What is AI governance?

AI governance is the organizational arrangement of policies, roles and decision processes for managing how AI is developed, acquired, used, reviewed and retired. NIST’s AI Risk Management Framework 1.0, GOVERN function, connects technical work with organizational policies and addresses the product lifecycle.

The practical question is who can decide what happens next. A tool may generate an answer, a classifier may flag it and a reviewer may correct it. Governance identifies the authority, responsibilities and records around those steps.

Make decisions traceable

For each AI-enabled workflow, identify its purpose, permitted information and intended users. Name who may approve use, who reviews unresolved problems and who can pause the workflow. A policy file without an accountable decision-maker leaves those questions unanswered.

Keep the reason for a decision with its supporting evidence. Approval to draft internal event copy is different from approval to publish customer-facing announcements. Record the boundary rather than treating “approved AI tool” as permission for every possible use.

Connect ownership with ongoing review

NIST’s GOVERN 2.1 category calls for documented roles, responsibilities and communication lines. Its GOVERN 1.5 and 1.6 categories address planned monitoring, periodic review and an inventory of AI systems, with review responsibilities and frequency defined.

Use an inventory entry that helps someone make a decision: the workflow, its owner, the tool or model used, permitted data, output destination and current review status. Link it to the tests, known limits and changes that matter for that use.

NIST describes GOVERN as a continuing function across AI risk management. It is not confined to a single approval before launch.

A decision record for a fictional workflow

The table below is an illustrative planning aid for a fictional team drafting event listings. It assigns responsibilities within this example; it does not prescribe job titles, company-size thresholds or a universal approval structure.

Proposed decisions, owners and records
DecisionExample ownerEvidence to retainReview trigger
Allow a draft taskWorkflow ownerApproved calendar, permitted inputs and draft-only scopeA new input source or intended use
Release a listingListing reviewerChecked date, location and wording, with the release decisionA calendar correction or unresolved claim
Change the systemTool administrator and workflow ownerChanged version or settings and the relevant checksNew behavior, access or failure handling
Pause or retire itWorkflow ownerReason, affected work and the replacement processUnresolved failures or discontinued use

A fictional event-listing draft

Imagine a community events team using AI to adapt an approved public calendar into draft listings. The tool receives the supplied event details and returns copy for review. It has no authority to publish a listing or invent a venue when information is missing.

If a draft changes an event date, the listing reviewer compares it with the calendar and withholds the incorrect listing. The workflow owner decides how to handle repeated failures or a proposed new data source. This is a hypothetical assignment of responsibilities, not a customer case or a measured reduction in incidents.

NIST describes its AI Risk Management Framework as intended for voluntary use. It offers a risk-management approach; adoption is not a determination that a particular organization meets applicable law.

Separate the framework you choose, the policy your organization adopts and any obligations that need assessment for the actual system, use and location. Do not replace a necessary specialist review with a checklist or assume an internal approval establishes legal compliance.

AI guardrails concern checks and constraints within the workflow. Governance determines who owns their configuration, approves exceptions and responds to failures. A successful automated check does not identify the person responsible for the release decision.

Match the arrangement to the actual use

NIST says framework users may apply functions according to their resources and capabilities, selecting categories or subcategories as appropriate. That flexibility is part of using its framework, not a guarantee that every selected approach is sufficient.

Choose responsibilities that can actually be carried out. A small drafting workflow may assign named people to clear decisions; a consequential application may need more specialized review. Assess the task, affected people, information access and unresolved risks rather than choosing a structure from employee count alone.

Specify how an exception is requested, who may decide it and where the reason is recorded. Set review triggers for changes in use, data, model behavior or output destination. Routine review and a response to a reported problem should both have an owner.

Plan for changes, failures and retirement

NIST’s GOVERN 1.7 category includes processes for decommissioning or phasing out AI systems. Governance therefore needs a stopping decision as well as a starting decision.

When a workflow is paused, identify pending work, who may resume it and what needs checking first. When it is retired, document the replacement process and relevant access or retained-record decisions. A closed inventory entry alone does not demonstrate that the old workflow stopped operating.

AI hallucination concerns false or context-inconsistent generated claims. Factual review can supply evidence for a governance decision, but it does not replace ownership or approval. Guardrails supply technical controls; governance assigns the surrounding responsibilities.

Sources and scope

This glossary uses NIST’s voluntary AI RMF 1.0 and its current framework page. NIST states that AI RMF 1.0 is being revised. Check the current edition when adopting its categories or supporting resources.