Employees often begin using generative AI before a company has decided what information may be entered, which tools are approved or who must review the output. An AI acceptable use policy turns that informal activity into a manageable business process.
The goal is not to ban useful technology or to approve every experiment centrally. It is to define boundaries that match the risk of the task. A short policy that employees understand and managers enforce is more useful than a long document nobody consults.
NIST’s Generative AI Profile specifically recommends transparent acceptable-use policies, clear responsibilities, system inventories and procedures for incident review. Its broader AI Risk Management Framework organizes work around Govern, Map, Measure and Manage. Small businesses can translate those ideas into a lightweight policy without pretending that one template satisfies every law or industry requirement.
Key takeaways
- List approved AI tools and permitted business purposes.
- Prohibit entry of passwords, private customer data, regulated records, confidential contracts and unpublished intellectual property unless a documented, approved arrangement protects them.
- Require human review before AI output affects customers, money, employment, safety, legal rights or public claims.
- Make disclosure rules specific to the channel and audience.
- Keep an AI-tool inventory, name an accountable owner and provide an incident-reporting route.
- Review the policy when tools, vendors, laws or business uses materially change.
What an AI acceptable use policy should accomplish
A useful policy answers six employee questions:
1. Which tools may I use?
2. What information may I enter?
3. What tasks are allowed or prohibited?
4. When must a person verify the result?
5. When must AI assistance be disclosed?
6. What should I do if something goes wrong?
If the policy cannot answer those questions quickly, employees will make inconsistent decisions. The document should connect to existing privacy, security, records, marketing, HR and intellectual-property rules rather than creating a separate universe for AI.
Use the SCOPE framework
SCOPE gives a small company a practical structure for policy design.
S — Systems
Maintain an inventory of approved tools, owners, intended uses and vendor terms. Record whether the tool trains on submitted data, how long it retains prompts, what administrators can see and whether enterprise controls are enabled.
An approved product is not automatically approved for every task. A public chatbot used for brainstorming a generic headline is different from the same chatbot receiving customer records or source code.
C — Content and confidentiality
Define information that must never be entered without specific approval. Typical restricted categories include authentication secrets, payment-card data, health information, government identifiers, private employee records, client-confidential files, legal advice, security configurations and unpublished product plans.
Use examples from the company’s real work. “Do not enter sensitive data” is too abstract. “Do not paste a client spreadsheet containing names, emails or order history” is actionable.
O — Oversight
Set review requirements according to impact. Low-risk brainstorming may need ordinary editing. Customer-facing claims require fact checking. Decisions affecting employment, access to essential services, legal rights, safety or significant spending need qualified human approval and may require legal or specialist review.
Human review should be meaningful. A reviewer needs the source material, authority to reject the output and enough time to check it.
P — Publication and provenance
State when the organization will disclose AI assistance and how it will preserve evidence. Disclosure may be appropriate when required by law, contract, platform rules, editorial policy or audience expectations.
For important outputs, keep the final approved version, key sources, reviewer and material changes. Do not retain sensitive prompts merely to create an audit trail; apply data-minimization and retention rules.
E — Escalation
Give employees one clear route to report accidental data exposure, harmful output, suspected copyright problems, security issues or a tool behaving outside its approved purpose. The policy should discourage concealment and define who can pause a tool while the issue is reviewed.
Classify uses by risk
Green: routine assistance
Examples include brainstorming generic ideas, improving the readability of non-confidential text, drafting internal checklists and summarizing public material that the employee verifies. Green does not mean “no review”; it means normal professional review is usually proportionate.
Amber: controlled business use
Examples include customer support drafts, marketing copy containing factual claims, code suggestions, translations, analysis of internal non-sensitive information and summaries used for management decisions. Require an approved tool, identified reviewer and source checks.
Red: prohibited or specialist approval
Examples include entering passwords or regulated personal records, generating fake reviews, impersonating a real person, making autonomous hiring or firing decisions, publishing unverified accusations, bypassing security controls or relying on AI as the sole decision-maker for legal, medical or safety-critical action.
Some red uses can become acceptable only through a separately approved system with contracts, security assessment, qualified oversight and applicable legal review. Others should remain prohibited.
A practical policy template
Purpose
This policy governs the use of artificial-intelligence tools for company work. Its purpose is to support useful experimentation while protecting people, confidential information, security, intellectual property and the accuracy of business decisions.
Scope
The policy applies to employees, contractors and temporary workers using company devices, accounts, information or work product. It covers public and enterprise AI services, embedded AI features, automated agents and AI-assisted software.
Approved tools and accounts
Use only tools listed in the company AI inventory for business work. Use the approved account type and settings. Do not create an unofficial integration, agent or API connection that can access company systems without authorization.
Data rules
Do not enter restricted or confidential information unless the tool, contract, purpose and access controls have been approved for that category. Remove unnecessary identifiers and use the least information required.
Human responsibility
The person using AI remains responsible for the output. Verify facts, calculations, citations, permissions and instructions before use. AI output must not be the sole basis for a high-impact decision.
Customer-facing and public content
Confirm claims against reliable sources. Do not fabricate testimonials, people, results or credentials. Follow applicable disclosure, copyright, advertising and platform rules.
Security and prohibited conduct
Do not use AI to evade access controls, create malware, expose credentials, impersonate people or perform unauthorized monitoring. Report suspected security weaknesses through the existing incident channel.
Incidents
Immediately report accidental disclosure, harmful or discriminatory output, suspected infringement, unauthorized tool use or unexpected system behavior to the named policy owner. Preserve relevant evidence without spreading the sensitive content further.
Enforcement and review
Violations are handled under existing security and conduct procedures. The owner reviews this policy at least annually and after material changes to tools, vendors, laws or use cases.
Build an approval workflow that people will use
Use a one-page request containing:
- Tool and vendor
- Business purpose
- Users and affected people
- Data categories entered or generated
- Integrations and permissions
- Expected benefit
- Main failure modes
- Required reviewers
- Retention and deletion settings
- Pilot limits and success measures
- Decision, owner and review date
Low-risk tools can follow a fast review. Tools connected to customer data, code repositories, HR systems or automated decisions need deeper assessment. The workflow should evaluate the specific use, not only the vendor’s general reputation.
A 30-day implementation plan
Days 1–5: discover actual use
Ask teams which tools they use and for what tasks. Create an inventory without framing the exercise as punishment. Hidden use is harder to manage than visible experimentation.
Days 6–10: define data and impact tiers
Map existing information classifications to AI prompts and uploads. Identify high-impact decisions and public channels that require stronger review.
Days 11–15: approve a starter set
Select a small number of tools with appropriate account controls. Document purposes, restrictions, owners and settings.
Days 16–20: publish and train
Teach employees with short scenarios from their daily work. Include a safe alternative for prohibited actions, such as using synthetic data or requesting an approved secure workspace.
Days 21–25: test incidents
Run a tabletop exercise: an employee accidentally pastes client data into a public AI service. Confirm who receives the report, who contacts the vendor, how access is reviewed and how affected parties or regulators would be considered.
Days 26–30: measure and improve
Review questions, exceptions and near-misses. Update ambiguous language. Measure approved adoption, review completion and incident response—not the number of prompts employees generate.
Common policy mistakes
Treating all AI use as equal
A grammar suggestion and an autonomous customer decision do not need the same controls. Risk tiers keep oversight proportionate.
Depending on a disclaimer
“AI may be inaccurate” does not replace verification, access control or accountability.
Writing rules without approved alternatives
If employees need a capability and the policy only says no, unsanctioned use may continue. Provide a safe tool or review route.
Ignoring embedded AI
AI appears inside office, CRM, design, analytics and security products. Inventory material embedded features, not just standalone chatbots.
Promising legal compliance
A general policy supports governance but does not prove compliance with privacy, employment, sector or AI laws. Obtain jurisdiction-specific advice for high-impact or regulated uses.
Frequently asked questions
Does a small business really need an AI policy?
If staff use AI for company work, basic rules reduce inconsistent handling of data, claims and approvals. The policy can be short and proportional to risk.
Should we ban public AI tools?
Not necessarily. A business can permit low-risk uses that involve no confidential information while reserving sensitive work for approved enterprise systems.
Who should own the policy?
Choose a manager with authority to coordinate operations, security, privacy and legal input. Very small businesses may assign the owner or operations lead, with external specialist advice when needed.
How often should it be reviewed?
Review at least annually and whenever a major tool, integration, law, vendor term or business use changes.
