What is an AI policy, and what goes in one?

If anyone in your business has pasted a customer email into a chatbot to draft a reply, you already have an AI practice. What you may not have is a rule about it. An AI policy is that rule, written down.

The short answer

An AI policy is a short written document that tells everyone in the business how AI tools may be used: what for, with what information, with whose approval, and with what checking. It is internal, it is written for staff rather than lawyers, and for a small business it should fit on one or two pages.

The National AI Centre, the federal body behind Australia’s Guidance for AI Adoption, treats the policy as the reference point for every AI activity in an organisation. Its reasoning is plain: without shared rules, teams use AI inconsistently, and inconsistency is where the avoidable risk comes from.

Why a small business needs one

Three reasons, none of them theoretical.

The tools are already in the building. The Australian Bureau of Statistics found around 12 percent of Australian businesses used AI in 2024-25, and around 11 percent of small and micro businesses. That counts deliberate, reported use. The National AI Centre’s register guidance separately names “shadow” AI use, meaning tools staff use that nobody approved or recorded, as one of the things good documentation is meant to reduce. A policy is how undeclared use becomes declared use.

The failure cases are ordinary. The Australian Cyber Security Centre’s guidance for small business describes two. In early 2025 a contractor working for an Australian organisation uploaded personal information, including names, contact details and health records, into an AI system. It became a notifiable data breach. In the same year a lawyer filed a court document containing false cases that an AI tool had generated and nobody had verified, and was barred from operating a law practice. Neither needed an attacker. Each is covered by one line of policy: what information may go in, and who checks what comes out.

Two regulators point at the same document. The privacy regulator, the Office of the Australian Information Commissioner, says the Privacy Act applies to personal information put into an AI system and to any personal information in what comes back, and that businesses should establish policies and procedures for their use of AI. The Australian Cyber Security Centre recommends an internal AI use policy that clearly defines what data cannot be uploaded. Neither is asking for a programme. Both are asking for something written.

The six things an AI policy has to cover

The National AI Centre’s guide closes with six points to confirm before a policy is published. They make the right skeleton. Each is set out below with the decision behind it and an example of wording at small business scale. The examples are illustrations to adapt, not legal advice.

1. What AI can and cannot be used for

The decision is where ordinary use ends. Most small businesses land on three tiers: fine without asking (drafting, summarising, tidying internal notes), needs approval first (anything a customer sees unreviewed, anything involving personal information, anything that decides something about a person), and not allowed at all.

Example: “Staff may use approved AI tools to draft, summarise and analyse. AI must not make the final decision about a customer, a job applicant or an employee.”

2. Who approves the higher-risk uses

One named person, with the authority to say no. A committee is too slow for a twelve-person business, and “check with management” is not a name.

Example: “New AI tools, and any use in the approval tier, are signed off by [name] before they start.”

3. What data staff can put into tools

This is the clause that prevents the breach above. The OAIC recommends, as best practice, that organisations do not enter personal information, and particularly sensitive information, into publicly available generative AI tools. The Australian Cyber Security Centre adds that some providers can use submitted data to train their models depending on settings and subscription type, and recommends removing or anonymising personal details before using an AI application.

Example: “Do not enter personal, confidential or financial information into any AI tool that is not on the approved list. Remove names and identifying details first wherever you can.”

4. When a person must check the output

AI tools produce confident errors. The Australian Cyber Security Centre’s advice is to train staff to verify outputs and to keep a person in the decision for anything high-stakes or sensitive.

Example: “Anything produced with AI is checked by the person sending it before it reaches a customer, a supplier or a regulator. You are responsible for what goes out under your name.”

5. How staff report a problem

Keep it short: who to tell, and how quickly. The purpose is to hear about a mistake while it is still small.

Example: “If an AI tool produces something wrong or harmful, or information goes somewhere it should not have, tell [name] the same day. Reporting early will be treated as doing the right thing.”

6. When the policy gets reviewed

Tools change faster than documents. Put a date on it.

Example: “This policy is reviewed every six months, and whenever a new AI tool is approved.”

A one-page outline you can copy

Eight headings are enough for most businesses under about fifty people:

1. Purpose and scope. Why the policy exists, and that it covers staff and contractors.
2. Approved tools. A pointer to the list, kept separately so the policy does not go stale.
3. Permitted, approval-needed and prohibited uses. The three tiers.
4. Data rules. What must never go into a tool, and what may go into the approved ones.
5. Checking output. Who is responsible for what leaves the business.
6. Customer-facing AI. If customers deal with a chatbot or an automated reply, they are told so. The OAIC expects public-facing AI tools to be clearly identified.
7. Reporting problems. Who to tell, and by when.
8. Owner and review date. One name, one date.

The National AI Centre’s free template covers the same ground in more depth: core principles, expected behaviours and internal rules, roles and responsibilities, governance and approval points, and monitoring and review. Start from it and cut. Its own advice on adapting the template is worth following: tailor it to how the business already works, use terms the team already knows, and test the draft with the people who will have to use it.

Where it sits next to the register and the privacy policy

Three documents get confused, and they do different jobs.

• The AI policy is the internal rulebook described above.
• The AI register is the list of AI systems the business actually uses, including the AI features embedded in software bought for something else. The National AI Centre publishes register templates as a Word document and a spreadsheet. The policy says “approved tools only”. The register is the list.
• The privacy policy is public and is about personal information. The OAIC says businesses should update their privacy policies and notifications with clear information about their use of AI. There is also a dated obligation: from 10 December 2026, entities covered by the Privacy Act that use personal information in automated decisions with the potential to affect a person’s rights or interests must describe that in their privacy policy.

Businesses turning over $3 million or less usually sit outside the Privacy Act, with exceptions that include health service providers and businesses that trade in personal information. Being under the threshold changes what is mandatory. It does not change what a customer expects you to have done with their details.

A policy, a register, a named owner and a review date together are the whole of AI governance at small business scale.

The mistakes that make a policy useless

• Banning everything. A ban with no approved alternative gives staff a reason to hide what they are doing, which is the opposite of what the document is for.
• Copying an enterprise policy. Twelve pages written for a bank will not be read in a twelve-person firm. Length is not rigour.
• Naming tools instead of behaviours. “Do not use tool X” is out of date within a quarter. Put behaviours in the policy and tool names in the register.
• No owner. A policy nobody is accountable for is a file, not a control.
• Writing it once. If the review date passes without anyone opening the document, it has stopped describing the business.

How to tell whether it is working

Three checks, answered from records rather than memory:

1. Can someone who joined last month tell you the data rule without looking it up?
2. Does the register match what is actually in use, including the tools people signed up for on their own?
3. Has anyone asked for an approval, or reported a problem, since the policy was written?

If the answer to the third is no after six months, either nothing risky has been attempted or nobody thinks the policy applies to them. It is worth finding out which.

When outside help is worth it

A first AI policy does not need a consultant. The government template is free, the six points above are the whole checklist, and most owners can produce a usable version in an afternoon. If you have not yet settled what AI is for in the business, start with AI for small business and write the policy before the first pilot.

Help earns its place in three situations: when personal information is involved and a client is asking you to evidence your controls, when AI is already built into how the business runs and the policy has to cover AI automation rather than staff typing into a chat window, and when the team needs training alongside the rules so the policy is understood rather than signed.

If that is where you are, our AI foundations work covers team training, use case ranking and tool selection, which is the groundwork a policy has to be written around. The AI agency page sets out how that fits with strategy, automation and ongoing operations as one practice. For a view of where AI would pay in your business specifically, the free AI use-case audit ranks the candidates in plain English, with no obligation attached.