Public information siteInstitutional framework and operating programs continue to evolve

Practical template · Updated July 23, 2026

AI Governance Checklist

Turn AI policy into operating evidence. This checklist gives each use case a purpose, accountable owners, release gates, human control, monitoring, incident response, and a real stop path.

A checkbox is not evidence

Record the owner, proof, decision, and next review.

Use one copy for one bounded AI workflow. For every item, identify who owns the answer, what evidence supports it, who approved the residual risk, and when the control will be checked again.

01

Choose one use case

Evaluate a specific workflow or decision—not “AI” in general.

02

Attach evidence

Link each answer to a test, record, owner, approval, or operating signal.

03

Stop on material gaps

High-consequence unknowns are review gates, not boxes to mark complete.

04

Set the next review

Controls age as models, data, vendors, people, and conditions change.

Eight control areas · Thirty-two checks

Move from purpose to evidence, operation, and correction.

01

Purpose, scope, and risk

A bounded use case with a clear benefit, consequence level, and decision boundary.

Suggested ownerExecutive sponsor and business owner

Evidence to keepUse-case brief, affected-party map, risk tier, prohibited-use list, approval record

Owner / evidence / exceptions / next action
02

AI inventory and accountability

Every AI system and material feature has a named owner and a current record.

Suggested ownerProgram owner, product owner, and technology owner

Evidence to keepAI inventory, system card, owner register, dependency map, retirement record

Owner / evidence / exceptions / next action
03

Data, privacy, and records

Only necessary, permitted, protected, and fit-for-purpose data enters the workflow.

Suggested ownerData owner with privacy and records support

Evidence to keepData map, permission basis, retention schedule, quality report, privacy assessment

Owner / evidence / exceptions / next action
04

Security, access, and change control

The system operates inside explicit access, action, supplier, and release boundaries.

Suggested ownerSecurity owner and technical owner

Evidence to keepThreat model, access review, vendor review, test results, release and rollback record

Owner / evidence / exceptions / next action
05

Human oversight, transparency, and recourse

People can understand AI involvement, exercise judgment, and correct consequential outcomes.

Suggested ownerDomain owner and operations owner

Evidence to keepReview protocol, user notice, override test, escalation path, complaint and correction records

Owner / evidence / exceptions / next action
06

Testing and acceptance evidence

Predefined evidence shows where the system works, where it fails, and whether it may launch.

Suggested ownerTechnical owner with domain and independent reviewers

Evidence to keepAcceptance plan, test set, threshold results, red-team findings, signed launch decision

Owner / evidence / exceptions / next action
07

Monitoring and incident response

The organization can detect change, contain harm, investigate, recover, and learn.

Suggested ownerOperations owner and incident owner

Evidence to keepMonitoring plan, dashboards, review log, incident playbook, post-incident actions

Owner / evidence / exceptions / next action
08

People, training, and review cadence

People know the rules, limits, escalation route, and when the system must be reviewed again.

Suggested ownerProgram owner and workforce or training owner

Evidence to keepRole-based training, completion records, job aids, review calendar, governance minutes

Owner / evidence / exceptions / next action

Ten questions before real-world use

Do not launch on optimism where evidence is required.

A “no” does not always end the project. It means the gap needs a named owner, a documented decision, and a control proportionate to the consequence before the workflow reaches people.

  1. 01

    The use case, affected people, expected benefit, and unacceptable outcomes are documented.

    YesNoNot applicable
  2. 02

    The consequence level and allowed automation level have been approved.

    YesNoNot applicable
  3. 03

    Every business, domain, technical, review, escalation, and stop role has a named owner.

    YesNoNot applicable
  4. 04

    The system, model, provider, data sources, integrations, and downstream actions are inventoried.

    YesNoNot applicable
  5. 05

    Data permission, privacy, retention, access, and supplier handling have been reviewed.

    YesNoNot applicable
  6. 06

    Security threats, access limits, tool permissions, rollback, and vendor risk have been tested.

    YesNoNot applicable
  7. 07

    Human review happens before consequential action and the reviewer can genuinely disagree.

    YesNoNot applicable
  8. 08

    Acceptance thresholds were set in advance and the system passed realistic and adversarial tests.

    YesNoNot applicable
  9. 09

    Users have appropriate notice, support, correction, appeal, and incident-reporting paths.

    YesNoNot applicable
  10. 10

    Monitoring, reapproval, incident response, pause, retirement, and evidence-retention rules are active.

    YesNoNot applicable

Governance is an operating cycle

Reopen the decision when the system or its context changes.

WhenReviewRequired record
Before approval

Define purpose, risk, people affected, owners, data, suppliers, and prohibited uses.

Use-case brief and accountable approval

Before launch

Complete testing, security and privacy review, reviewer training, monitoring, rollback, and user communication.

Signed launch decision with unresolved limits

During operation

Review outcomes, overrides, complaints, incidents, drift, access, suppliers, cost, and control performance.

Dated operating review and action log

After material change

Reassess purpose, model, data, prompts, integrations, permissions, population, vendors, and consequence.

Reapproval, rollback, or retirement decision

Ground the checklist in recognized public guidance

Use this operating layer alongside—not instead of—formal requirements.

NIST AI Risk Management Framework

NIST organizes AI risk work through Govern, Map, Measure, and Manage. The checklist converts those functions into use-case owners, evidence, gates, and recurring decisions.

Review NIST AI RMF Review the NIST playbook

OECD AI Principles

The OECD principles address human rights and democratic values, transparency, robustness, safety, accountability, and inclusive benefit. They help teams test the broader purpose behind a control.

Review the OECD AI Principles

References indicate source alignment only. They do not imply affiliation, endorsement, certification, or approval by NIST, OECD, or any other organization.

AI governance checklist FAQ

Know what the checklist can—and cannot—prove.

What should an AI governance checklist include?

A useful checklist connects purpose and risk to named owners, an AI inventory, data and privacy rules, security boundaries, meaningful human oversight, acceptance testing, user transparency, monitoring, incident response, training, change control, and retirement. Each control should identify the evidence that proves it is operating.

Who should own AI governance?

Executive leadership should assign accountability, but no single department can operate the whole system. Business, domain, technical, security, privacy, legal or compliance, operations, and affected-user perspectives have different responsibilities. Every use case still needs one named owner with authority to pause or stop it.

How often should an AI governance checklist be reviewed?

Review before approval, before launch, on a defined operating cadence, after an incident, and after a material change to purpose, model, data, prompts, integrations, permissions, population, or supplier. Higher-consequence systems generally need more frequent review and tighter escalation thresholds.

Is an AI governance checklist the same as an AI audit?

No. A checklist helps a team design and operate controls. An audit independently evaluates whether required controls exist, work as intended, and have reliable evidence. A completed checkbox without evidence or independent challenge is not an audit result.

Can I download this AI governance checklist as a PDF?

Yes. Use the Print or save as PDF button and select your browser’s PDF destination. The page is formatted for printing and leaves the checkboxes, owner fields, evidence prompts, and launch gate visible.

Does completing this checklist prove legal or regulatory compliance?

No. This is general educational guidance, not a certification, legal opinion, regulatory approval, or substitute for review of a specific system, jurisdiction, industry, contract, or affected population. Use qualified professional review where the use case requires it.

Build the connected governance system

Use the checklist with human-centered design and accountable standards.