Deployment Labs
Deployment NotesAI governance

AI Governance Policy: A Copy-Ready Template for Who Approves AI and How

Deployment LabsUpdated 15 min read

Most companies that wrote an AI policy wrote the staff half: which tools are allowed and what data can go in. The other half is missing. Nobody wrote down who decides when a team wants to connect an AI tool to the customer database, or which uses need more than a manager's nod, or who can switch one off. That's what an AI governance policy settles. In a 2026 ISACA poll of more than 3,400 professionals, only 38% said their organization had a formal, comprehensive AI policy, and 39% didn't know whether they had a documented process for shutting down or overriding an AI system. This article gives you the governance policy to close that gap, ready to copy.

What is an AI governance policy?

An AI governance policy is the document leadership signs that decides how your company approves, owns, checks and stops its uses of AI: who approves each kind of use, which uses count as high risk, what every use needs before it goes live, and how often it's reviewed.

It's written for the people who make those decisions, not for every employee. Microsoft's own guidance puts it this way: "AI governance policies provide a structured framework for responsible AI use within your organization."

This guide is for COOs, IT and security leads, HR and operations leads, and partners at firms of 20 to 1,000 people that use AI rather than build it. It explains what the policy covers, gives you a template, and shows how to adopt it in a month. It isn't legal advice.

AI governance policy vs AI usage policy: what's the difference?

An AI usage policy is the page of rules every employee follows; an AI governance policy is the set of decisions behind it: who approves new uses, how risk is judged, who owns each use and who can stop it. You need both, and each should point to the other.

Two more documents sit close by. A framework is the whole system, and a council charter says how a decision-making group runs. Here's how the four divide the work:

Document Who reads it What it settles Our guide
AI usage policy Every employee Approved tools, data rules, checking output, reporting problems AI policy template
AI governance policy Leaders, use owners, the AI lead Who approves what, risk tiers, owners, vendor checks, the off switch, reviews This article
AI governance framework Leadership, auditors, customers The whole system, mapped to NIST, ISO 42001 or the EU AI Act AI governance framework
AI council charter Council members Who sits on the group that decides, how often it meets AI council

A company of 50 people can fold the governance policy and the council charter into one document. A company of 500 usually can't, because the people who approve AI are no longer in one room every week.

Do you need an AI governance policy?

Yes, once more than a handful of people use AI for work: without one, every new request is decided by whoever happens to be asked, and nobody can say who owns a tool when it goes wrong. No US law requires a document with this name, but several now reward having one.

The ISACA poll shows how common the gap is. Of the professionals it surveyed, 90% believed employees were using AI in their organization. Only 38% had a formal, comprehensive AI policy, up from 28% in 2025. A further 30% had a limited policy, and 25% had none. Over half (56%) didn't know how long it would take to halt an AI system during a security incident.

The laws that reach a mid-size company point the same way:

  • Texas. The Texas Responsible AI Governance Act presumes a business used reasonable care. It also says a defendant can't be found liable if it discovers a violation through "an internal review process" while substantially complying with NIST's Generative AI Profile or another recognized AI risk framework. A governance policy is where that review process is written down. Our Texas AI law guide covers the rest.
  • Colorado. The new law, which takes effect January 1, 2027, excludes a general-purpose chatbot only if nobody configured or intended it for consequential decisions and it is "subject to an acceptable use policy" that keeps its output out of them. Deployers must keep records of covered decisions for at least three years. The governance policy is what makes sure both happen. See our Colorado AI law guide.
  • ISO/IEC 42001. If a customer asks for certification, the standard requires top management to establish an AI policy, according to Bastion's summary of the clauses: one that fits the organization's purpose, frames its AI objectives and commits to meeting requirements and to continual improvement. More in our ISO 42001 guide.

What should an AI governance policy include?

A useful AI governance policy settles ten things: scope, principles, roles, risk tiers, what a new use needs before launch, the list of uses, vendors, decisions about people, incidents and the off switch, and reviews. Each one maps to a part of NIST's AI Risk Management Framework, the main US reference.

Section What it settles NIST AI RMF
1. Purpose and scope Which AI it covers, including AI features inside software you already buy GOVERN 1.1
2. Principles The few commitments every decision is checked against GOVERN 1.2
3. Roles Who signs, who runs it day to day, who owns each use GOVERN 2.1, 2.3
4. Risk tiers How much review each kind of use gets GOVERN 1.3
5. Approving a new use What a request must say, who decides, how fast GOVERN 1.4
6. The list of uses One record of every AI use and its owner GOVERN 1.6
7. Vendors Data, retention and training terms checked before purchase GOVERN 6.1
8. Decisions about people Hiring, pay, discipline and firing get the most review GOVERN 1.1
9. Incidents and the off switch How problems are reported and uses are stopped GOVERN 4.3, 6.2
10. Review and retirement When the policy and each use are reviewed, and how a use is retired GOVERN 1.5, 1.7

NIST's wording behind three of these is worth quoting, because it's what an auditor or a customer's questionnaire will check you against. GOVERN 1.3 asks for processes "to determine the needed level of risk management activities based on the organization's risk tolerance". GOVERN 1.6 asks for "mechanisms" to "inventory AI systems". GOVERN 2.3 says "executive leadership of the organization takes responsibility for decisions about risks associated with AI system development and deployment." NIST says the framework is "intended for voluntary use".

AI governance policy template: copy this

Copy the ten sections below into a document, replace everything in square brackets with a real name, role or number, and have the executive sponsor sign it. It's written for a company that uses AI rather than builds it. Keep it next to your AI usage policy, which staff follow day to day.

1. Purpose and scope

This policy sets how [Company] decides which uses of AI it adopts, who owns each one, how they are checked and how they are stopped. It covers every AI tool, AI feature and AI agent used for [Company] work by employees, contractors and our systems, including AI features inside software we already pay for. The staff rules for using approved tools are in the AI Usage Policy.

2. Principles

Every decision under this policy is checked against four commitments:

  1. We use AI where it does a job better, faster or more cheaply, and we can say which.
  2. A named person owns every use and answers for its results.
  3. People's data goes only into tools approved for that data.
  4. AI output alone never decides anything about a person's job, pay or access to our services.

3. Roles

  • Executive sponsor: [Role, e.g. COO]. Signs this policy, approves Tier 3 uses and answers to the leadership team for AI risk.
  • AI lead: [Name, role]. Keeps the list of uses, routes requests, answers within [10] working days, runs the reviews and keeps the records. Backup: [Name].
  • Use owner: the manager whose team relies on a use. Checks it works, keeps it within its approval and can switch it off.
  • Reviewers: [IT/security lead] for data, access and vendors; [HR lead] for anything touching staff or applicants; [counsel, internal or external] for Tier 3.
  • [AI council / leadership team]: meets [monthly] to decide Tier 3 requests and review incidents.

4. Risk tiers

Every use is placed in one tier when it's requested and moves up if its data, reach or purpose changes.

Tier What puts a use here Who approves Review
1. Standard An approved tool, used as the usage policy allows, with no new data or connections The use owner, recorded on the list Annual
2. Elevated A new tool or vendor, confidential or client data, a connection to a company system, or output sent outside the company The AI lead, with IT/security Every [six] months
3. High Decisions or recommendations about people (hiring, pay, discipline, firing, credit, housing, care), health data or financial records about individuals (bank, credit, payroll), an agent that can act on its own, or anything public-facing that speaks for us The executive sponsor, after review by [HR / counsel / IT] Every [three] months

5. Approving a new use

Anyone can request a new use through [the form / channel]. The request says what the use is for, which tool and plan, what data goes in, who is affected, what could go wrong and who will own it. The AI lead places it in a tier and answers within [10] working days with yes, no or "yes, if". Tier 2 and 3 uses don't go live until the owner can show how to switch them off.

6. The list of uses

The AI lead keeps one list of every approved use: the tool and plan, what it does, its tier, its owner and backup, the data it receives, its vendor terms and its next review date. A use that isn't on the list isn't approved. The list is checked against what's actually running at least [twice a year].

7. Vendors

Before we buy or switch on an AI tool or feature, [IT/security lead] reads the vendor's terms on data use, retention, model training and sub-processors, and records them on the list. We recheck them at renewal and whenever a vendor announces a change. A vendor switching on a new AI feature inside software we already use counts as a new use.

8. Decisions about people

Any use that informs a decision about an applicant, employee or customer is Tier 3. A person reviews every such decision and can overrule the output. We give the notices and explanations the law requires, and we keep records of covered decisions for [three] years or longer where a law requires it. Before a Tier 3 use goes live, [HR lead / counsel] confirms which laws apply in each state and country where it's used.

9. Incidents and the off switch

Anyone who sees an AI use go wrong reports it to [channel] the same day; reporting fast is never the problem. The use owner can pause the use at once and tells the AI lead. The AI lead logs every incident and its outcome, and Tier 3 incidents go to the executive sponsor within [one] working day. Each Tier 2 and 3 owner tests the off switch [once a year].

10. Review and retirement

The AI lead reviews each use on the date its tier sets, and whenever its vendor, data or purpose changes. A use that no longer earns its place is retired: access removed, data deleted or returned under the vendor's terms, and the list updated. The executive sponsor reviews this policy every [12] months and after any change in the law that applies to us.

Approved by [Name, role] on [date]. Next review: [date].

Fig. 1The three risk tiers in the template, with who approves and how often each is reviewed. Fill in your own roles and dates; the squares are illustrative.

How do you sort AI uses into risk tiers?

Sort each use by three questions: whose data goes in, who is affected by the output, and whether the AI can act on its own; the answer to the riskiest question sets the tier. Most uses land in Tier 1, which is the point: low-risk work gets a fast yes, so the review effort goes where it matters.

Three examples from a company of 200 people:

  • A marketing lead drafts blog outlines in the company's approved chat assistant. Approved tool, no confidential data, a person edits everything before it's published. Tier 1: the use owner records it on the list.
  • The finance team wants an AI feature in its expense software to read receipts. A vendor switching on a new feature, with company spending data but no individual's bank, credit or payroll records. Tier 2: the AI lead reads the vendor's terms with IT and says yes within ten days.
  • HR wants a tool that ranks applicants for interview. A recommendation about people, reached by hiring laws in several states. Tier 3: HR and counsel review it, the executive sponsor decides, and it's reviewed every quarter.

Microsoft's governance guidance uses the same logic for review dates: "Schedule quarterly risk assessments for high-risk AI workloads and annual assessments for lower-risk systems." For how to rate likelihood and impact within a tier, use our AI risk register.

Who owns the AI governance policy?

One executive signs the policy and answers for it, an AI lead runs it day to day, and every use has its own named owner who can switch it off. Spreading ownership across a committee with nobody in charge is the most common way a governance policy stops working.

The executive sponsor is usually the COO, or whichever executive owns the AI rollout. That matches NIST's GOVERN 2.3, which puts responsibility for AI risk decisions on executive leadership rather than on IT. The AI lead is the person staff actually go to: they keep the list, answer requests and chase reviews. In a company of 50 that might be a few hours a week of an operations lead's time. In a company of 500 it's often a dedicated role. If decisions need more than one person, our guide to setting up an AI council covers who should sit on it.

A governance policy is only as real as its off switch

Most governance policies list principles and approval steps, then stop. The test of whether one works comes later, on the day a use goes wrong. Who notices? Who can stop it, and how fast? ISACA's numbers suggest most organizations couldn't answer: 56% didn't know how long it would take to halt an AI system during a security incident.

That's why the template asks every Tier 2 and 3 owner to show how to switch their use off before it goes live, and to test it once a year. Our framework treats this as part of owning anything that runs: an owner and a backup, an off switch, a tested undo and review dates. The lesson on the off switch and a tested undo walks through setting one up for a tool, an agent or a way of working. It's free, and it takes one meeting to apply to your Tier 3 list.

The same idea runs through our framework as a whole: a governance policy that slows every request to a committee gets routed around, while one that gives low-risk uses a fast yes and puts the review effort on the few risky ones keeps AI deployment moving and under control at the same time.

Fig. 2What the policy asks of every AI use. The approval step is where the owner and the off switch get named; NIST's GOVERN 1.7 asks for a safe way to retire a system too.

How to write and adopt an AI governance policy

The template does the drafting. These six steps make it true for your company, in about a month.

01 · Name the executive sponsor and the AI lead

Pick the one executive who will sign and answer for the policy, and the person who will run it day to day, with a backup. Without these two names, every later step stalls. This gives you the roles for section 3.

02 · List the AI already in use

Before you can tier anything, you need the list: every tool, every AI feature vendors have switched on inside software you pay for, and who uses each. The free Shadow AI Audit builds a first version in minutes, and our guide to building an AI inventory shows what to record. This becomes the list in section 6.

03 · Place every use in a tier

Run each use through the three questions: whose data, who is affected, can it act. Most will be Tier 1. Flag the Tier 3 uses first, because they need owners and reviews before anything else. This gives you the tier table in section 4, filled in.

04 · Fill in the brackets and agree the review dates

Replace every bracket with a real name, role, channel or number. Don't leave "TBD": a policy with blanks gets read as optional. Agree the review dates for each tier with the people who will do the reviews.

05 · Have counsel review sections 4 and 8

Ask counsel to check the Tier 3 definition and the section on decisions about people against the laws in every state and country where you hire or serve customers. Colorado, Texas, Illinois and New York City all have rules that can reach AI used on applicants or employees. If you need help with this step, our AI governance service covers it.

06 · Sign it, publish it and tell people where the front door is

The executive sponsor signs and dates it. Publish it where managers can find it, link it from the AI usage policy, and tell everyone where to send requests and how fast they'll get an answer. The first Tier 3 review date goes in the calendar the same day.

What's still open

  • NIST's framework is being revised. NIST says the AI RMF 1.0 "is being revised as part of the White House AI Action Plan" and gives no date. The subcategories cited here may be renumbered or reworded.
  • Colorado's rules. The new Colorado law takes effect January 1, 2027, and by the same date the attorney general must adopt rules that clarify what a business has to tell someone after an adverse decision. The records and notice steps for Tier 3 uses may need more detail once those rules are final.
  • Federal preemption. Executive Order 14365, signed December 11, 2025, calls for "a minimally burdensome national standard – not 50 discordant State ones" and set up a Justice Department task force to challenge state AI laws. Until a court or Congress changes them, the state laws above apply as written.

We'll update this article when NIST publishes its revised framework or a state law cited here changes. The date at the top of the page shows when we last checked.

Sources

Questions

Common questions

Start with step one

Write the staff rules that sit under the governance policy.

The governance policy decides who approves AI. Staff still need the page of rules they follow day to day. The free AI Usage Policy generator writes it from your answers about the tools you use and the data you handle.