Applied AIUpdated 9 min read
AI use policy for small businesses: a one-page template

In this article
An illustrative example. A sales rep pastes the customer list, balances and all, into a free AI chat and asks it to rank who to call about payment first. The answer is good and saves her an hour. Nobody at the company knows she did it, which account that file ended up in or whether the vendor can use it to train its model.
She didn't do anything wrong on purpose. Nobody told her what was allowed and what wasn't. That's what an AI use policy is for, and for a company of 10 to 50 people it fits on one page.
Your team already uses AI, with or without permission
In Microsoft and LinkedIn's survey of 31,000 workers in 31 countries, 78% of people who use AI at work bring their own tools, and at small and medium-sized companies that number reaches 80%. On top of that, 52% of AI users are reluctant to admit using it for their most important tasks (Microsoft and LinkedIn, 2024).
Companies are behind. In ISACA's survey of more than 3,400 audit, risk and security professionals, 90% believe employees at their organization already use AI, but only 38% have a formal, comprehensive policy, and 25% have none at all (ISACA, 2026).
The case everyone cites is Samsung. In 2023, engineers there uploaded internal source code to ChatGPT, and Samsung responded by banning generative AI on company devices and networks (Bloomberg, 2023). Banning it after the leak is the most common reaction, and it doesn't always work: if the company doesn't offer an approved tool, people keep using their own, just quietly.
The wrong-inbox test
Almost all the risk is in what gets pasted. So the policy is built around one question anyone can ask in three seconds, without knowing anything about technology. We call it the wrong-inbox test: before you paste something into an AI tool, picture that text landing by mistake in the inbox of someone outside the company. What happens?
- Nothing: it's public information. It can go into any approved tool.
- It would be awkward, but not serious: it's internal. Only in the approved tool, on the company account.
- It would hurt a customer, an employee or the company: it's confidential. It doesn't go into an AI chat.
| Level | Examples | Where it can be used |
|---|---|---|
| Public | Published catalog, website copy, FAQs, job postings | Any approved tool |
| Internal | Processes, draft proposals, monthly sales totals, product lists | Only the approved tool, on a company account |
| Confidential | Social Security numbers and IDs, health information, bank accounts, passwords, payroll, contracts with confidentiality clauses, customers with names and balances | No AI chat. Only systems the company connects and approves in writing |
The test doesn't replace the table, but it makes it usable. Nobody memorizes a list of 20 data types; everyone can picture an email sent to the wrong person.
The template: eight clauses on one page
Copy the text, fill in the brackets and cut whatever doesn't apply. If it runs past one page, nobody reads it.
| Clause | Model text |
|---|---|
| 1. Scope | This policy covers how we use AI tools at [Company]. It applies to employees, contractors and interns, on company or personal devices, whenever they work with company information. |
| 2. Approved tools | We use [tool] on the company account. To use a different one, ask [owner], who will answer within [5] business days. Personal accounts aren't used for work. |
| 3. Data | Before you paste anything, run the wrong-inbox test. Public data: yes. Internal: only in the approved tool. Confidential: never in an AI chat. |
| 4. Human review | Anything that goes to a customer, vendor or agency with a price, deadline, figure, condition or legal detail is reviewed by a person before it's sent. Whoever sends it owns it. |
| 5. Transparency | If a customer asks whether we use AI, we tell the truth. Assistants that talk to customers identify themselves as automated. |
| 6. What we don't hand off | AI doesn't decide on hiring or firing, approving credit, setting special pricing or answering formal complaints. It can help prepare; a person decides. |
| 7. Incidents | If you pasted something you shouldn't have, or AI produced something wrong that already went out, tell [owner] the same day. Reporting it promptly is never penalized; hiding it is. |
| 8. Review | We review this policy every [6] months. Version [1], dated [date]. Owner: [name]. |
Why human review is the clause that matters most
A language model can write something false with complete confidence. In 2023, a federal judge in New York fined two lawyers and their firm $5,000 for filing a brief that cited six court decisions ChatGPT had made up (Mata v. Avianca, 2023). Nobody checked them before signing. The AI guide for law firms covers what else you shouldn't hand off.
People who use AI every day already know this: 86% treat what it produces as a starting point, not a final answer (Microsoft, 2026). Clause 4 just puts that habit in writing and makes clear who's accountable.
Why the incident clause doesn't punish
If reporting a mistake gets you in trouble, nobody reports, and you find out when the customer complains. Data pasted in the wrong place can be contained if you know the same day: delete the conversation, change a password, let the customer know. A week later, you can't.
Why just one tool
One approved tool on a company account is easier to manage than five: you know who has access, you can remove it when someone leaves and you review a single contract. If the team needs another one, clause 2 says how to ask, which keeps everyone from picking their own.
Checklist: 10 things to check before picking an AI tool
Before you approve a tool, go through its terms with this checklist. Most of it is on the vendor's privacy or security page; if you can't find something, ask in writing.
- It has a business plan, with accounts the company manages instead of each person.
- It states in writing whether it uses your data to train its models and whether that's off by default. OpenAI, for example, says it doesn't train on business data from its business plans by default (OpenAI, 2026).
- It says where your data is stored and for how long.
- You can delete conversations and files, and it says how long until they're actually gone.
- You can add or remove a user in minutes, for when someone joins or leaves.
- It supports signing in with the company account and two-factor authentication.
- It offers a data processing agreement you can sign or accept.
- It lets you export your information if you switch vendors someday.
- If it connects to email or files, you choose which folders and with what permissions.
- It has clear per-user pricing, and you know what you'd pay if the whole team used it.
If a tool fails the first two items, don't approve it for internal data, no matter how good it is.
How to roll it out in a week
- Ask what people use todayA three-question survey, no penalties: which tool you use, what for and what kind of data you give it. With 52% reluctant to admit it (Microsoft and LinkedIn, 2024), the no-penalty promise is what gets you honest answers.
- Pick one toolUsing the checklist, choose the one that covers the most common uses from the survey. Just one, with a company account for whoever needs it. If you don't know which tasks yet, start from the 30 AI tasks by department.
- Fill in the templateFill in the brackets, name an owner and adjust the data table with examples from your business: your documents, your customers, your files.
- Walk the team through it in 30 minutesOne meeting with three real cases from the survey: what's fine, what isn't and why. Run the wrong-inbox test out loud on each one.
- Send it and get sign-offHave each person confirm they've read it. Put the first review date on the calendar.
When this template isn't enough
- You handle health information, data about minors or regulated financial data. One page won't cover what privacy law requires where you operate: have a lawyer review the data clause.
- You use AI to make decisions about people, like screening résumés or approving credit. That calls for bias testing and records this template doesn't include. NIST's voluntary AI Risk Management Framework is a good place to start (NIST, 2023).
- You sell a product or service that uses AI with your customers. What you promise and how you answer belongs in your contracts, not just an internal policy.
- A large customer asks you to follow their own AI rules. Their contract overrides your template.
- You're a team of two or three. You don't need a document: agree on the tool, the wrong-inbox test and review before sending, and pin it in the team chat.
This template isn't legal advice. It's a starting point so your team works with clear rules while a lawyer reviews what your industry and your state or country require.
From policy to systems
A policy tells people what not to paste. But the risk exists because people have to paste: the data they need for AI to help them lives in spreadsheets, emails and chats, and the only way to hand it over is to copy it.
The next step is having sales, inventory and customer information in one system with per-person permissions. Then, when you want to use AI on that data, you connect it to that system with controlled access, and nobody has to paste the receivables list into a chat. That's an AI integration, and AI training helps your team use it within these rules.
Sources
- Microsoft and LinkedIn, "2024 Work Trend Index: AI at Work Is Here. Now Comes the Hard Part," 2024 (opens in a new tab)
- ISACA, "AI Use Accelerates, While Governance and ROI Lag, Says New ISACA Research" (2026 AI Pulse Poll), 2026 (opens in a new tab)
- Mark Gurman, "Samsung Bans Staff's AI Use After Spotting ChatGPT Data Leak," Bloomberg, 2023 (opens in a new tab)
- Mata v. Avianca, Inc., U.S. District Court for the Southern District of New York, opinion and order on sanctions, 2023 (opens in a new tab)
- Microsoft, "2026 Work Trend Index," 2026 (opens in a new tab)
- OpenAI, "Enterprise privacy at OpenAI," 2026 (opens in a new tab)
- NIST, "AI Risk Management Framework" (AI RMF 1.0), 2023 (opens in a new tab)