AI Education
How to Write a One-Page AI Policy for Your Business (With a Template)
Most small businesses use AI and most have no rules for it. A policy you can write in an hour, with the template and the reason behind every line.
A one-page AI policy tells everyone in your business which AI tools they may use, on which accounts, what must never be pasted into them, who checks the output before it goes to a customer, what an AI agent may do without a person's approval, and what to do when something goes wrong. For a business with fewer than 50 people it takes about an hour to write, and the template in this article is ready to copy. Most small businesses don't have one, and the court rulings and data settings of 2026 make that a bigger gap than it used to be.
One note before you read on: I'm not a lawyer. I spent 15 years in corporate risk and security, and this is general information. For a specific legal matter, talk to your lawyer. The data settings and legal points below were checked on their original pages on October 27, 2026.
Why a one-page policy, and why now?
Because the tools arrived before the rules did. Two-thirds of US small businesses now use AI, according to the US Chamber of Commerce's 2026 report (US Chamber of Commerce, October 2026). Among the 1,256 owners Goldman Sachs surveyed in its 10,000 Small Businesses program, 76% use AI, and the hurdle they named most often was data privacy and security, at 50% (Goldman Sachs, March 2026). That's one program's members, but the worry matches what I hear from owners: everyone uses it, nobody has written down how.
The legal profession shows what the gap looks like with numbers. In the 8am 2026 Legal Industry Report, 69% of legal professionals use general AI tools for work, only 9% said their firm has an AI policy that is actively enforced, and 43% have no policy and no plans for one (8am, March 2026). If lawyers haven't written the rules, most dental practices and web shops haven't either.
Three things changed in 2026 that make the policy worth an hour of your time:
- Chats aren't confidential. In February 2026, a federal judge in New York ruled that a defendant's conversations with an AI chatbot weren't protected by attorney-client privilege, partly because the provider's privacy policy allowed the chats to be disclosed (Harvard Law Review, March 2026). What your staff paste into an AI tool can end up in a dispute.
- The data settings differ by plan. On the consumer plans of ChatGPT, Claude and Gemini, your chats can be used to train the models unless you change a setting. The business plans mostly don't do that by default. Which plan a person uses is a policy decision, not a personal one.
- AI agents act now. Tools that read your inbox, edit your Business Profile or browse for you have arrived for small businesses this year, and they need a rule about what they may do without a person saying yes.
I'd add a fourth, which is older: the Samsung case. In 2023 Samsung restricted generative AI on company computers after staff uploaded sensitive code to ChatGPT (CNBC, May 2023). A one-page rule in 2022 would have been cheaper than the ban.
What goes into a one-page AI policy?
Seven sections, each one a decision you make once so your team doesn't have to make it every day.
| Section | What it decides | Why it's there |
|---|---|---|
| 1. Approved tools and accounts | Which tools, on which plan, logged in how | The business plans don't train on your chats by default; personal accounts may |
| 2. The never list | What goes into no AI tool, ever | Chats can be subpoenaed, reviewed by vendor staff or used for training |
| 3. The redact-first list | What may go in after names and numbers are replaced | Most useful work needs real figures, not real identities |
| 4. Checking | Who checks names, numbers, facts and tone before anything leaves | AI tools state wrong things with confidence |
| 5. Agents | What an AI may do without a person's approval | Sending, paying, deleting and publishing can't be undone |
| 6. After a mistake | What to do, who to tell, no blame | People hide mistakes they'll be punished for |
| 7. Legal hold and records | When nobody deletes anything, and what you keep | A likely dispute changes the rules; the EU asks for a record of training |
The order matters less than the length. A policy that fits on one page gets read. A twelve-page one gets signed and forgotten.
The template
Copy this into a document, replace the brackets, and delete anything that doesn't apply. It's written for a business with one owner and a handful of staff or freelancers.
AI USE POLICY, [BUSINESS NAME]
Version 1, [DATE]. Owner: [NAME]. Review date: [DATE + 6 MONTHS].
1. Approved tools and accounts
We use [ChatGPT Business / Claude Team / Gemini in our Google Workspace / other] for work.
Work happens in the company account only, never in a personal account.
Training is switched off on every account used for work (see the settings sheet).
Anyone who wants to use another tool for work asks [NAME] first.
2. The never list
These never go into any AI tool, on any plan:
- customer names together with contact, payment or account details
- passwords, login codes, API keys
- contracts, NDAs, offers and anything a client gave us in confidence
- staff records, pay, health information, HR matters
- patient, legal or financial records we hold for other people
- anything our lawyer has told us is privileged
3. The redact-first list
These may go in after names, account numbers and exact figures are replaced
with placeholders like Client A, Supplier B and rounded numbers:
- financial figures and forecasts
- supplier terms and prices
- internal disputes and complaints
- draft contracts, with the parties and amounts removed
4. Checking
Everything an AI tool writes is a first draft.
Before it goes to a customer, a supplier or the public, [NAME / the person sending it]
checks every name, number, date, price and fact against the source.
AI text we publish to inform the public is reviewed by [NAME], who takes
editorial responsibility for it. We say so where our customers would expect it.
Images, voices or videos that could be mistaken for real people or events are labelled as AI-generated.
5. AI agents and connected apps
An AI tool may read, sort, summarise and draft without asking.
It never sends, publishes, deletes, pays, signs or changes customer-facing details
without a named person approving that specific action.
We connect an AI tool to our email, files, accounts or Business Profile only after
[NAME] has read what the tool does with that data and written it on the settings sheet.
If an agent does something we didn't ask for, we disconnect it the same day.
6. After a mistake
If something confidential went into an AI tool, or an AI tool did something wrong,
tell [NAME] the same day. We fix it first and talk about it second. Nobody is in trouble
for reporting a mistake. We keep a one-line log of what happened and what we changed.
7. Legal hold and records
If [NAME] says a legal matter is likely, nobody deletes AI chats, files or logs about it
until told otherwise. We keep this policy, the settings sheet and the date each person
read it as our record of AI training.
8. Payments and identity
No AI, no caller and no email changes a bank account, a payment or a login on its own.
We confirm payment changes by phone on a number we already have.
I have read this policy. Name: ____________ Date: ________The settings sheet it refers to is a simple table: tool, account, who owns it, training setting, date checked. The next section tells you what to put in it.
The three settings to check before you publish it
Open each account your team uses for work and check one setting. Ten minutes for all three.
| Tool | The setting | What it means | Source |
|---|---|---|---|
| ChatGPT Free, Plus, Pro | "Improve the model for everyone" under Settings, Data controls. On by default. | Switch it off, or the chats may train OpenAI's models. Business, Enterprise, Edu and the API aren't used for training by default. | OpenAI Help |
| Claude Free, Pro, Max | The training choice in Privacy Settings | Say no, or chats can be used for training and kept for five years. Team, Enterprise and the API fall under separate commercial terms. | Anthropic, August 2025 |
| Gemini (personal accounts) | Keep Activity | On means Google can use your chats and uploads to improve its services, including training, with some human review; reviewed chats are kept up to three years. Off disables Business Profile, notebooks and most connected apps. | Google Gemini Apps Privacy Hub, October 2026 |
OpenAI says it doesn't use content from ChatGPT Business, Enterprise, Edu or Healthcare workspaces to train its models by default, and the same goes for the API (OpenAI Help). ChatGPT Business costs $25 per user a month billed monthly, or $20 billed annually (OpenAI Help). Anthropic asks users of Claude Free, Pro and Max to choose whether their chats can be used for training, keeps them for five years if they say yes, and excludes Team, Enterprise and the API from that choice (Anthropic, August 2025). Google's Privacy Hub says that with Keep Activity on, your activity is used to improve its services, including training generative AI models, a subset is read by human reviewers, and reviewed chats are kept for up to three years (Google, Gemini Apps Privacy Hub, updated October 2026).
In my view this is the decision that matters most in the whole policy: the cheapest privacy upgrade a small team can make is moving work from personal accounts to a business plan. I went through the court cases behind that in is ChatGPT private?, and the Gemini conditions in Gemini's new business hub.
What does the law expect of a small business?
Less than the headlines suggest, and more than nothing. Two EU rules touch a small business directly, and one Swiss one.
AI literacy. Article 4 of the EU AI Act asks providers and deployers of AI systems to take measures to support the AI literacy of their staff. It has applied since 2 February 2025, and the July 2026 amendment kept the obligation while dropping any required "sufficient" level (Regulation (EU) 2026/1744). The European Commission's own Q&A says a company whose employees use ChatGPT to write advertising text or translate documents is in scope, and that those employees should be informed about the specific risks, such as hallucination. No certificate is needed, no specific governance structure is mandated, and organisations can keep an internal record of their training or guidance (European Commission AI Literacy Q&A, updated October 20, 2026). The Commission also says the rule applies to companies outside the EU when the AI system is used in the EU or affects people there. Section 7 of the template, with the dates people read it, is the record.
Labelling. Since 2 August 2026, Article 50 of the AI Act requires deployers who publish AI-generated or manipulated text to inform the public on matters of public interest to disclose that, unless a person with editorial responsibility has substantively reviewed or controlled the content. Deepfakes must be disclosed as artificially generated or manipulated, and anyone building a chatbot has to tell users they're dealing with AI. The Commission adopted guidelines on 20 July 2026 and published a voluntary Code of Practice with an icon set (Cooley, August 2026). For most small businesses the practical reading is section 4 of the template: a named person reviews what you publish, and anything that could pass for a real photo or recording is labelled. The disclosure line at the end of every article on this site is my own version of that.
Switzerland. The Federal Data Protection and Information Commissioner has said the revised data protection law (nFADP) applies directly to AI: people have a right to know when they're talking to a machine, the purpose of AI-based processing has to be transparent, and high-risk processing needs an impact assessment (FDPIC, updated May 2025). If you serve EU customers, the GDPR applies as well.
This is a plain-English summary, not legal advice. If you're in a regulated field, send the finished policy to your lawyer before you roll it out.
What should an AI agent never do without approval?
Anything you can't undo and anything that speaks for your business. Section 5 of the template draws the line at reading, sorting, summarising and drafting on one side, and sending, publishing, deleting, paying, signing and changing customer-facing details on the other.
The reason is a risk that didn't exist when AI was a chat window. Once a tool reads your emails, web pages or files, instructions hidden in that content can steer it. The security organisation OWASP lists prompt injection as the number one risk in its 2025 Top 10 for AI applications (OWASP), and Google's help page for its own agent warns that a website, email or document can contain hidden instructions that mislead it (Google Help). Google's agent asks for confirmation before sending, deleting, buying, paying or submitting forms. Your policy should ask for the same from every agent, whoever makes it. I used that rule for Tasks in Gemini and it hasn't needed changing since.
Section 8 belongs next to it. The same week your team learns that an AI can draft a convincing email, they should learn that so can a criminal. A payment change confirmed by a call to a number you already have stops most of the fraud that AI has made cheaper to attempt.
How to write it in an hour
- Ten minutes: list the tools. Who uses what, on which account, for which jobs. The Quick Win table.
- Ten minutes: fix the settings. Switch off training on every personal account still used for work, and decide whether the work moves to a business plan.
- Fifteen minutes: fill in the template. Keep the never list short enough to remember. Add your own items, for example patient records for a clinic or listing data under contract for an estate agent.
- Ten minutes: name the people. One owner for the policy, one checker for anything published, one person to tell after a mistake. In a three-person business that's the same name three times, and that's fine.
- Ten minutes: tell the team. A short meeting or a two-minute video. Each person reads the page and signs or replies "read". Put the date in the settings sheet.
- Five minutes: put the review in your calendar. Six months. The tools will have changed by then.
If the decisions in the policy are the hard part, the AI Prompts for Business Decisions pack has 28 prompts for risk assessment and structured choices, including a risk check you can run on the policy itself.
How to keep it alive
A policy that nobody looks at after the signing is a document, not a rule. Three habits keep it working:
- Log mistakes without blame. The one-line log in section 6 is the most useful page you'll have in a year: it tells you which rule people actually break and which tool causes it.
- Re-check the settings quarterly. Vendors change their defaults. Google's agent was renamed within five months of launch and ChatGPT's Team plan became Business; the training switches move too.
- Review the page every six months. Add the tool everyone started using, remove the one nobody opens, and update the never list when you take on a new kind of client.
If you want the method behind the checking rule, my guide to how lawyers are using AI shows what happens when nobody checks: more than 2,000 court cases with fake AI citations. For the EU literacy duty in more depth, see why learning AI matters in 2026.
Frequently asked questions
Does a small business need an AI policy?
If anyone uses AI for work, yes. It takes about an hour, it decides which accounts are used and what never goes into them, and in the EU it doubles as the record of AI literacy that Article 4 of the AI Act asks for. Goldman Sachs found data privacy and security was the top AI concern of small business owners, at 50%.
What should an AI usage policy include?
Seven things: approved tools and accounts, a never list, a redact-first list, who checks AI output before it leaves the business, what an AI agent may do without approval, what to do after a mistake, and the legal hold. A line on confirming payment changes by phone is worth adding.
Is it safe to use ChatGPT at work?
On a business plan with training switched off and a short never list, for most everyday work, yes. On a personal Free, Plus or Pro account, your chats may be used for training unless you switch off "Improve the model for everyone", and a 2026 court ruling found AI chats aren't privileged.
Do I have to train my staff on AI under the EU AI Act?
Article 4 asks providers and deployers to take measures to support their staff's AI literacy. The Commission says a company whose staff use ChatGPT for advertising text is in scope, that no certificate or specific level is required, and that an internal record of your training or guidance is enough. It applies to non-EU companies whose AI use affects people in the EU.
Do I have to label AI-generated content?
In the EU, since 2 August 2026, if you publish AI-generated text to inform the public on matters of public interest and nobody with editorial responsibility has reviewed it. Deepfakes must always be disclosed, and chatbots must tell people they're talking to AI. Ordinary business communications that a person has reviewed are not caught by the text rule.
What should employees never put into AI tools?
Customer personal and payment details, passwords and keys, contracts and anything given in confidence, staff and health records, records you hold for other people, and anything your lawyer has called privileged. Financial figures and supplier terms can go in once names and exact numbers are replaced.
How often should I update my AI policy?
Every six months, with a quarterly check of the training settings. Vendors change defaults and product names often, so the settings sheet goes stale faster than the policy.
Where to go from here
Run the three-question Quick Win today, then block an hour this week for the template. Send the finished page to your lawyer if you work in a regulated field, and put the six-month review in your calendar before you forget. If you'd like to see where your business stands more broadly, the free AI Readiness Quiz takes about 5 minutes.
If you'd like one short, practical note on AI and risk a week, The Risk Memo goes out every Tuesday.
Researched with AI tools, written and fact-checked by Dominic.
Sources
- US Chamber of Commerce, Empowering Small Business 2026 (Oct 2, 2026)
- Goldman Sachs 10,000 Small Businesses Voices, AI Presents a Major Opportunity for Small Businesses, But Support Is Needed to Close the Implementation Gap (Mar 17, 2026)
- 8am, Generative AI data from the 8am 2026 Legal Industry Report (Mar 20, 2026)
- Harvard Law Review Blog, United States v. Heppner (Mar 23, 2026)
- CNBC, Samsung bans use of AI like ChatGPT for staff after misuse of chatbot (May 2, 2023)
- OpenAI Help Center, Data controls in ChatGPT
- OpenAI Help Center, ChatGPT Business overview
- Anthropic, Updates to our consumer terms (Aug 28, 2025)
- Google, Gemini Apps Privacy Hub (updated Oct 23, 2026)
- European Commission, AI Literacy - Questions & Answers (updated Oct 20, 2026)
- Regulation (EU) 2026/1744, Digital Omnibus on AI
- Cooley, EU AI Act: Transparency Obligations Take Effect 2 August 2026 (Aug 3, 2026)
- FDPIC (EDÖB), Update: Current data protection legislation is directly applicable to AI (May 8, 2025)
- OWASP GenAI Security Project, LLM01:2025 Prompt Injection
- Google Gemini Apps Help, Manage your tasks & workflows in Gemini Apps (accessed Oct 27, 2026)