlier synthetic meeting. The recipient-isolation criterion fails, so production activation is paused.
  • Reassess: the supplier ships a repair. A new test passes recipient isolation and per-action confirmation, but production remains restricted until logging and incident-routing evidence is complete.

Verification and Two Prune boundary

Workflow verification reconciles the inventory, intake, tier, evidence register, decision, configured controls, tests, change triggers, exceptions, and exit actions. Missing, failed, stale, and contradictory evidence should remain visible. Periodic sampling can check that actual uses still match approved records, while a separate control-evidence process tests whether named controls operate over time.

This workflow is a governance template, not legal advice, a jurisdiction-specific compliance determination, or supplier certification. Two Prune can simulate vendor-risk analysis and capture evidence of a participant's research, triage, verification, judgment, and communication. It does not procure or approve vendors, operate customer controls, 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]