AI at work: what's safe to use

The real risk is already in the building

Most small businesses don't have an AI problem; they have a personal-login problem. Someone in the office is already pasting a client's email into a free chatbot to draft the reply, or dropping a spreadsheet of names in to "tidy it up". None of it is malicious. All of it is happening on accounts the business can't see, with information the business is responsible for.

On a free personal plan that content has left the business. Depending on the vendor's settings it may be used to train the model, kept for a long time and reviewed by people, and it isn't covered by any agreement your business holds. If the staff member leaves, the chat history goes with them. That's the risk to fix first, and it's mostly a matter of which login people use.

Free and personal plans versus business plans

The vendors draw a clear line between consumer and business products. The table is our reading of their published policies on 31 October 2026; they change, so check the links at the bottom before relying on it.

The vendors' own published claims. "Business tier" means they commit not to train on your content and give you controls and a contract. It doesn't make anything safe to paste, and it doesn't change your obligations under the Privacy Act.

Copilot's strength is also its trap: it can only show a person what that person already has access to, so if your SharePoint permissions are a mess, it will cheerfully surface the pay spreadsheet to anyone who asks the right question. Tidy the permissions before you switch it on.

What never goes in

Whichever plan you're on, these stay out of the prompt box:

• Client records, and anything that identifies a client alongside details about them: a name with a diagnosis, a name with an address and a job.
• Health information. The Privacy Act treats it as sensitive information, with extra protection.
• Tax file numbers, Medicare numbers, NDIS participant numbers, licence numbers, dates of birth.
• Bank account and card details, yours or anyone else's.
• Passwords, MFA codes, API keys, anything that lives in the password manager.
• Other people's confidential information: anything under an NDA, a supplier's pricing, a tender.
• Staff HR matters: pay, performance, medical certificates.
• Whole documents "to summarise", when a copy with the identifiers taken out would do the same job.

The Privacy Act basics for a small business

The Australian Privacy Principles apply to businesses with an annual turnover over $3 million, and to some smaller ones regardless of size, above all health service providers: physios, OTs, psychologists, GPs and most NDIS providers delivering health-type supports are covered from their first dollar, as are businesses that trade in personal information. If you're a small trade or professional business under the threshold, the Act mostly doesn't bind you; your contracts and your reputation still do. Health practices in Victoria and New South Wales also sit under state health-records laws with no turnover threshold.

If the Act does apply to you, the parts that matter for AI are these:

• Health information is sensitive information. You need consent to collect it, and you can only use it for the purpose you collected it for.
• Collection notices. People must be told what you collect and what you do with it. If an AI tool is part of what you do with it, that belongs in the notice and in your privacy policy.
• Overseas disclosure. Most AI tools process data outside Australia. The Act makes you responsible for what happens to it there.
• Security. You must take reasonable steps to protect the information, which is hard to argue if it's sitting in a staff member's personal chatbot history.

The OAIC published guidance in 2024 on using commercially available AI products. Its plainest line is the one to remember: as a matter of best practice, don't enter personal information, and particularly sensitive information, into publicly available generative AI tools.

What changes on 10 December 2026

The Privacy and Other Legislation Amendment Act 2024 adds a transparency rule about automated decision-making, and it starts on 10 December 2026. In plain terms: if your business uses a computer program, which includes an AI tool, to make a decision about a person using their personal information, or to do something substantially and directly related to making that decision, and the decision could reasonably be expected to significantly affect the person's rights or interests, your privacy policy must say so. It has to describe the kinds of personal information used and the kinds of decisions involved.

For most small businesses that's a narrower change than it sounds. Drafting a quote with AI isn't a decision about a person. Letting a tool decide who gets a service, score a job application, or approve or reject a claim is. If you're anywhere near that line, get it into the policy before December, and read the OAIC's material on the amendments rather than our summary; we're not lawyers.

AI scribes in health

Ambient scribes that listen to a consultation and write the note are the most useful and the most sensitive AI tool a practice can adopt. Ahpra and the National Boards' guidance on meeting your professional obligations when using AI in healthcare comes down to a few things you'd expect:

• You remain responsible. The note is yours, whatever wrote the first draft.
• Informed consent before a consultation is recorded, and a way for the patient to say no without fuss.
• Review every output before it goes in the record. Scribes make confident mistakes.
• Know where the audio and the text go, who holds them, for how long, and whether they're used for training. That's a question for the vendor, in writing, before you sign.
• Keep the clinical judgement. The tool documents; it doesn't diagnose.

In practice: a business account with a signed data agreement, a one-line consent script at the front desk and on the intake form, a review-before-save habit, and a sentence in the privacy policy that says you use it.

A one-page AI-use policy

It doesn't need to be longer than a page, and it won't be read if it is.

1. Approved tools and accounts. Name them. Business logins only; personal accounts are not used for work, full stop.
2. What's allowed. The green list below, in your own words.
3. What's banned. The never list above, in full.
4. Review before it leaves. A person reads every AI-drafted email, quote, note or document before it's sent, saved or acted on, and that person is responsible for it.
5. Who to ask. One name for "is this OK?", and a no-blame way to report a mistake, such as pasting the wrong thing.
6. When this gets reviewed. Every six months, or when a tool changes.

Then tell staff what it says. Thirty minutes with the tools open beats a PDF nobody opens.

Good first uses

• Quote drafting. Turn a job note or a description of the site into a clear quote for you to check and price.
• Inbox triage. Sort what came in overnight, flag what needs you, draft replies you approve before they go.
• Document templates. Letters, onboarding checklists, position descriptions, a plain-English version of your terms.
• Meeting notes. A Teams recording turned into decisions and actions, with everyone told it's recording.
• Rewriting. Your own email or your own service page, made shorter and plainer.

We set up quote drafting and inbox triage for an electrical contractor this year. Small and unglamorous, which is why they're the right place to start.

Traffic lights

Where we fit

Our AI set-up for small business is the practical version of this guide: tools chosen and set up on business accounts (Microsoft 365 Copilot, Claude or ChatGPT, whichever fits how you work), a one-page AI-use policy written for your business, a staff session, two or three automations first, and a monthly check-in if you want one. From $1,500 one-off, with the optional check-in at $200 a month. For health and NDIS clients the AI-use policy is written alongside the privacy policy, so the two agree.