AI acceptable use policy: a one-page template for small teams

An AI acceptable use policy tells everyone who works for you which AI tools they may use, what data may go into them, what a person must check before AI output is used, and who to tell when something goes wrong. For a small team, one page is enough when it is paired with a short register recording how each approved tool is set up.

Write it after you know what people already do; the shadow AI post covers that discovery step. A policy alone changes little; an approved path that is easier than the workaround changes behavior. This article is informational and not legal advice; employment, privacy, and client-contract questions belong with counsel.

A small team's policy should answer ten questions:

• Who it covers. Employees, contractors, and temporary staff, on any device. NIST's Generative AI Profile (AI 600-1, July 2024) (opens in a new tab) suggests acceptable use policies cover third-party personnel such as contractors and consultants (GV-6.1-010).
• Approved tools and plans. AI 600-1 also suggests keeping approved lists of generative AI technology and service providers (GV-6.1-007). In a small business, that list is the register below.
• Data rules. What never goes into a prompt, and what may go only into specific tools.
• Human review. What a person must read before AI output reaches a client, a public page, or a decision about someone.
• Prohibited uses. The few things nobody does with AI here.
• Disclosure. When customers are told AI was involved.
• Requests. How to ask for a new tool or feature, and how fast you will answer.
• Mistakes. How to report pasted data or a bad output, and that prompt reporting is expected.
• Records. What approved tools log about each person's use, who can see it, and how long it is kept.
• Ownership. Who maintains the policy and when it is next reviewed.

Copy the text below and replace everything in square brackets. It is a starting point, not a finished legal document.

[COMPANY] AI ACCEPTABLE USE POLICY
Owner: [name, role]    Effective: [date]    Next review: [date]

1. Scope
This policy applies to anyone doing work for [Company], including
employees, contractors, and temporary staff, on any device or account.
"AI tool" means software that generates or summarizes text, images,
audio, or code, including AI features inside tools we already use.

2. Approved tools
Use only the tools and plans listed in the Approved AI Tools Register,
signed in with your [Company] account. Do not use personal AI accounts
for [Company] work. To request a new tool or feature, ask [owner] via
[channel]. We aim to answer within [number] business days.

3. Data rules
We sort data into three classes. The register says which classes each
approved tool may receive.
  General: public information, and our own documents that contain no
    client or personal details. Any approved tool.
  Client and personal: client names, project details, correspondence,
    and personal information about customers or staff. Only tools the
    register approves for this class.
  Never: the items below. Not in any AI tool, approved or not.
    - passwords, API keys, access codes, or security answers
    - full payment card or bank account numbers
    - Social Security, driver's license, or other government ID numbers
    - [health, legal, or other sensitive categories you handle]
    - anything a client contract says must stay in our systems

4. Human review
A person reads and takes responsibility for AI output before it is sent
to a client, published, used in a quote or invoice, or used in a
decision about a person. AI may draft. People decide.

5. Prohibited uses
Do not use AI to make or screen hiring, pay, or discipline decisions
unless counsel has reviewed the use and [owner] has approved it. Do
not use AI to impersonate a real person or to [add your own].

6. Disclosure
[When customers are told AI was involved, for example: chat and phone
assistants identify themselves as automated.]

7. Mistakes and incidents
If data went into an AI tool that is not approved for it, or AI output
caused or nearly caused a problem, tell [owner] the same day. Prompt
reporting is expected.

8. Records
[What approved tools record about use (prompts, outputs, who used
them), who can see those records, and how long they are kept.]

9. Acknowledgment
I have read this policy and will follow it.
Name: ____________  Signature: ____________  Date: ________

Section 3 lets the register approve a tool for a class of data rather than case by case. Section 8 is the most likely to need counsel, because it describes what the business can see about each person's use; the shadow AI post explains why that logging should be proportionate and disclosed. Section 7 is where an AI incident response plan starts, so name the same person and channel in both.

Section 5 keeps AI out of hiring, pay, and discipline decisions until counsel has reviewed the use and the owner has approved it, because some places regulate automated hiring tools directly. New York City's Local Law 144 (opens in a new tab), for example, bars employers and employment agencies from using an automated employment decision tool unless it has had a bias audit within one year of use, and requires a public audit summary and notices to candidates. Rules elsewhere differ and, as of October 2026, are still changing.

For each tool, record the plan your team actually uses, who administers it, and its training and retention settings as checked on a given date. It can also serve as the inventory of AI systems that NIST's voluntary AI Risk Management Framework (opens in a new tab) calls for (GOVERN 1.6); NIST notes that the framework is being revised (opens in a new tab) as of October 2026. The rows below are hypothetical and use generic tool types, not real vendor settings.

The Never class applies whatever the register says: a row approves a tool for some classes of data, never all of them.

The same brand can carry different terms depending on the plan and the account you sign in with. As of October 2026, the vendors' documentation says:

• Anthropic says it does not use inputs or outputs from its commercial products (opens in a new tab) to train models by default, with exceptions for submitted feedback and explicit opt-in. On its consumer plans (opens in a new tab), chats can be used for training when the user allows it in model improvement settings, among other listed cases such as safety review.
• Google's Workspace privacy hub (opens in a new tab) says Workspace content is not human reviewed or used for generative AI model training outside your domain without permission.
• Microsoft's Copilot privacy documentation (opens in a new tab) says that for work accounts, prompts, responses, and data accessed through Microsoft Graph are not used to train foundation large language models, and that admins can set retention policies. (Microsoft 365 Copilot is now called Microsoft Copilot.)

Terms change and settings differ by workspace, so check your own admin console, record the date, and recheck when a vendor announces a change. For other vendors, read the data-controls or privacy page before approving the tool; the AI vendor evaluation checklist lists the documents to request.

A workable sequence for a small team:

1. Set up the approved accounts and settings first, so nobody has to stop using a tool before its replacement exists.
2. Walk the team through the policy using examples from your own work: a client email, a quote, a meeting summary.
3. Collect signed acknowledgments.
4. After the first month, ask what people still do outside the approved tools and why. Treat the answers as requests, not confessions.
5. Update the register and repeat.

The misuse-simulation exercise in how to red-team an AI system is a structured way to run step 4: it tests the approved path the way busy people will use it.

In this hypothetical, an assistant at a nine-person bookkeeping firm pastes a client's payroll export into a personal chat account to reformat the columns. She tells the owner that afternoon, as section 7 asks.

• Record. Which tool, which plan, what data, and when.
• Check the settings. On a consumer plan, training and retention depend on her account settings. She checks them, deletes the conversation, and confirms in the vendor's documentation what deletion actually removes.
• Check obligations. Payroll data is personal information. The owner checks the client contract and calls counsel about any notice duty before deciding what to tell the client.
• Fix the cause. The approved assistant could do the same job, but nobody had shown her. The owner adds "reformat spreadsheets" to the approved uses and shows the team how.

The useful outcome is a faster approved path and a policy that rewarded the early report.

This template does not settle employment questions (monitoring, discipline, and how acknowledgments are used), sector rules for health or financial data, state privacy and AI laws that are still changing, or the confidentiality and subprocessor terms in your client contracts. Have counsel check the final version against each.

A policy also does not decide which AI projects are worth doing; AI use case prioritization scores candidate tasks for that. A workflow customers depend on needs acceptance criteria and a named owner, as why AI pilots stall explains.

For help evaluating the tools in your register, Cleland & Co.'s AI Strategic Advisor engagement covers vendor evaluation and written governance criteria your team can apply, alongside the rest of an AI roadmap.