How to write a one-page AI policy for a small business

One page, four rules. Which tools are approved, what never gets pasted in, who to ask, and what happens when something slips through.

Short answer: One page. Say which tools are approved, name the four categories that must never be pasted in, say who to ask when someone is unsure, and say what happens if something goes in by accident. Anything longer will not be read, and a policy nobody remembers is not a control.

Key takeaways

Why most small business AI policies fail

They are adapted from enterprise templates written for organisations with an IT function, a security team and an approval workflow. They run to eight pages, reference roles that do not exist in your business, and end up in a folder nobody opens.

Then someone pastes a client's bank details into ChatGPT at 4pm on a Thursday, because at that moment the policy is not in the room.

A single printed page on a desk beside a laptop

A small business policy has a different job. It is not a compliance artefact and it will never be audited by anyone. It is four things people can remember while working, written so a new starter can follow them on their second day.

Before you write: find out whether you are regulated

Worth ten minutes, because it changes the tone of the document rather than its length.

If you operate in New Zealand or the European Union, you are covered by privacy law regardless of size. There is no small-business exemption in either. Your policy should say so plainly, because "we have to" is a more durable motivator than "we should".

If you operate in Australia, you are exempt under $3 million annual turnover, but eleven categories of business are covered regardless, and the AML/CTF expansion on 1 July 2026 pulled in accountants, lawyers, conveyancers, real estate professionals and trust and company service providers even below that threshold. Do not assume; check the number and the categories.

If you operate in the United States, a firm of three to twenty people usually sits below the volume thresholds in most state privacy laws, but sector rules and the FTC Act still apply.

If you deploy AI in the EU, note also the AI Act's Article 4 on AI literacy, which from 2 February 2025 required providers and deployers to take measures to ensure "a sufficient level of AI literacy of their staff and other persons dealing with the operation and use of AI systems on their behalf". That article has since been amended by the Digital Omnibus on AI and the operative wording is in flux, so check its current status rather than quoting the original text. The direction of travel is that training your staff on the tools they use is an expectation, not a courtesy.

Our companion piece works through all four jurisdictions in detail.

The four rules

Four short handwritten notes pinned side by side

1. Which tools are approved. Name them. "ChatGPT, Claude and Microsoft Copilot are approved for general work" is more useful than a principle, because a principle requires a judgement call and a name does not. If people must choose their own, they will use whatever they already have logged in.

Say which account to use, too, and this is the rule that costs money rather than words. A business plan with terms you have actually read is a materially different position from a dozen personal free accounts you have never seen. If the business has not bought accounts, the policy is asking staff to carry a risk the business declined to pay to remove.

2. What never goes in. Be specific enough to act on at speed. A workable list for most firms:

Four categories, concrete enough to recognise in the moment. "Do not share confidential information" is not a rule anyone can apply, because everything at work feels a bit confidential, and a rule that flags everything teaches people to ignore it.

There is a reason to lead with credentials even though privacy law is mostly about the other three. A leaked API key is not a privacy incident, it is an unlocked door that works immediately for whoever finds it.

3. Who to ask. One named person. Not a role, a person. When someone is unsure at 4pm they need to know who to message, and "the data owner" is not a person anyone can message.

In a business of three to twenty people this is usually the owner, and that is fine. The point is that the question has somewhere to go, so the default answer stops being "probably fine, I'll just send it".

4. What happens when it goes wrong. State plainly that reporting an accidental paste is expected and will not be punished. If people fear consequences, you will find out about incidents months later or never.

This clause does more work than the other three combined, and it is the one most often cut. Some jurisdictions make it operationally necessary: in New Zealand, a privacy breach that has caused or is likely to cause serious harm must be reported to the Privacy Commissioner and the affected people as soon as you are practically able. You cannot meet a duty like that if your own staff will not tell you.

What to leave out

Long definitions of artificial intelligence. Legal recitals about legislation that may not apply to you. Check first, and if it does apply, cite it in one line rather than a page. Approval workflows requiring roles you do not have. And anything you will not actually enforce, because an unenforced rule teaches people that the rest are optional too.

Writing it

Half an hour, once. Ask which tools people already use. Genuinely ask, without consequences, or the answers will be wrong. Write the four rules. Send it around with a short note explaining why, not just what. Revisit in six months, or sooner if you start operating somewhere new.

A policy is not a control

This is the honest limit, and it is worth being direct about it in a document that exists to sell you something.

A policy tells people what to do. It does not do anything at the moment someone is working quickly and the chat box looks like any other text box. Every study of security training points the same way: effect decays within weeks, and it decays fastest under exactly the time pressure that produces the risky paste.

A person typing quickly at a keyboard, mid-task

That gap is a tooling problem rather than a documentation one. Prinn runs on each person's computer and checks a prompt before it is sent. When it finds something from those four categories it shows what it found and offers three choices: redact it, stop, or send it as is. The person decides. It never blocks anyone, and prompt content never leaves the machine, including to us.

An owner sees the category, the tool, the date, the device and the choice made, so you learn where the risk actually sits without reading anyone's prompts. That is also what tells you whether the policy is working, which is otherwise unknowable.

It is reliable on credentials, financial details and identity numbers. Names, home addresses and dates of birth written into ordinary prose are harder, and it is still improving at those.

Prinn is not legal advice and does not make you compliant with anything.

The one-page template

Copy this, replace the bracketed parts, delete nothing else. If it grows past one page, something needs to come out.

[Business name] AI use policy Last updated: [date]. Questions: [name], [email or phone].

Why this exists. We use AI tools because they make our work faster and better. This page exists so we can keep doing that without putting client information somewhere it should not be.

1. Approved tools. You may use [tool], [tool] and [tool] for work. Use the business account we provide, not a personal one. If you want to use something else, ask first. The answer is usually yes.

2. Never paste these into any AI tool.

If you need AI help with a document that contains these, remove or replace them first.

3. If you are not sure, ask [name]. No question is too small, and asking is never held against you.

4. If something goes in by accident, tell [name] the same day. You will not be in trouble. We need to know quickly so we can respond. For credentials that means rotating them immediately, and for client information there may be people we are required to notify. Telling us late is the only thing that causes a problem.

What we do on our side. We provide business accounts with terms we have reviewed, and [we run Prinn on work computers, which checks prompts before they are sent and warns you if something sensitive is about to go out / describe your own measure here].

Related

Sources