Is an AI use policy legally required in Australia?

No law says a small business must have one. The obligations that make one worth having already exist though: the Privacy Act for covered businesses, confidentiality clauses in client contracts, and professional duties in regulated fields all apply to what staff type into an AI tool. A written policy is how a business shows it took those obligations seriously, and it is increasingly requested in supplier questionnaires and cyber insurance proposals.

Does the Privacy Act apply to my small business?

Generally not if annual turnover is $3 million or less, but the exceptions matter: private health service providers, businesses that buy or sell personal information, and a few other categories are covered regardless of turnover, and any business can opt in. Even under the threshold, client contracts and commercial expectations usually demand the same data handling, which is why the template keeps its privacy clause either way.

Can we just ban AI tools instead?

You can write that policy, but it rarely reflects reality. Staff who find AI useful will keep using it on personal devices where the business has no visibility, which is the worst possible arrangement: all of the data risk, none of the productivity, no audit trail. Naming one sanctioned tool with real data protections and banning the rest is both more enforceable and more honest.

What is wrong with staff using free ChatGPT accounts for work?

Free consumer AI accounts sit entirely outside the business: the business cannot see what was entered, cannot produce records later, and depending on the account's settings the input may be used to improve the service. A work-account tool such as Microsoft 365 Copilot keeps prompts and responses inside the business's own environment and under its existing controls. The distinction is the account and its terms, not the underlying model.

Do we need client consent to use AI on their information?

It depends on what the tool does with the data and what the client was told at collection. Using an in-tenant tool where the data never leaves your environment is a different proposition to entering client details into a consumer tool that stores them elsewhere. Where the Privacy Act applies, use has to line up with the purpose of collection or reasonable expectations, and the OAIC has published specific guidance on commercial AI products. When in doubt, minimise the data or ask.

How often should the policy be reviewed?

Every six months is the template default, and sooner on two triggers: adopting a new AI tool, or a change in the law. AI capability and Australian privacy reform are both moving quickly enough that an annual-only review will drift out of date. The review itself is short if nothing changed; the point is that a named owner looks.

A free AI use policy template for Australian small business

In most small businesses the AI policy conversation happens in the wrong order. Staff started using ChatGPT, Gemini or Copilot months ago, on personal accounts, on their own phones, with whatever text was handy. Some of that text was client names, quotes, contract clauses and complaint emails. The policy question is not whether to allow AI. It is already in the building. The question is whether anyone has written down the rules.

Below is a complete AI use policy for an Australian small business. It is published in full on this page, no email address required, no download gate. Copy it, replace the bracketed placeholders, have a director sign it, and you have a defensible policy this week. The sections after the policy explain how to adapt it and why naming one sanctioned tool makes the whole thing enforceable.

Why a two-page policy beats no policy

The risk in a small business is rarely the AI tool itself. It is unmanaged input. A free consumer chatbot account may use whatever is typed into it to improve the service, sits outside the business's control, keeps history on a personal login, and leaves no record the business can produce later. When the person leaves, the history leaves with them. None of that is visible to the owner until something goes wrong.

A short written policy fixes the part of this that can be fixed on paper: it names the tools staff may use, draws a bright line around the data that must never be pasted into any of them, and makes a human responsible for anything AI-written that leaves the business. Increasingly it is also simply asked for. Supplier questionnaires from larger customers, cyber insurance proposals and tender responses now routinely ask whether an AI use policy exists. "No" is becoming an expensive answer.

The policy, in full

Replace the bracketed placeholders, delete any clause that does not apply, and put a real name against clause 8. It is written for a business using Microsoft 365; swap the tool names if that is not you.

1. Purpose and scope

This policy sets out how [Business Name] and its people use artificial intelligence (AI) tools, including generative AI assistants and chatbots. It applies to all directors, employees and contractors, on any device, wherever work is performed, including work done on personal devices or personal accounts. Using AI for [Business Name] work in a way this policy prohibits is a breach of this policy even if the tool itself is free or personal.

2. Approved tools

The following AI tools are approved for business use, using a [Business Name] work account only: [Microsoft 365 Copilot / Copilot Chat, signed in with a work account]. Any other AI tool, and any approved tool accessed through a personal account, may not be used for business information. Staff may request approval of an additional tool from [role]; the request must cover what data the tool will receive and where that data is stored.

3. Information that must never be entered into an AI tool

The following must not be entered into any AI tool, including approved tools, unless [role] has approved the specific use in writing:

Personal information about clients, staff or any individual: names combined with contact details, dates of birth, addresses, financial details, government identifiers such as TFNs or Medicare numbers.
Health information about any individual, which the Privacy Act treats as sensitive information carrying stricter obligations.
Client-confidential material: contracts, pricing, disputes, legal advice, or anything received under a non-disclosure agreement.
[Business Name] financial records, credentials, passwords, API keys or security configurations.
Any information subject to a court order, regulatory investigation or legal professional privilege.

Where a task genuinely needs AI help with material like this, the approved path is the sanctioned work tool under clause 2, which keeps the data inside [Business Name]'s Microsoft 365 environment, and only with the data minimised to what the task needs.

4. Privacy Act obligations

[Business Name] [is / may be] covered by the Privacy Act 1988 (Cth) and the Australian Privacy Principles. Where the Act applies, personal information may only be used for the purpose it was collected for, or a related purpose an individual would reasonably expect. Entering personal information into an AI tool that stores or learns from it can amount to a disclosure to the tool's operator. Staff must treat any use of personal information in an AI tool as a use or disclosure that needs a lawful basis, and ask [role] before proceeding rather than after.

5. Confidentiality

Obligations of confidence to clients, suppliers and staff apply in full when using AI tools. Entering confidential information into an unapproved tool is treated as an unauthorised disclosure, in the same way emailing it to a personal address would be. Contractual confidentiality clauses in client agreements bind [Business Name] regardless of which tool a staff member used.

6. Human review of AI output

AI output is a draft, not an answer. Before AI-generated content is sent to a client, published, or relied on for a decision:

A named person must review it and takes responsibility for it as if they wrote it.
Factual claims, figures, names, dates and legal or regulatory statements must be checked against a source other than the AI tool.
AI must not be the decision-maker for matters significantly affecting an individual, including recruitment, performance and credit decisions; it may assist, a person decides.

7. Incident and breach reporting

Anyone who becomes aware that information covered by clause 3 has been entered into an unapproved tool, or that an AI tool has produced output that may have caused harm, must report it to [role] the same day. Reporting an honest mistake promptly will not attract disciplinary action; concealing one will. Where personal information is involved, [role] will assess whether the incident is a data breach requiring assessment under the Notifiable Data Breaches scheme and act on that assessment.

8. Ownership and review

This policy is owned by [name, role]. It will be reviewed every six months, or sooner if the business adopts a new AI tool or the law changes. Questions about whether a specific use is permitted go to the owner before the use, not after. Adopted on [date]. Signed: [director].

Adapting the template to your business

The template is deliberately conservative on privacy. Whether the Privacy Act formally covers your business turns mostly on turnover: businesses with annual turnover of $3 million or less are generally outside it, but there are important exceptions, including private health service providers and businesses that trade in personal information, and any business can opt in. Two practical notes on tailoring clause 4:

If your turnover exceeds $3 million, or you handle health information, state plainly that the Act applies and keep clause 4 as written. There is no judgement call to make.
If you sit under the threshold, resist the urge to delete the clause. Client contracts, professional obligations and plain commercial sense still require the same handling, larger customers increasingly flow privacy obligations down through supplier terms, and privacy reform has been moving in one direction. A policy built on the exemption is a policy you may have to rewrite; one built on good handling is not.

Beyond privacy, the clauses that most often need local tailoring are clause 2 (name the tools people actually have licences for) and clause 3 (add the specific data types your industry handles: patient records, trust account details, student information). The federal government's Voluntary AI Safety Standard is a useful checklist if you want to go further than this template, particularly its guardrails on accountability and record-keeping.

Why naming Copilot as the approved tool changes enforcement

A policy that only bans things loses to convenience. If the sanctioned path is slower than pasting into a free chatbot, the free chatbot wins quietly. The reason this template names Microsoft 365 Copilot as the approved tool is that it changes the enforcement problem from policing to defaulting: staff get an assistant inside Word, Excel, Outlook and Teams that is genuinely more useful for work tasks than a consumer tool, because it can see their files and mail, and the business gets commercial data protections that consumer accounts do not offer. Prompts and responses stay within the Microsoft 365 service boundary and are not used to train the underlying models.

It also makes clause 3 workable rather than aspirational. Inside the tenant, what Copilot can reach is governed by the permissions, sensitivity labels and audit trail the business already runs, so "never put client data in an AI tool" becomes "client data stays in the tenant, where the controls are". Frontrow has written separately on what that looks like in practice at /insights/is-microsoft-365-copilot-safe-company-data, and on the buying decision itself at /insights/copilot-for-business-australia.