How to Write an AI Use Policy Your Small Business Will Actually Follow

Your employee pastes a customer email into an AI assistant to make it sound friendlier. Your office manager uses a different tool to summarize meeting notes. Someone else asks a personal chatbot to rewrite a proposal because the company account was not handy.

Congratulations: your business already has an AI policy. It is just unwritten, inconsistent, and being invented by whoever opens a prompt box first.

A useful AI policy for a small business can fit on one page. It should answer six questions: Which tools may we use? What data stays out? What work may AI help with? What requires a person to review it? What should we record? Who owns questions and incidents?

The point is not to turn a ten-person company into a federal agency. It is to replace guesswork with house rules people can remember while they are doing the work.

Your policy should open a safe lane, not build a brick wall

Owners often approach an AI policy from one of two extremes.

The first is “do whatever helps.” That sounds nimble until confidential information lands in the wrong account or a confident draft reaches a customer without anyone checking it.

The second is a sweeping ban. That may feel safe, but it often drives useful work into personal accounts and side channels where the owner has even less visibility.

The better move is a clear operating lane: approved tools, permitted work, visible boundaries, and a simple escalation path.

NIST’s AI Risk Management Framework Playbook is deliberately voluntary and adaptable. Organizations can use the suggestions that fit their context instead of treating the framework as one giant checklist. That is the right spirit for a small-business policy: proportionate enough to use, specific enough to matter.

Write the six AI House Rules

Forget the grand policy language for a moment. Sit down with the people who actually use these tools and make six operating decisions.

1. Give the team one approved doorway

Name the tools, account types, and access method people may use for company work.

“AI tools are allowed” is not a rule. A consumer chatbot on a personal email address and a company-managed product with administrative controls are not the same doorway.

Your policy should say:

• which products are approved;
• whether company-managed accounts are required;
• who may add a new tool;
• where employees request access; and
• whether browser extensions, meeting bots, and AI features inside existing software need separate approval.

This is not a popularity contest. The approved list should follow your vendor privacy review, the information the tool will touch, and the control the business actually has.

2. Draw a data boundary people can recognize

“Do not share sensitive information” is the sort of sentence everyone agrees with and nobody can apply at 4:47 on a Friday.

Name the categories in plain language. Depending on your business, the keep-out list may include passwords, payment details, government identifiers, health information, employee records, confidential contracts, unpublished financial information, customer lists, or material covered by a nondisclosure agreement.

Then define what is acceptable. A fictional customer scenario may be fine. An anonymized paragraph may be fine. A real complaint containing a name, phone number, account history, and private context may not be.

The FTC warns that customers can reveal confidential business documents and their own users’ information when using model services, and that provider privacy commitments matter because those services may gain access to sensitive business data. Its guidance is a useful reminder that the prompt box is still a data destination, not a magical scratch pad. Treat it accordingly.

3. Approve jobs, not just software

A tool can be acceptable for one task and reckless for another.

Define the work AI may assist with. Good early candidates often include brainstorming, restructuring rough notes, drafting internal outlines, summarizing approved non-sensitive material, classifying low-consequence information, or preparing a first draft for review.

Then name the work that needs extra permission or stays human-led: legal conclusions, hiring decisions, financial commitments, safety advice, final customer promises, access changes, or actions that cannot be easily reversed.

The distinction is simple: approving a hammer does not approve every place someone might swing it.

4. Put a human checkpoint before consequence

“Review AI output” is too vague. Review it for what?

Match the checkpoint to the stakes:

Output	Minimum checkpoint
Internal brainstorm or rough outline	User checks relevance before relying on it
Summary of source material	User checks important details against the source
Customer-facing draft	Named employee checks facts, tone, promises, and private information
Advice involving money, people, contracts, safety, or regulated work	Qualified human owns the decision; AI does not make the final call
Action that changes an external system	Explicit approval and an audit trail before execution

NIST describes its AI framework as a way to manage risk across the design, development, use, and evaluation of AI products and systems. Its Generative AI Profile adds actions for risks unique to generative AI. In everyday business language, a good-looking answer is not the end of the process. You still need to decide who checks it, what “good” means, and what happens when it is wrong. Those controls should match the use case.

5. Keep receipts where the stakes justify them

Not every prompt deserves a museum exhibit.

For harmless brainstorming, elaborate logging may create more burden than value. For work that informs an important decision or reaches a customer, keep enough evidence to understand what happened.

That might include:

• the source documents used;
• the final human editor or approver;
• material corrections made;
• the date and tool used;
• the action taken; and
• any incident or unexpected behavior.

The record should help you answer a practical question later: How did this result become business action?

Joint NSA, CISA, FBI, and international guidance emphasizes protecting data throughout the AI-system lifecycle. Its technical scope is larger than an ordinary office policy, but the operating lesson travels well: data security is not a one-time choice at signup. It follows the information through use, storage, access, change, and disposal. Your policy should do the same at the scale of your business.

6. Name one owner and one escape hatch

If an employee is unsure, who answers?

If someone pastes the wrong data into a tool, what happens next?

If a provider changes its terms, who reviews the approved list?

Put a name or role in the policy. Give that person authority to approve tools, answer edge cases, pause use, coordinate an incident, and update the document. Then give the team a no-drama reporting instruction: stop, preserve the relevant details, and contact the owner immediately.

People hide mistakes when the process feels punitive or mysterious. A useful policy makes early reporting the fastest route out of trouble.

Turn those decisions into one readable page

Your first policy does not need to predict every future use of AI. It needs to make today’s work safer and tomorrow’s question easier to answer.

A practical one-page structure looks like this:

1. Purpose: We use approved AI tools to support work while protecting customers, employees, company information, and the quality of our decisions.
2. Approved doorway: Use only the named tools through company-approved accounts. Request new tools from the policy owner.
3. Data boundary: Never enter the listed categories unless the owner has approved the exact workflow and controls.
4. Permitted work: AI may assist with the named low-consequence tasks. The listed high-consequence work stays human-led or requires explicit approval.
5. Review and records: A person remains responsible for the final result. Customer-facing and consequential work gets the stated review and evidence.
6. Questions and incidents: When unsure—or after a mistake—stop and contact the named owner through the stated channel.

Add an effective date and a review date. Link to a separate approved-tools list if that list changes often. Keep one canonical copy so employees do not discover three competing versions in shared drives.

The FTC’s current small-business cybersecurity guidance encourages owners to discuss practical security advice with staff, not merely store it somewhere online. A short conversation is part of the control. A policy that nobody has practiced is just a PDF with excellent intentions.

Test the policy with a 20-minute rollout drill

Do not launch the document with “Please read and acknowledge.” Put three believable situations on the screen and ask the team what the policy requires.

Try these:

The awkward email. A customer sends a long complaint with contact details and account history. Can an employee paste it into the approved tool to draft a response? If not, can the content be minimized or anonymized first? Who reviews the final reply?

The tempting shortcut. A team member finds an AI meeting assistant that promises instant summaries. Can they connect it to tomorrow’s customer call? Who reviews the provider and approves the account?

The polished mistake. AI drafts a proposal containing a service promise the company never made. Who catches it? What should be recorded? Does the incident change the prompt, the review step, or whether AI belongs in that task?

If the team cannot answer quickly, the policy is not clear enough yet. Rewrite the rule, not the employee.

A policy cannot carry the whole operation

The document is a starting control, not a compliance certificate.

It does not replace industry-specific legal advice, contractual duties, vendor due diligence, account configuration, access controls, staff training, or a secure implementation. A healthcare practice, financial firm, school, government contractor, and neighborhood service company may need very different boundaries.

It also does not make every AI project ready. Use the AI readiness checklist for a specific initiative and the AI agent versus workflow guide before giving software more authority.

Your policy tells people how to enter the conversation safely. Good system design determines what happens after they do.

Give your team clarity before you give the tools more power

The best small-business AI policy is not the one with the most impressive vocabulary. It is the one an employee can remember when a real deadline, real customer, and very inviting prompt box appear at the same time.

Approve the doorway. Draw the data boundary. Name the jobs. Put a person before consequence. Keep the right receipts. Give questions and mistakes somewhere to go.

If those six decisions expose larger questions about tools, workflows, ownership, or risk, AI consulting can turn the policy from a defensive document into a practical operating plan for using AI with more confidence.