What is AI governance?
AI governance is the set of decisions an organization makes about the AI systems it builds and uses: which systems exist, who owns each one, what each is for, which providers are allowed, and how often those decisions are reviewed. It also covers the evidence that the decisions are followed in practice.
The decisions are the easy part to write down. A policy can say that every AI system has an owner, that only approved providers are used, and that each system is reviewed twice a year. The harder part is showing, at any point in time, that the policy matches what is running. That requires a current list of systems, a record of who decided what and when, and records from the controls that carry the decisions out.
The three frameworks most teams meet are the NIST AI Risk Management Framework (AI RMF 1.0), the international standard ISO/IEC 42001, and the EU Artificial Intelligence Act (EU AI Act). Each one describes that evidence in its own terms. For how they line up, see NIST AI RMF vs ISO 42001 vs EU AI Act.
How is AI governance different from AI security?
AI governance decides what is allowed, and AI security enforces those decisions on each request, tool call, and connection. Governance answers "should this system exist, who owns it, and which providers may it use?" Runtime security answers "should this specific action happen right now?"
The two layers depend on each other. A provider decision that nothing enforces is an instruction that people may or may not follow. A runtime rule that nobody owns is a configuration that nobody reviews when the system changes. For a side-by-side view of the two layers, see AI governance vs AI runtime security, and for the security side on its own, see what is AI security.
What do NIST AI RMF, ISO/IEC 42001, and the EU AI Act have in common?
All three deal with the same four things: a list of the AI systems you run, a named owner and purpose for each, decisions about which providers and uses are acceptable, and evidence that those decisions hold. They differ in legal force and in how much documentation they expect. The following table maps the four to NIST AI RMF and EU AI Act references. It leaves out ISO/IEC 42001 references, because that standard's text is sold rather than published openly.
| Shared ask | NIST AI RMF 1.0 | EU AI Act |
|---|---|---|
| An inventory of AI systems | GOVERN 1.6 | No general inventory duty. Article 49 requires registration of some high-risk systems in the EU database |
| An owner and a purpose per system | GOVERN 2.1, MAP 1.1 | Article 26(1) and 26(2) for high-risk deployers |
| Decisions about third-party providers | GOVERN 6.1 | No direct equivalent; Article 26 covers systems a deployer uses |
| Evidence that decisions hold | MEASURE 2.4, MANAGE 4.1 | Articles 12, 19, and 26(6) on logs |
The NIST AI RMF is voluntary guidance published by the National Institute of Standards and Technology. ISO/IEC 42001 specifies requirements for an AI management system, for any organization that provides or uses products or services that use AI. The EU AI Act, Regulation (EU) 2024/1689, is law, and many of its obligations apply only to high-risk systems. Because the shared controls overlap, one piece of work can support all three.
What is an AI system, for governance purposes?
An AI system is an application or agent together with the model providers and tools it uses. A support agent that calls a hosted model, reads from a ticketing system through an MCP (Model Context Protocol) server, and posts to a chat tool is one AI system, not four separate entries.
This definition matters because the unit you govern decides what an owner is accountable for. If you inventory providers alone, you know that people use a given model API, but not which application sends it customer data. If you inventory applications alone, you miss the providers and tools that carry the risk. Grouping an actor with what it uses gives an owner one record to review: the purpose, the providers, the tools, and the data classes the system is meant to handle.
The EU AI Act has its own legal definition of an AI system in Article 3. Use that definition when you classify a system under the Act. The working definition in this article is for inventory and ownership, and it covers more than the legal one, because a governance program also tracks tools that the Act doesn't regulate.
Why do AI inventories go stale?
AI inventories go stale because most are built from surveys, and people adopt AI tools faster than anyone fills in forms. A spreadsheet that lists the AI tools each team reported in March doesn't list the coding agent an engineer installed in April or the MCP server that agent started using in May.
A survey also can't say how complete it is. It records what people remembered and chose to report, and nothing in it shows what was left out. An observed inventory works the other way around. It records what devices connect to and what applications and agents report, so it captures adoption without anyone filing a request. It can also report its own coverage, for example as the share of enrolled devices that are reporting.
Observation has its own limit: it sees only what it can reach. Unmanaged devices, systems that never report, and AI features inside approved SaaS products stay invisible to it. The workable approach is to use observation as the base and ask people for the fields that need judgment, such as owner, purpose, and risk tier. For a full comparison, see observed vs survey-based AI inventory, and for the procedure, see how to build an AI inventory that stays current.
What does AI governance evidence look like?
AI governance evidence is a record that shows a governance decision exists and that a control acted on it. An auditor asks for the record, not for a description of the process. The following table lists common types of evidence and where each one comes from.
| Evidence | What it shows | Where it comes from |
|---|---|---|
| AI inventory with stated coverage | Which AI systems exist, and how complete the list is | Device observation and the agents that report to you |
| Owner and purpose per system | Who is accountable, and what the system is for | People, recorded on the inventory |
| Review dates | That someone checked the record within the review interval | The review action, with date and reviewer |
| Provider decisions and violations | Which providers are approved or prohibited, and use after a prohibition | Decision records, and device observations after the decision time |
| Runtime decision records | That controls allowed, denied, or held specific actions, with reasons | The enforcement layer: SDK, coding-agent hooks, or gateway |
| Change history | Who changed an owner, a decision, or a policy, and when | The audit log of administrative changes |
| Retirement records | That a system was taken out of service, and by whom | Deletion records that keep the system's history |
Evidence ages. A screenshot of a dashboard shows the state on the day it was taken, and nothing after. Evidence that comes from the control itself, exported on a schedule, stays current. For the framework view, see producing compliance evidence for AI agent activity and how to export AI governance evidence.
How do you start with AI governance?
Start with the inventory, because every other control attaches to it. The following steps take you from no record to evidence you can export. Each step links to a guide.
- Build an observed inventory. Turn on discovery for managed devices, and connect the applications and agents you build, so the list comes from what runs. See how to build an AI inventory.
- Assign an owner and a purpose to each AI system. Pick one person, not a team, and write down what the system is for. See how to assign owners to AI systems.
- Decide which providers are approved or prohibited. Record each decision against the provider or the specific client, and enforce it where people use AI. See how to approve or prohibit AI providers.
- Review each system on a schedule. Confirm the owner, purpose, and providers at least every 180 days, and sooner when a system changes. See how to run periodic AI system reviews.
- Export the evidence. Map the records to the frameworks you report against, and send them to your governance, risk, and compliance (GRC) tool. See how to export AI governance evidence.
To check your program item by item, use the framework-neutral AI governance checklist.
Which guides cover each part of AI governance?
Each part of an AI governance program has its own guide. The following table groups them by the question they answer.
| Question | Guides |
|---|---|
| Which frameworks apply, and what do they ask? | NIST AI RMF guide, ISO/IEC 42001 guide, EU AI Act for deployers, NIST AI RMF vs ISO 42001 vs EU AI Act |
| What should I check? | AI governance checklist, NIST AI RMF checklist, ISO 42001 Annex A checklist |
| How do I run the program? | AI inventory, AI system owners, approved AI providers, AI system review, evidence export |
| How do I handle specific cases? | AI risk tiers, AI purpose drift, AI system log retention, AI governance for coding agents |
| Which approach or product fits? | AI governance vs runtime security observed vs survey AI inventory OneTrust AI Governance vs Arcjet , IBM watsonx.governance vs Arcjet |
What does Arcjet not cover in AI governance?
Arcjet supplies evidence for some framework requirements, and it doesn't make an organization compliant with any framework. Its inventory covers 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.
The following work stays with your organization:
- Written AI policies, risk tolerance, and the risk management process itself.
- Staff training and AI literacy.
- Impact assessments, including a fundamental rights impact assessment under Article 27 of the EU AI Act.
- Legal classification of each system under the EU AI Act. Arcjet records the risk tier and role that you choose.
- Notices to workers, customers, and other people affected by an AI system.
- Model quality, fairness, and accuracy testing for models you train.
- AI systems that Arcjet can't observe, such as tools on unmanaged devices.
Where Arcjet fits
Arcjet builds the governance record from what it already observes and enforces. Discovery finds the AI providers that people reach from managed devices. The SDK and coding-agent hooks report the applications and agents you build and run. Guard records each decision it makes on a tool call, prompt, or model use. Arcjet groups each actor with the providers and MCP servers it uses into an AI system, suggests those groupings from what it observes, and asks people to confirm them.
On top of that inventory, team administrators record an owner, a purpose, a risk tier, and the EU AI Act role for each AI system. An owner must be a member of the team, and an owner who leaves the team reads as no owner, automatically. Provider decisions approve or prohibit a provider or a specific client, such as the ChatGPT app versus the OpenAI API, and the Discovery agent enforces them on managed devices. An observation of a prohibited provider after the decision is a violation attributed to a person. Each owner gets a weekly digest of violations and overdue reviews, and a notice when observed use drifts from the declared purpose. Every change to an owner, purpose, review, or provider decision is written to the audit log.
The evidence page maps these records to NIST AI RMF 1.0, ISO/IEC 42001, and the EU AI Act, plus frameworks that you define. It lists every requirement and marks each one as supported or not covered. You can export it as JSON or CSV, or push it to Vanta or Drata. The export is the point-in-time record, so keep it with your audit files.
To see the runtime side in practice, start with the Arcjet coding agents guide and publish a starter policy in dry run.
Frequently asked questions
Is AI governance a legal requirement?
It depends on where you operate and what your AI systems do. The EU AI Act is law and places duties on providers and deployers, many of them for high-risk systems only. The NIST AI RMF is voluntary, and ISO/IEC 42001 is a standard you choose to adopt. Customers and auditors often ask for the same controls even where no law requires them.
Who owns AI governance in a software company?
A senior leader usually owns the program and its policies, often the CTO, CISO, or a head of risk. Each AI system then needs one named owner, a person rather than a team, who is accountable for its purpose, providers, and controls. Engineering produces most of the runtime evidence, and security or compliance assembles it for audits.
What is the difference between AI governance and model governance?
Model governance covers the models you train: data lineage, validation, versioning, and performance monitoring. AI governance is broader. It covers every AI system you build or use, including third-party providers, coding agents, and chat assistants, and the decisions about who owns them and which uses are allowed. A company that trains no models still needs AI governance.
Do I need a GRC tool for AI governance?
A governance, risk, and compliance (GRC) tool helps you track controls, policies, and audits across frameworks, but it doesn't observe your AI systems. It needs evidence from the systems that do. Many teams keep the control list in a GRC tool such as Vanta or Drata and feed it inventory, ownership, and runtime records exported from the tools that produce them.
How long does it take to set up AI governance?
An observed inventory can list the providers and agents in use within days of deploying discovery and connecting your agents. Assigning owners and making provider decisions takes longer, because people have to make those calls, and the first full review cycle runs for the review interval, such as 180 days. Written policies, training, and impact assessments are separate work on their own timeline.
AI runtime security in your code
Protect your AI agent workflows with Arcjet
See your AI inventory, owners, provider decisions, and framework evidence in one place, built from what Arcjet observes and enforces.