AI governance checklist: 20 controls for teams shipping AI

Inventory, owners, provider decisions, reviews, runtime controls, and logs, each mapped to NIST AI RMF and EU AI Act references.

13 min read
In short: This framework-neutral checklist covers 20 governance controls that map to the NIST AI RMF and the EU AI Act: an inventory with stated coverage, one owner and a purpose per AI system, provider decisions enforced on devices, reviews every 180 days, runtime policies, human oversight, logs, and exported evidence. Each item names the record that shows it's done.

What is an AI governance checklist?

An AI governance checklist is a list of the governance controls that the major AI frameworks have in common: an inventory, an owner and purpose per system, provider decisions, reviews, runtime controls, human oversight, logs, and evidence. Each item names the NIST AI RMF and EU AI Act references it supports, so one piece of work can count toward more than one framework.

This checklist has 20 items. It covers governance: what exists, who owns it, what is allowed, and how you show it. For runtime threats such as prompt injection and data leaks, use the AI security checklist instead. For the background on each framework, see what is AI governance and NIST AI RMF vs ISO 42001 vs EU AI Act.

How do you use this AI governance checklist?

Work through the items in order, and for each one, find the record that shows it's done. If you can't point to a record, the item is open, even if the practice exists.

Each item lists its framework references in a short form:

  • N is the NIST AI Risk Management Framework (AI RMF 1.0), with its subcategory identifier, such as GOVERN 1.6.
  • E is the EU Artificial Intelligence Act (EU AI Act), with its article. Most EU AI Act references apply to high-risk systems only.

This checklist doesn't cite ISO/IEC 42001 clauses or controls, because that standard's text is sold rather than published openly.

References are pointers, not quotations. Check each one against the framework text before you rely on it in an audit. The "Evidence" column names the record that shows the item is done and where that record comes from.

An AI system is an application or agent together with the model providers and tools it uses. Each item applies per AI system unless it says otherwise.

How do you check your AI inventory?

Check that the inventory lists every AI system you build or use, and that it states how complete it is. Every other item attaches to an entry in this list.

ControlWhy it mattersReferencesEvidence
  1. Keep an inventory of every AI system you build or use

You can't assign an owner, decide on a provider, or review a system that isn't on the list.

N GOVERN 1.6

An inventory built from device observation and the agents that report to you

2. Measure how complete the inventory is

An inventory that can't state its coverage can't show what it missed.

N GOVERN 1.6

Coverage as the share of enrolled devices that are reporting, plus a written list of sources it doesn't reach

For how to build the inventory, see how to build an AI inventory that stays current.

How do you check AI system ownership and purpose?

Check that each AI system has one named owner, a written purpose, the data classes and tools it may use, and a risk tier. These fields turn an entry in a list into something a person is accountable for.

ControlWhy it mattersReferencesEvidence
  1. Name one owner per AI system, and let ownership lapse when the owner leaves

A team can't be accountable, and an owner who has left the company can't answer for anything.

N GOVERN 2.1, E Art. 26(2)

Owner per AI system, tied to team membership, with changes in the audit log

4. Document each system's purpose

The purpose is the baseline every later review and drift check compares against.

N MAP 1.1, E Art. 26(1)Purpose recorded on the AI system
  1. Declare the data classes and tools each system may use

A declaration can be compared with what the system does; a free-text purpose can't.

N MAP 1.1, E Art. 26(4)Declared data classes and tools on the AI system
  1. Record a risk tier and, if you have EU exposure, the AI Act role

The tier sets how much review and control a system gets, and the role sets which EU AI Act duties apply.

N GOVERN 1.3, E Art. 6

Risk tier and EU AI Act role on the AI system; the classification itself is a legal judgment you make

For what an owner is accountable for, see how to assign owners to AI systems. For tiers, see how to classify AI systems by risk.

How do you check decisions about AI providers?

Check that every AI provider in use has an approve or prohibit decision, that the decision is enforced where people use AI, and that use after a prohibition is counted.

ControlWhy it mattersReferencesEvidence
7. Approve or prohibit every AI provider in use

An undecided provider is neither allowed nor blocked, so nobody knows whether its use is a problem.

N GOVERN 6.1

Decision per provider or client, with date and decider, and the share of observed providers with a decision

8. Enforce provider decisions on managed devices

A written list of approved tools doesn't stop anyone from opening a prohibited one.

N MANAGE 2.4

Device enforcement of each decision, from the endpoint agent on managed devices

  1. Count use after a prohibition as a violation, attributed to a person

Violations show whether a decision holds, and a name lets someone follow up.

N MANAGE 3.1

Violations: observations of a prohibited provider after the decision time, per person

For decision levels and exceptions, see how to approve or prohibit AI providers.

How often should you review AI systems?

Review each AI system at least every 180 days, and sooner when it drifts from its purpose or starts using a new provider. The review record is the evidence, not the meeting.

ControlWhy it mattersReferencesEvidence
  1. Review each system at least every 180 days, and record the review

Owners, purposes, and providers change, and a fixed interval catches changes nobody reported.

N GOVERN 1.5

Last review date and reviewer per AI system, and the list of overdue reviews

11. Review sooner on drift or a new provider

A system that changed last week shouldn't wait months for its next check.

N MANAGE 2.3, MEASURE 3.1

Drift notices and new-provider observations, with the review that followed

For the review procedure, see how to run periodic AI system reviews. For how drift is detected, see how to detect AI purpose drift.

Which runtime controls belong in an AI governance checklist?

The runtime controls that belong here are the ones that carry governance decisions out: policies on agent actions, testing before enforcement, change approval, human oversight, and sensitive-data detection. Governance decides what is allowed, and these controls produce the records that show the decision holds.

ControlWhy it mattersReferencesEvidence
  1. Put a runtime policy on every consequential action an agent can take

A documented purpose is only enforced when a control checks each action against it.

N MEASURE 2.4, MAP 4.2

Share of AI systems with at least one live policy, and the decision records it produces

13. Test policies in dry run before enforcing them

Dry-run results show what a rule would have blocked, so the decision to enforce is based on real traffic.

N MEASURE 2.4

Dry-run decision records, and the test a rule passed before going live

  1. Require a second approver to publish a policy change
One person shouldn't be able to loosen a control alone.No direct referencePolicy revision history naming the author and the approver
15. Route held actions to the system's owner

Human oversight works when the person deciding knows what the system is for.

N MAP 3.5, E Art. 14, 26(2)

Hold decisions with the owner's approval or rejection recorded

  1. Detect sensitive data in prompts, responses, and tool calls

You can't compare declared data classes with real use unless you detect the data.

N MEASURE 2.10

Sensitive-information findings by type, such as personal data, secrets, and health data

For the threat side of runtime controls, see the AI security checklist. For coding agents specifically, see AI governance for coding agents.

Which records and evidence do you need to keep?

Keep decision logs for at least six months, record retirements, export evidence on a schedule, and write down what your tools don't cover. These items make the rest of the checklist provable after the fact.

ControlWhy it mattersReferencesEvidence
  1. Keep decision logs at least six months, or longer where other law requires

The EU AI Act sets at least six months for logs of high-risk systems, and other rules can require longer: HIPAA keeps required documentation for six years.

E Art. 19, 26(6); N MANAGE 4.1

Retained decision records, in the product or in your security information and event management (SIEM) system

18. Record retirements, with date and person

A retired system should stop working and still leave its history behind.

N GOVERN 1.7

Retired AI systems with date and who retired them, with keys revoked

19. Export evidence to your GRC tool on a schedule

A governance, risk, and compliance (GRC) tool tracks controls, and it needs current evidence from the systems those controls govern.

N GOVERN 1.4Dated JSON or CSV exports, or a scheduled push to the GRC tool
  1. Write down what your tools don't cover, and own it

An auditor trusts a record that states its gaps more than one that doesn't.

All three frameworks

A list of requirements marked not covered, each with the person or process that covers it

For retention by framework, see how long to keep AI system logs. For exports, see how to export AI governance evidence.

What does Arcjet not cover in this checklist?

Arcjet supplies evidence for many items in this checklist, and it doesn't make an organization compliant with any framework. The following parts stay with your organization:

  • The risk tier and EU AI Act role in item 6 are legal judgments. Arcjet records the values you choose.
  • Item 20 is your own work. Arcjet lists the requirements it doesn't cover, and you decide who covers them.
  • Items 1 and 2 cover only what Arcjet observes: managed macOS and Windows devices with the Discovery agent, applications and agents that report to Arcjet, and coding agents with hooks installed.
  • Written AI policies, staff training, AI literacy, impact assessments, and notices to affected people aren't in this checklist, and Arcjet doesn't produce them. The framework checklists list them.

For the full item-by-item view of each framework, use the NIST AI RMF checklist, the ISO 42001 Annex A checklist, and the EU AI Act checklist.

Where Arcjet fits

Arcjet produces the evidence for most of this checklist from what it already observes and enforces. Discovery and the agents that report to Arcjet build the AI inventory and state its coverage (items 1 and 2). Team administrators record the owner, purpose, declared data classes and tools, risk tier, and EU AI Act role on each AI system, and every change goes to the audit log (items 3 to 6). Provider decisions are enforced on managed devices by the Discovery agent, and violations are attributed to people (items 7 to 9).

Reviews are due every 180 days. The weekly owner digest lists violations and overdue reviews, and owners get a notice when observed use drifts from the declared purpose (items 10 and 11). Guard decides tool calls, a rule must pass a test before it goes live, dry run records what a rule would do, a second approver publishes policy changes, and Hold decisions go to the system's owner (items 12 to 15). Sensitive-information detection covers personal data, secrets, and health data (item 16).

Arcjet keeps decisions for 30 days by default, with opt-in long-term retention, and a log drain exports decisions and captures to your own systems (item 17). Retiring an actor revokes its keys and keeps the record (item 18). The evidence page maps all of this to NIST AI RMF 1.0, ISO/IEC 42001, and the EU AI Act, lists every requirement as supported or not covered, and exports as JSON or CSV or pushes to Vanta and Drata (items 19 and 20).

Frequently asked questions

How is an AI governance checklist different from an AI security checklist?

An AI governance checklist covers decisions and accountability: which AI systems exist, who owns them, which providers are allowed, how often they are reviewed, and what evidence you keep. An AI security checklist covers runtime threats such as prompt injection, data leaks, and abuse of agent tools. Most teams need both, and the runtime controls in the governance checklist are the ones that carry governance decisions out.

Does completing this checklist make me compliant with the EU AI Act?

No. The checklist covers governance controls that the frameworks have in common. The EU AI Act also requires legal classification, AI literacy measures, transparency notices, and, for high-risk systems, duties such as fundamental rights impact assessments and provider documentation. Use the EU AI Act checklist and legal advice for those.

Why review AI systems every 180 days?

Neither the NIST AI RMF nor the EU AI Act sets a fixed interval. NIST AI RMF GOVERN 1.5 asks for planned periodic review with a defined frequency, and 180 days keeps records current without making review a constant task. Review sooner when a system drifts from its declared purpose or starts using a new provider.

Should an AI system owner be a person or a team?

A person. A team can't be held accountable for a decision, and NIST AI RMF GOVERN 2.1 asks for documented roles and responsibilities. Tie the owner to team membership so that ownership lapses when the person leaves, and the system shows as unowned until someone takes it over.

What evidence does an auditor accept for AI governance?

Records from the systems that do the work: an inventory with its coverage, owner and review fields with dates, provider decisions and violations, runtime decision records, and the audit log of changes. Dated exports of those records hold up better than screenshots, because a screenshot shows one moment and can't be traced back to its source.

AI runtime security in your code

Protect your AI agent workflows with Arcjet

Arcjet produces evidence for most items on this checklist from what it observes and enforces, and lists the requirements it doesn't cover.