How to Create an AI Acceptable Use Policy for a Small Business

The Direct Answer

A useful small-business AI acceptable use policy should answer twelve questions:

1. Who does the policy cover?
2. Which AI tools and account types are approved?
3. Which business purposes are allowed?
4. What information may be entered?
5. What information is prohibited or requires approval?
6. Which activities are prohibited?
7. Which outputs require human review?
8. Who owns the final decision and result?
9. How should AI-assisted work be disclosed or documented?
10. How are security, privacy, and intellectual-property concerns handled?
11. How are mistakes, suspected exposure, and policy violations reported?
12. Who maintains the policy as tools and business uses change?

NIST’s voluntary AI Risk Management Framework emphasizes defined policies, roles, responsibilities, and human oversight. The U.S. Small Business Administration advises small businesses to review AI outputs and avoid entering sensitive or proprietary information into tools that have not been approved for that purpose.

1. Define the Policy Scope

Begin by stating who and what the policy covers. Ambiguous scope encourages employees to assume the rules apply only to company-paid tools or formal automation projects.

The scope should also explain that using a personal or free account does not remove responsibility. Business information remains business information even when an employee enters it through a personal device or account.

2. Establish Approved Tools and a Tool-Request Process

A practical policy should identify how tools become approved rather than relying on a static list that becomes outdated.

A tool may be approved for one use and prohibited for another. For example, a public chatbot might be approved for brainstorming non-confidential headlines but not for customer records, contracts, source code, or internal financial analysis.

Minimum tool review questions

• Who owns and administers the account?
• What data does the tool receive?
• Can the vendor use submitted information to train or improve models?
• What retention, deletion, export, and access controls are available?
• Can employees connect external systems or upload files?
• Which users receive which permissions?
• How is access removed after role changes or offboarding?
• What happens if the tool is unavailable or produces unreliable results?
• Does the product support the organization’s privacy, security, contractual, and regulatory requirements?

3. Create Clear Data and Confidentiality Rules

The policy should translate the business’s information classifications into actions employees can follow.

CISA’s AI-data-security guidance highlights the importance of protecting the confidentiality, integrity, and trustworthiness of data used by AI systems. The FTC’s business data-security guidance similarly recommends limiting information collection, restricting access, and protecting sensitive information based on business need.

Use data minimization

Even with an approved tool, provide only the information required for the task. Remove unnecessary names, account numbers, addresses, credentials, customer identifiers, and internal details. Replace real data with approved test data when possible.

Do not assume anonymization is automatic

Deleting one obvious identifier may not remove the ability to recognize a customer, employee, transaction, or project from context. The policy should require additional review when a combination of details could reveal the person or organization.

Treat connected systems as data access

A connected AI assistant may access far more information than a person intends to place into one prompt. Review drive, email, CRM, calendar, browser, code-repository, and customer-system permissions before enabling a connector.

4. Define Permitted, Restricted, and Prohibited Uses

Provide examples that reflect the actual work performed by the business.

The policy should separate assistance from authority. AI may help prepare a draft, extract information, or identify patterns without receiving authority to make the final business decision.

Examples by function

5. Define Human Review and Accountability

“Human in the loop” is too vague unless the policy states who reviews, what they check, and what authority they retain.

NIST’s AI RMF Core specifically calls for policies and procedures that distinguish human roles and responsibilities in AI oversight. Its Generative AI Profile also recommends documenting system requirements, ownership, and human-oversight responsibilities.

6. Address Accuracy, Sources, Disclosure, and Intellectual Property

Verify material claims

Require users to verify facts, calculations, citations, quotations, dates, product details, customer information, and professional claims against reliable sources. AI-generated citations should not be trusted without opening and checking the source.

Do not fabricate evidence

Prohibit invented testimonials, clients, results, statistics, licenses, credentials, research, quotes, cases, or experience. AI-assisted content must follow the same truthfulness standards as any other business communication.

Define disclosure expectations

Not every internal use requires an AI label. The policy should identify when disclosure is appropriate because of customer expectations, contractual terms, platform rules, professional obligations, material automation, synthetic media, or the nature of the decision.

Review ownership and licensing

Users should not assume generated content is automatically exclusive, protectable, or free of third-party rights. Review the vendor’s current terms and the source material supplied to the tool. Obtain qualified advice for consequential intellectual-property questions.

7. Define Security and Access Controls

Security requirements should be proportional to the information and authority involved. A tool that only drafts from public material presents a different exposure than an agent that can read customer records and change a live system.

8. Establish Incident and Exception Reporting

Employees should know how to report a suspected exposure or harmful output without hiding the mistake.

Examples of reportable events

• sensitive information entered into an unapproved tool;
• unexpected access to files, email, customer records, or other systems;
• an incorrect or misleading output sent externally;
• an automated action outside the approved scope;
• a suspected prompt injection, data leak, account compromise, or malicious file;
• discriminatory, unsafe, or materially harmful output;
• use of a prohibited tool or activity;
• a vendor change that may affect approved use.

Define the immediate response

1. Stop the affected use or automation when safe to do so.
2. Preserve relevant records without spreading the sensitive information further.
3. Notify the designated owner or security contact.
4. Document the tool, account, information, action, user, and time.
5. Follow applicable legal, contractual, privacy, security, and customer-notification procedures.
6. Review the root cause and update controls, training, or tool approval.

The policy should also establish an exception-request process. Some valid use cases will not fit the default rules, but exceptions should be approved by named roles and documented with scope, time limit, controls, and review date.

9. Assign Ownership and a Review Schedule

AI-use policies become obsolete when no one owns them.

Review the policy on a defined schedule and after material events such as a new vendor, new integration, sensitive-data use, new automated action, significant incident, regulatory change, or expansion into a higher-impact use case.

10. Roll Out the Policy in Practical Steps

StartLab’s AI Readiness Assessment helps evaluate whether a business has the process, information, ownership, team participation, and safeguards required for a focused AI project. The guide to internal AI knowledge assistants addresses source ownership, permissions, citations, and human review for that specific use case.

One-Page AI Acceptable Use Policy Outline

Frequently Asked Questions