A short list is enough to start. The people using AI every day need to know the six rules, and someone needs to check now and then that they're followed. That keeps the risk down while your team gets on with using AI for the work it's good at.
1. Client information going into the wrong tool
What goes wrong. Someone pastes client details into a free or personal AI account to save ten minutes. The Australian Cyber Security Centre describes a real case from early 2025: a contractor uploaded people's names, contact details and health records into an AI system, which led to a serious data spill and is considered a notifiable data breach. Some AI providers may also use what you type to train their models, which can depend on your settings or subscription.
The rule. "Client names, contact details, health and financial information, passwords and anything confidential only go into our approved tools, signed in with a work account. Even there, include only what the task needs."
How to check. Confirm every approved tool is on a business plan with training on your data switched off, and ask the team every quarter whether anyone's still using a personal account.
2. Confident wrong answers going out
What goes wrong. AI can produce answers that sound confident but are wrong, incomplete or misleading, and the National AI Centre notes these errors aren't always easy to spot. The Cyber Security Centre cites a 2025 case where a lawyer filed a court document with AI-generated cases that didn't exist and was later barred from running a law practice.
The rule. "Anything AI helped with gets checked by a person before it's sent, published or relied on. Check the facts, figures, names, dates and tone. The person who sends it is responsible for it."
How to check. Pick one piece of AI-assisted client work each month and ask who reviewed it and what they changed. If the answer is "nothing, it looked fine", that's a conversation to have.
3. Unapproved tools and personal accounts
What goes wrong. Whatever your policy says, some of your team are probably using AI already. The National AI Centre calls this shadow AI, and says it's often a sign of unmet need, time pressure or curiosity. The risk is that you can't see what's going in or set any rules for it.
The rule. "Use only the AI tools on our approved list. If you want to try a new one, ask [name] first. We'd rather say yes to a safe tool than have you work around us."
How to check. Give people a good approved tool and training on their own work, then ask openly what they're using. One client, a commercial property firm of around 70 people, had a few people using AI in patches and most of the team unsure where to start. They chose Claude Team for their property management and sales teams. We ran a session for everyone, then hands-on workshops for those teams using their own emails and leases. Within the first month, staff showed or told us about 18 different ways they were using it.
4. AI shaping decisions about people
What goes wrong. AI can reflect or magnify bias in the data it learned from, and the National AI Centre says this risk is higher in hiring, pricing, eligibility and customer decisions. A biased shortlisting tool can reject many qualified people before anyone notices. There's also a legal date to know: from 10 December 2026, businesses covered by the Privacy Act must explain in their privacy policy when computer programs use personal information to make decisions, or do something substantially and directly related to making them, that could significantly affect people's rights or interests.
The rule. "AI doesn't make final decisions about people, including hiring, pay, performance, credit or pricing for an individual. Any use of AI in those decisions is approved by [name] first and a person makes the call."
How to check. List every tool that touches decisions about people, including AI features inside your recruitment or finance software. If you're covered by the Privacy Act, check your privacy policy against that list before 10 December.
5. People not knowing they're dealing with AI
What goes wrong. A customer chats with your website bot thinking it's a person, or gets an AI-written reply that nobody reviewed. The Guidance for AI Adoption says people should know when they're interacting with AI, and the OAIC, Australia's privacy regulator, says public-facing AI tools such as chatbots should be clearly identified as AI.
The rule. "Any chatbot or automated message a client sees says it's AI. If a client asks how we use AI, we tell them."
How to check. Look at every place a client could meet AI in your business, like your website, inbox, phone line and booking system, and read it as a client would.
6. AI connected to your systems
What goes wrong. Once AI can read your email and files, or take actions in them, mistakes travel further. The Cyber Security Centre warns about prompt injection, where malicious inputs disguised as normal requests trick AI into giving sensitive, wrong or unsafe answers. Anthropic's safety guidance says those instructions can sit in websites, emails or documents the AI reads. The Cyber Security Centre also flags supply chain risk, because you rely on the provider's systems and any weakness in them can affect your business.
The rule. "New connections between AI and our systems are approved by [name]. AI only gets access to what the person using it can already see. Anything AI does on its own, like sending emails or deleting files, needs a person to approve it."
How to check. Keep a list of what each AI tool can connect to. Turn on one connection at a time, check what AI can see, then add the next. Review the list when a provider announces new features.
The six rules on one page
| Risk | The rule in short | Who checks | How often |
|---|---|---|---|
| Client information in the wrong tool | Approved tools only, work accounts, minimum detail | Policy owner | Quarterly |
| Wrong answers going out | A person checks before anything leaves | Whoever sends it, plus a manager spot check | Every time, spot check monthly |
| Unapproved tools | Approved list only, ask before trying a new one | Policy owner | Quarterly |
| Decisions about people | AI never decides alone; approval first | Policy owner and a director | Before any new use, and before 10 December 2026 |
| People not knowing it's AI | Label it, and be open when asked | Whoever owns the channel | When anything changes |
| AI connected to systems | Approve each connection, people approve actions | Policy owner | Each new connection, and quarterly |
Put these six rules into your AI use policy. Our free AI use policy template already covers most of them, so you'll mostly be filling in names.
When a use needs more than a rule
Some uses need a proper look before they start. The National AI Centre says risk is higher when AI:
- interacts directly with customers or the public
- influences important decisions about people
- runs with little or no input from people
- runs continuously or at scale
- uses personal, sensitive or confidential information.
Run any new use like that through the National AI Centre's seven screening questions first. If any answer is yes, it may need stricter controls, a named owner and a plan for checking it over time. Our guide to an AI governance framework for directors shows how that fits on one page.
Common questions
It's how a business finds what could go wrong with its use of AI and decides what to do about each risk. For a small business that usually means a short list of risks, one rule for each, and someone checking the rules are followed.
For most, it's client information going into unapproved tools, and wrong answers going out unchecked. In our experience those two come up most often, and both have simple rules.
Not to start. A short list of rules and a named owner covers most owner-led businesses. If you build your own AI tools, use AI in decisions about people, or work in a regulated industry, you'll need more, and the National AI Centre's implementation guidance goes into detail.
Give them a better approved option and make it easy to ask questions. Banning AI tends to push it onto personal phones and accounts, where you can't see it.