An AI governance policy sets organizational control — who owns AI risk, how systems get classified and approved, and how oversight works. An AI usage policy sets employee behavior — which tools staff may use and how. Regulators increasingly expect both; treating them as one document leaves gaps in either compliance or daily practice.
Most enterprises that get an audit finding on AI don’t get it because nobody wrote a policy. They get it because one policy tried to answer two different questions at once: “how does the organization control AI?” and “what am I allowed to do with AI?” Those need separate owners, audiences, and review cycles — the EU AI Act, ISO/IEC 42001, and the NIST AI RMF all point toward that split, even though none of them uses the word “usage policy.” Fruggr’s own overview of AI governance covers the strategic “why”; this piece covers the document-level “what goes where.”
A year ago, a single-page “acceptable use of ChatGPT” memo was often the entire AI governance program at a mid-sized enterprise. That’s no longer enough once AI systems influence credit decisions, hiring, insurance underwriting, or safety-relevant processes — the “high-risk” or “consequential” uses EU, US, and Canadian rules now single out for structured oversight, not just behavioral guidelines.
What an AI Governance Policy Actually Covers
An AI governance policy is a leadership-facing document. It answers “how does the organization control AI?” — not “what am I allowed to type into a chatbot?” ISO/IEC 42001’s clause on AI policy (5.2) requires the policy to be relevant to the organization’s actual AI use, to provide a framework for setting measurable AI objectives, to commit explicitly to meeting applicable legal and regulatory requirements, and to commit to continual improvement as the organization’s AI footprint changes. In practice, that translates into five recurring sections:
- Scope and applicability — which systems, business units, and AI use cases (build, buy, or embedded-in-a-vendor-product) the policy governs.
- Roles and accountability — a named owner (increasingly a Chief AI Officer, AI governance committee, or existing risk/compliance function), with clear decision rights rather than a diffuse “everyone is responsible” statement.
- Risk classification and control requirements — how a system gets tiered (e.g., prohibited, high-risk, limited-risk) and what controls attach to each tier before deployment.
- Approval workflow and change control — what has to happen before a new AI use case goes live, and what triggers re-review when it changes.
- Monitoring, incident response, and audit — how the organization detects drift or failure after deployment, not just at launch.
The NIST AI RMF’s Govern function describes essentially the same list from a risk-management angle: defining and communicating an AI risk policy, appointing accountable roles, setting the organization’s risk tolerance, and putting board- or management-level review in place. NIST frames Govern as the function that makes the framework’s other three functions — Map, Measure, Manage — actually operate as an ongoing capability rather than a one-time project.
What Belongs in the AI Usage Policy Instead
The usage policy is the employee-facing counterpart. It answers “what am I allowed to do with AI, day to day?” — and it should be short enough that people actually read it. Typical contents:
- Approved tools and prohibited platforms — which AI assistants and browser extensions staff may use for work, and which are off-limits.
- Data-handling rules — what categories of data (customer PII, source code, unreleased financials, regulated health or financial data) may or may not be entered into an AI tool.
- Rules for personal accounts — whether an employee’s personal ChatGPT or Gemini login may touch company data at all.
- Output verification — the expectation that AI-generated content gets checked before it’s used in a customer-facing or regulated context, not published as-is.
- Incident reporting — what to do if an AI tool produces something wrong, biased, or exposes data it shouldn’t have.
Without a governance policy behind it, a usage policy has no teeth — a list of rules with no named owner or consequence structure. Without a usage policy, a governance policy never reaches daily workflow. Fruggr’s AI Governance Platform FAQ covers related employee-facing questions, including Shadow AI, that a usage policy typically needs to answer.
What the Regulations Actually Point To
No single regulation uses the phrase “AI governance policy,” but three major frameworks converge on requiring the governance-level document specifically, not just a usage memo:
| Framework | What it requires | Which document it points to |
|---|---|---|
| EU AI Act, Article 4 | Providers and deployers must take measures to support AI literacy among staff operating AI systems, tailored to their role and context — in force since February 2, 2025 | Both — literacy training sits under governance ownership, but usage policy is what staff actually read |
| ISO/IEC 42001, clause 5.2 | A documented AI policy, approved by top management, providing a framework for AI objectives and a compliance commitment | Governance policy |
| NIST AI RMF, Govern function | A defined AI risk-management policy, named accountable roles, and management/board-level oversight of AI risk | Governance policy |
The EU AI Act is the one framework here that explicitly reaches individual staff (Article 4 applies to “providers, deployers and affected persons”), which is why it doesn’t resolve the governance-vs-usage split on its own — it needs both a governance structure that assigns responsibility for literacy training, and a usage-facing artifact proving staff received it. Fruggr’s EU AI Act overview covers the regulation’s broader risk-tiering structure.
How This Changes for US and Canadian Organizations
The EU AI Act’s staff-literacy angle doesn’t have a direct US federal equivalent, so North American organizations tend to anchor their governance policy to the NIST AI RMF’s Govern function instead — voluntary, but increasingly the reference framework state regulators and industry auditors point to when asking “where’s your AI governance policy.” Three state- and sector-specific wrinkles matter for scoping the policy correctly:
- Colorado repealed its original AI Act (SB 24-205) after a federal court paused enforcement in 2026, replacing it with SB 26-189, a narrower “automated-decision-making technology” law covering consequential decisions (employment, housing, credit), effective January 1, 2027. The Attorney General’s rulemaking was still in progress as this piece was drafted — verify final rule language before finalizing risk tiers tied to this law. Fruggr’s Colorado AI Act reset guide covers the replacement law in full.
- California SB 53 (the Transparency in Frontier AI Act) imposes governance and disclosure obligations, but its core duties apply specifically to large “frontier” AI developers, not to every enterprise deploying third-party AI tools — a governance policy shouldn’t assume SB 53 applies without checking the enacted thresholds against the organization’s actual role (developer vs. deployer).
- Canada’s federally regulated banks and insurers face OSFI Guideline E-23 on model risk management, which reaches AI/ML models specifically and has its own named accountability and validation requirements — a governance policy for a Banking/Insurance organization operating in Canada needs a section that maps to E-23 directly, not just a general “we follow NIST” statement.
None of this changes the governance-vs-usage split described above — it changes which regulatory column the governance policy needs to cite as its “compliance commitment,” per ISO 42001’s own wording.
A Practical Structure You Can Adapt This Quarter
A governance policy doesn’t need to be long to be complete. A workable first version runs six sections: (1) purpose and scope, (2) a risk classification scheme with three or four tiers, (3) named roles — an owner, an approval body, and an escalation path, (4) an approval and change-control workflow tied to those tiers, (5) monitoring and incident-response expectations, and (6) a review cadence, typically annual or triggered by a material change in AI use. Keep the usage policy genuinely separate and shorter, referencing the governance policy rather than restating it — that separation is what lets each document stay current without a full rewrite every time the other changes.

Key takeaways
- An AI governance policy and an AI usage policy answer different questions — “how does the organization control AI?” versus “what am I allowed to do with AI?” — and need separate owners, not one combined memo.
- ISO/IEC 42001 (clause 5.2), the NIST AI RMF (Govern function), and the EU AI Act (Article 4) all point toward a governance-level policy document specifically, even though none of them uses that exact term.
- US and Canadian organizations should anchor governance policy to the NIST AI RMF by default, then layer in sector- or state-specific requirements — Colorado’s SB 26-189, California’s SB 53, or OSFI E-23 for federally regulated financial institutions — rather than assuming one framework covers all of them.
Next step: See how Fruggr’s compliance module maps your AI system inventory to EU AI Act, NIST, and ISO 42001 requirements automatically, so your governance policy’s risk tiers stay backed by evidence instead of a spreadsheet: AI Regulatory Compliance Software.
FAQ
Both, if AI touches anything beyond low-stakes internal drafting. The usage policy alone has no accountability structure behind it; the governance policy alone rarely reaches daily employee behavior. Regulators and auditors increasingly expect to see both as separate artifacts.
Increasingly a named accountable role — a Chief AI Officer, an AI governance committee, or an existing risk/compliance function extended to cover AI — rather than a diffuse “IT and Legal share it” arrangement. The NIST AI RMF’s Govern function treats named accountability as a prerequisite, not a nice-to-have.
At minimum annually, and immediately after any material change — a new high-risk use case, a new jurisdiction, or a regulatory update such as Colorado’s 2026 replacement of its original AI Act. ISO/IEC 42001 requires a continual-improvement commitment rather than a “set once” policy.
Article 4 itself applies to providers and deployers within the EU AI Act’s scope, which can include non-EU organizations whose AI systems are used in the EU. Outside that scope, there’s no direct US or Canadian equivalent, though the underlying idea — staff need to understand the AI systems they operate — shows up in the NIST AI RMF’s Govern function and in OSFI E-23’s accountability requirements for regulated financial institutions.
This article was drafted with AI assistance and reviewed by our editorial team before publishing.