Skip to content

AI · 11 min read

How to Write an AI Use Policy and Governance Framework for Your Business

A step-by-step guide to creating an AI use policy and a right-sized governance framework for small and mid-sized businesses, built on the NIST AI RMF.

Red file folders neatly arranged on a shelf
On this page 13 sections
  1. Why your business needs an AI policy now
  2. AI policy vs. AI governance framework
  3. The NIST AI Risk Management Framework in plain English
  4. Step 1: Assign ownership
  5. Step 2: Take an inventory of current AI use
  6. Step 3: Classify uses by risk
  7. Step 4: Write the acceptable use rules
  8. Step 5: Build a simple approval process
  9. Step 6: Monitor, train and improve
  10. What laws and standards should you know about?
  11. Common mistakes to avoid
  12. The bottom line
  13. FAQs

To write an AI use policy, decide who owns AI decisions, list the AI tools and uses already in your business, sort them by risk, then set clear rules on approved tools, data that must never be shared, human review and disclosure. Wrap those rules in a light governance framework, such as one based on the NIST AI Risk Management Framework, so the policy is enforced, measured and updated.

Many businesses already have AI in daily use without any written rules. This guide explains the difference between an AI policy and an AI governance framework, how to use the NIST framework without turning it into a bureaucracy, what a practical policy should contain and which laws and standards are worth knowing about. It is written for owners and executives of small and mid-sized US businesses, not for compliance specialists.

Why your business needs an AI policy now

The main reason is simple: your people are probably already using AI. In Microsoft’s 2024 Work Trend Index, based on a survey of 31,000 knowledge workers in 31 markets, 78% of AI users said they were bringing their own AI tools to work, rising to 80% at small and medium-sized companies. When that happens without guidance, three problems follow:

  • Data leaves your control. Customer records, financial figures, source code or confidential contracts get pasted into tools whose terms your business has never reviewed.
  • Quality and accountability blur. AI-generated content reaches customers without review, and nobody is sure who is responsible when it is wrong.
  • Legal exposure grows quietly. Existing laws on consumer protection, discrimination, privacy and intellectual property apply to AI-assisted work just as they apply to any other work.

The US Federal Trade Commission made the last point explicit when it announced Operation AI Comply in September 2024, stating that “there is no AI exemption from the laws on the books.”

A policy does not slow adoption down. Done well, it speeds it up, because employees know what they are allowed to do and new tools can be approved through a known process instead of case-by-case debate.

AI policy vs. AI governance framework

The two terms are often used interchangeably, but they do different jobs.

AI use policyAI governance framework
PurposeTells people what they may and may not do with AIDefines how the organization decides, oversees and improves its use of AI
AudienceEvery employee and contractorLeadership, system owners, IT, legal and risk
Typical lengthA few pages, written in plain languageA set of roles, processes, registers and review cycles
Examples of contentApproved tools, prohibited data, review and disclosure rulesRisk tiers, approval workflow, AI inventory, monitoring, incident handling
ChangesUpdated when rules changeRuns continuously as new uses are proposed and reviewed

A small business can start with the policy and a minimal framework: one owner, one register and one approval route. The framework grows as the number and risk of AI uses grows.

The NIST AI Risk Management Framework in plain English

The NIST AI Risk Management Framework (AI RMF), released on January 26, 2023, is the most widely referenced US framework for managing AI risk. It is intended for voluntary use, so adopting it is a choice, not a legal requirement. NIST also notes that the AI RMF 1.0 is being revised as part of the White House AI Action Plan, so it is worth checking for updates when you adopt it.

The framework document organizes AI risk management into four core functions:

FunctionWhat it meansWhat it looks like in a smaller business
GovernCulture, policies, roles and accountability for AI riskA named AI owner, the written policy and an approval process
MapUnderstanding the context of each AI use and its potential impactsAn AI inventory recording what each tool does, for whom and with what data
MeasureAssessing and tracking risks and trustworthinessTesting outputs before launch and spot-checking quality afterward
ManagePrioritizing and acting on risks, including incidentsFixing or retiring problem uses and learning from incidents

NIST describes Govern as “a cross-cutting function that is infused throughout AI risk management and enables the other functions.” In practice, that means your policy and ownership come first, and everything else depends on them.

The framework also lists characteristics of trustworthy AI: valid and reliable; safe; secure and resilient; accountable and transparent; explainable and interpretable; privacy-enhanced; and fair with harmful bias managed. These make a useful checklist when reviewing any proposed AI use.

The Generative AI Profile

In July 2024, NIST published NIST AI 600-1, a companion profile for generative AI. It names 12 risks that are unique to or made worse by generative AI. Several map directly to everyday business use:

  • Confabulation — confidently stated but false content, often called hallucination.
  • Data privacy — leakage or misuse of personal information.
  • Information security — new attack routes, including manipulation of AI systems.
  • Intellectual property — questions around copyrighted or confidential material in inputs and outputs.
  • Harmful bias or homogenization — outputs that disadvantage groups or flatten diversity of views.
  • Human-AI configuration — problems in how people use and rely on AI, such as over-trusting its output.
  • Value chain and component integration — risks inherited from third-party models, data and vendors.

You do not need to address all 12 in depth on day one. Use the list to make sure your policy has at least one clear rule for each risk that applies to your business.

Step 1: Assign ownership

Name one accountable executive for AI, often the COO, CTO or head of operations in a smaller company. Their job is not to approve every prompt, but to own the policy, the approval process and the AI inventory. For larger organizations, add a small cross-functional review group that includes IT or security, legal or compliance, and a representative from the business teams using AI.

Also decide who owns each AI system once it is in use. Every tool or workflow that touches customers, employees or sensitive data should have a named owner responsible for its quality, cost and incidents.

Step 2: Take an inventory of current AI use

You cannot govern what you cannot see. Run a short, non-punitive survey asking teams which AI tools they use, for what tasks and with what kinds of data. Check software subscriptions and expense reports, and ask your key software vendors which AI features are now switched on in the products you already pay for.

Record the results in a simple register. For each use, capture the tool, the owner, the purpose, the data involved, whether outputs reach customers or affect decisions about people, and where a human reviews the output. This register becomes the backbone of your governance framework.

Step 3: Classify uses by risk

Not every use of AI deserves the same scrutiny. A tiered approach keeps low-risk productivity uses fast while giving real attention to the uses that matter. A simple three-tier model works for most smaller organizations:

TierTypical examplesControls
LowDrafting internal emails, brainstorming, summarizing public informationApproved tools only; no confidential data; user checks output
MediumCustomer-facing content, code generation, analysis of internal business dataApproved tools with business-grade data terms; mandatory human review before use; owner named in the register
HighDecisions affecting people’s employment, credit, housing, insurance, health or legal rights; autonomous actions in business systemsFormal review before launch; testing for accuracy and bias; documented human oversight; legal review; ongoing monitoring

Some uses may belong in a fourth category: prohibited. Examples might include entering regulated personal data into consumer AI tools or using AI to make final decisions about individuals without human review. Where AI systems can take actions on their own, such as agents connected to your business software, treat them as at least medium risk. Our guide to AI agents for business explains the additional controls they need.

Step 4: Write the acceptable use rules

This is the part every employee reads, so keep it short and concrete. A practical AI use policy usually covers the following.

Approved tools

List the AI tools that are approved, for which tiers of use, and under which accounts. Business or enterprise accounts often come with different data handling terms from free consumer versions, so specify which one employees must use. Explain how to request a new tool.

Data rules

Be explicit about what must never be entered into AI tools that are not approved for it. Typical categories include customer personal information, employee records, payment and financial account data, health information, passwords and credentials, unreleased financial results, source code and anything covered by a confidentiality agreement.

Human review and accountability

State that people remain responsible for work they produce with AI assistance. Require review of AI output before it is sent to customers, published, used in code that ships or used in decisions about people. Remind staff that AI can produce plausible but false information.

Transparency and disclosure

Decide when customers or other people should be told that they are interacting with AI or receiving AI-generated content. Some laws already require notice in specific situations, as the legal section below shows, so check this with counsel for your sector and states.

Intellectual property and confidentiality

Address the use of third-party content in prompts, ownership of AI-assisted work and the need to respect client confidentiality. Your contracts with clients may already restrict how their information can be processed.

Prohibited uses

List uses that are not allowed at all, such as creating misleading content, impersonating real people or making final automated decisions about individuals in high-risk areas.

A one-page policy outline

If you need a starting structure, the following sections cover the essentials:

  1. Purpose and scope — why the policy exists and who it applies to, including contractors.
  2. Ownership — the accountable executive and how to reach them.
  3. Approved tools — the current list and how to request additions.
  4. Data rules — categories that must never be entered into unapproved tools.
  5. Human review — when AI output must be checked and by whom.
  6. Disclosure — when to tell people that AI is involved.
  7. Prohibited uses — what is not allowed.
  8. Risk tiers and approval — how new uses are assessed.
  9. Incident reporting — how to report problems.
  10. Review cycle — when the policy will next be updated.

Step 5: Build a simple approval process

A good approval process is quick for low-risk requests and thorough for high-risk ones. For each new tool or use, the requester describes the purpose, data and users. The AI owner assigns a risk tier and routes it: low-risk requests can be approved in days, while high-risk uses go to the review group.

For vendor tools, ask a consistent set of questions:

  • Is our data used to train the vendor’s models, and can that be turned off?
  • Where is data stored and processed, and for how long is it retained?
  • What security certifications or independent assessments does the vendor hold?
  • Which third-party models or subprocessors does the product rely on?
  • What logging and administrative controls are available to us?
  • How will the vendor notify us of incidents or material changes to the product?

The answers belong in your register next to the tool.

Step 6: Monitor, train and improve

Monitoring and incidents

Governance is not complete when a tool is approved. For medium- and high-risk uses, define what “good” looks like and check it regularly: accuracy on a sample of outputs, complaints or corrections, cost and usage. Review the register at least a few times a year to remove unused tools and catch new ones.

Set up a clear route for reporting AI problems, such as a wrong answer given to a customer, a suspected data leak or biased output. Treat these like other operational incidents: contain, investigate, fix and record what changed. NIST’s Manage function covers exactly this loop.

Training people, not just publishing rules

A policy that nobody understands will not be followed. Give every employee a short briefing on the rules, with realistic examples from their own work. Give heavier users practical training on writing good prompts, checking outputs and recognizing the limits of the tools. Make it easy to ask questions, and treat early mistakes as a chance to improve the policy rather than a reason to punish people for being candid about their AI use.

What laws and standards should you know about?

AI regulation is changing quickly, and obligations depend on your industry, location, customers and how you use AI. The following are useful reference points, not a complete list, and you should consult legal counsel about how they apply to your business.

  • Existing US law. Consumer protection, anti-discrimination, privacy and sector rules continue to apply when AI is involved, as the FTC’s Operation AI Comply announcement emphasized.
  • State AI laws. Colorado’s governor signed SB26-189, on automated decision-making technology, on May 14, 2026. According to the Colorado Division of Real Estate’s summary, it applies to consequential decisions made on and after January 1, 2027, and requires deployers to give notice of its use and to provide information to consumers after adverse outcomes. Other states have their own rules, and the landscape continues to shift.
  • The EU AI Act. If you offer AI-enabled products or services in the European Union, the European Commission’s AI Act page states that the Act entered into force on August 1, 2024, with prohibited practices applying from February 2, 2025. Following the Digital Omnibus changes, the Commission lists December 2, 2027 as the start date for obligations on high-risk systems in sensitive areas such as employment and education.
  • ISO/IEC 42001. Published in December 2023, ISO/IEC 42001 specifies an AI management system for organizations that develop, provide or use AI. It is certifiable and may matter if enterprise customers ask about your AI practices.

For most small and mid-sized businesses, a policy built on the NIST AI RMF gives a sensible foundation that can be extended toward specific legal requirements or ISO/IEC 42001 as needed.

Common mistakes to avoid

  • Banning AI outright. Blanket bans tend to push use out of sight rather than stop it.
  • Copying a large company’s policy. A long, legalistic document written for a different organization will not be read or followed.
  • Treating the policy as a one-time project. Tools, laws and uses change quickly, so schedule regular reviews.
  • Ignoring AI inside existing software. Many tools you already use have added AI features that may process your data differently.
  • Leaving ownership vague. Without a named owner, the register goes stale and approvals stall.

The bottom line

An AI use policy tells people how to use AI safely. A governance framework makes sure the policy is followed, measured and improved. Start with ownership, an inventory and a short set of clear rules, use the NIST AI RMF’s Govern, Map, Measure and Manage functions to structure what follows, and scale oversight to the risk of each use.

If you would like help turning this into a policy and governance model that fits your business, explore our Strategy & Consulting services and our AI Transformation approach, or start a conversation with our team.

Sources

FAQ

Questions, answered.

What should an AI use policy include?

A practical AI use policy covers its purpose and scope, the accountable owner, approved tools, data that must never be entered into unapproved tools, when AI output needs human review, when to disclose AI use, prohibited uses, how new tools are approved, how to report incidents and when the policy will be reviewed. Keep it short and written in plain language.

What is the difference between an AI policy and AI governance?

An AI use policy tells employees what they may and may not do with AI. AI governance is the wider system that decides, oversees and improves how the organization uses AI, including ownership, an inventory of AI uses, risk tiers, approval workflows, monitoring and incident handling. The policy is one part of the governance framework.

Is the NIST AI Risk Management Framework mandatory?

No. NIST states that the AI Risk Management Framework is intended for voluntary use. Many organizations adopt it because it offers a clear, widely recognized structure built on four functions: Govern, Map, Measure and Manage. Specific laws may impose their own obligations, so check with legal counsel about what applies to your business.

Do small businesses need an AI policy?

Yes, in most cases. Surveys show employees often bring their own AI tools to work, which can expose confidential or personal data and put unreviewed content in front of customers. A short policy with a named owner, approved tools and clear data rules reduces those risks and makes it easier to adopt AI with confidence.

Should we ban employees from using ChatGPT and other AI tools?

Blanket bans tend to push AI use out of sight rather than stop it. A more effective approach is to approve specific tools under business accounts with suitable data terms, set clear rules about what information must never be entered, require human review for important outputs and provide a quick way to request new tools.

How do you classify AI use cases by risk?

A simple approach uses tiers. Low-risk uses such as drafting internal emails need only approved tools and basic care. Medium-risk uses such as customer-facing content need human review and a named owner. High-risk uses, such as decisions affecting people's employment, credit, housing or health, need formal review, testing, documented oversight, legal input and ongoing monitoring.

What is ISO/IEC 42001?

ISO/IEC 42001, published in December 2023, is an international standard that specifies an AI management system for organizations that develop, provide or use AI-based products or services. It can be certified, which may be useful when enterprise customers ask for evidence of responsible AI practices. Many smaller businesses start with a lighter NIST-based approach first.

How often should an AI policy be updated?

Review the policy and your AI inventory on a regular schedule, such as every few months, and whenever you adopt a significant new tool, launch a higher-risk use, experience an AI incident or become subject to new legal requirements. AI tools and regulations change quickly, so a policy that is never revisited soon becomes outdated.

Have something ambitious in mind?

Tell us where you are and where you want to be. We'll come back with honest thinking on how to get there.