Every small business with more than a handful of employees now has AI tools in use across the team, frequently without the owner's explicit knowledge and almost always without a written policy governing that use. The result is a quiet compounding exposure: client data pasted into consumer chat tools, AI-drafted correspondence sent without review, and no record of what tools are being used for what purposes.
An AI policy is the single highest-leverage governance document a small business can adopt in 2026. It is not a long document, the best ones run two to four pages, and it is not a legal formality. It is an operational policy that tells employees what they may and may not do with AI tools, what data is and is not safe to input, and what review the output requires before it leaves the company.
Why an AI Policy Matters Now
The exposure from unmanaged employee AI use falls into four categories. Each one has produced real, documentable incidents at small businesses over the last 18 months.
| Exposure | Typical Incident Pattern |
|---|---|
| Client data leakage | Employee pastes client records into a consumer LLM for a drafting task |
| Confidential business data leakage | Pricing, payroll, or strategic information pasted into a chatbot for analysis |
| Unreviewed AI output in client-facing work | AI-drafted emails, proposals, or advice sent without human review |
| Compliance and regulatory exposure | AI tools used in regulated workflows (HIPAA, financial services) without vendor vetting |
The default setting on most consumer-tier LLMs historically allowed the provider to use submitted content for model training. Several providers have changed defaults and offer explicit opt-out, but the responsibility to verify and configure those settings rests with the business. A written policy converts this from an employee choice to a business standard.
The Four Elements of a Workable AI Policy
A workable policy has four sections. Longer policies exist, but the marginal value of length drops quickly, clarity and enforceability are more important than comprehensiveness.
Approved Tools List
A short list of AI tools the business has vetted and explicitly approves for work use. For most small businesses this includes the enterprise or business tier of one major LLM (Claude for Work, ChatGPT Enterprise, or Microsoft Copilot), the AI features of the accounting platform in use, and any purpose-built tools specific to the business. Tools not on the list are not approved, including consumer tiers of the tools that appear on the list at a higher tier.
Data Classification and Permitted Use
Three tiers: Public (marketing content, general business drafts) — any approved tool. Internal (operational data, non-client business information) — approved tools with enterprise data controls only. Confidential (client data, payroll, financial records, legal matters) — only tools with explicit data-handling commitments and no training use. The classification governs the tool.
Human Review Requirement
Every AI-generated output that leaves the company — emails, proposals, content, client deliverables — requires explicit human review before sending. The review is not optional and is not a checkbox. The policy should name who the reviewer is for each category of output.
Incident Reporting and Remediation
A clear path for employees to report accidental data exposure without fear of retaliation. Incidents reported promptly can often be mitigated (vendor-side deletion, notification, remediation); incidents hidden until discovery are substantially worse. The policy should name who to notify and on what timeline.
The Data Classification Question, In Practice
The hardest part of the policy for most small businesses is the data classification. Employees often cannot tell at the moment they're pasting whether a piece of text is Public, Internal, or Confidential. The policy has to make this easy.
When in doubt, classify data as Confidential. If an employee cannot immediately determine whether something is Public or Internal, the policy should default to the stricter treatment. This default has the effect of pushing borderline cases toward safety rather than toward disclosure.
The three-tier classification is simplest when illustrated with examples specific to the business. The table below is a starting template for a professional services firm:
| Classification | Examples | Approved AI Tools |
|---|---|---|
| Public | Marketing copy, blog outlines, website content, general concepts | Any approved tool |
| Internal | Internal procedures, non-client operational data, team communication drafts | Enterprise/business-tier LLMs with data controls |
| Confidential | Client records, financial data, tax information, legal matters, payroll, employee data | SOC 2 certified purpose-built tools only; no general LLMs |
The Human Review Discipline
The second-hardest part is enforcing human review. The temptation to send an AI-drafted email without reading it carefully is real, and it grows as the team gets comfortable with the tool's output quality. The policy has to anticipate this.
Weak Human Review Enforcement
- Review is implied but not explicitly required in writing
- No named reviewer for different output categories
- AI drafts sent directly from the tool without editing
- No record of which outputs went through AI drafting
- Errors caught by the recipient, not by internal review
Strong Human Review Enforcement
- Policy explicitly names review as required for every client-facing output
- Reviewer named by role for each output category (partner, senior associate, etc.)
- AI drafts always edited in the final medium (email, document) before sending
- Optional: low-friction marker or tag indicating AI drafting occurred for later review
- Errors caught internally at measurable rate before going out the door
Sample Policy Structure
A two-to-four-page AI policy that covers the essentials without overbuilding. The structure below is a starting template that can be adapted to the specific business.
- 1. Purpose and scope — what the policy covers and whom it applies to
- 2. Definitions — what counts as AI, what tools are in scope
- 3. Approved tools list — specific named tools and tiers permitted
- 4. Prohibited tools list — specific named tools and tiers not permitted
- 5. Data classification — three tiers with examples and governing tools
- 6. Human review requirement — explicit for every output category
- 7. Vendor and data privacy — SOC 2, training-data policies, retention
- 8. Incident reporting — contact, timeline, remediation expectations
- 9. Periodic review — policy reviewed and updated at least annually
- 10. Employee acknowledgment — signed confirmation on hire and after updates
The single most effective policy provision is the approved-tools list paired with the prohibited-tools list. A policy that says "do not use unapproved AI tools for client work" is weaker than a policy that names ChatGPT Enterprise as approved and names consumer ChatGPT as prohibited. Specificity removes ambiguity and is enforceable.
Regulatory and Industry Considerations
Several industries now have specific AI governance requirements beyond general best practice. Small business owners in these sectors should adapt the policy to reflect applicable rules:
| Industry | Applicable Consideration |
|---|---|
| Healthcare and medical | HIPAA applies to any AI tool processing protected health information; Business Associate Agreements required |
| Financial services and accounting | Client data handled under professional ethics rules; SOC 2 and state-level CPA society guidance apply |
| Legal services | Client confidentiality and attorney-client privilege considerations; state bar opinions on AI tool use |
| Education | FERPA applies to student data; school-specific AI policies frequently layered above state rules |
| Regulated employment (HR) | EEOC has issued guidance on AI in hiring; bias-audit requirements in some jurisdictions |
A 30-Day AI Policy Adoption Plan
An AI policy can be drafted, reviewed, and adopted inside 30 days for most small businesses. The plan below is what I recommend to Florida small business owners.
Week 1 — Inventory
Survey every employee anonymously about which AI tools they currently use, for what tasks, and with what types of data. The inventory is almost always longer than expected.
Week 2 — Draft
Draft the policy using the ten-section template. Keep it under four pages. Name specific tools and tiers — the approved list and the prohibited list.
Week 3 — Review
Review with any legal counsel, the firm's accountant (for professional-services firms), and a representative from each functional team. Adjust based on feedback; the goal is a workable policy, not a perfect one.
Week 4 — Rollout
Distribute the policy, host a 30-minute all-hands walkthrough covering data classification and the approved/prohibited lists, and collect signed acknowledgments. Update the onboarding packet so new hires receive the policy on day one.
Ongoing — Review Annually
AI tools change quickly. A policy written in 2026 will need review in 2027. Schedule the review now so it does not slip.
Frequently Asked Questions
Yes — possibly more than a larger company needs one. A small team has less informal oversight than a larger one, and a single incident of accidental client data exposure can be materially damaging. The policy does not need to be long; it needs to be explicit about what tools are approved, what data goes where, and who reviews what. A two-page policy for a five-person team is appropriate and protective.
A full ban is usually unrealistic and counterproductive — employees will use tools that help them work faster whether they're banned or not, and a policy that drives use underground is worse than a policy that channels it. The better approach is to approve specific business-tier tools with appropriate data controls and explicitly prohibit consumer-tier versions of those same tools. Employees get the productivity; the business gets the data protections.
Start with SOC 2 Type II certification — this is the baseline for any tool processing business data. Beyond that, verify the training-data policy (is your input used to train models?), the data-retention policy (how long is input stored?), and the geographic storage location (which jurisdiction's law applies?). A reputable vendor publishes this information clearly; if you cannot find it, that is itself a signal.
The policy should name the consequence, but the first response to most incidents should be remediation rather than discipline. An employee who pasted a client record into a consumer LLM and reports it immediately enables a vendor-side deletion request and a client notification that may prevent material harm. An employee who hides the incident makes remediation impossible. The policy should encourage reporting with explicit non-retaliation language — and reserve discipline for willful or repeated violations.
Sources
- Gartner. Workplace AI Usage Survey, 2026 — Employee Adoption and Policy Coverage Rates. gartner.com
- International Association of Privacy Professionals. AI Governance Framework for Small and Medium Business, 2026. iapp.org
- American Institute of Certified Public Accountants. AI Tool Vendor Due Diligence Guide, 2026. aicpa.org
- Service Organization Control 2 (SOC 2) Certification Framework. American Institute of CPAs.
- U.S. Equal Employment Opportunity Commission. Technical Assistance Document — AI in Employment Decisions. eeoc.gov
- Health and Human Services Office for Civil Rights. HIPAA and Artificial Intelligence — Guidance, 2026. hhs.gov
Discussion
Comments
Leave a comment