What Staff Should Never Paste Into an AI Tool
A staff member pastes a whole customer email thread into an AI tool to get a quick reply drafted. It worked, and nobody noticed. That's the moment the policy line is for.
Official guidance explains the risk plainly: "uploading customer or staff details into generative AI platforms without proper anonymisation can expose sensitive or private information".1
Why the line exists
Two facts sit behind every rule below.
First, what you type may not stay with you. Guidance for government staff says: "When you use public tools, your inputs and outputs may be shared with the tool provider."2
It also gives a blunt rule of thumb. "Assume anything you enter into a public generative AI tool could be made public."2
Second, some providers may reuse what they receive. Australia's cyber security agency says whether they do "can depend on the configuration settings or the type of subscription."1 So the answer can differ between account types. That's why the approved tools list matters.
The categories to name in your policy
Write these out in full. "Sensitive data" alone won't stop anyone.
- Customer details. Names, addresses, phone numbers, emails, account or order numbers.
- Staff details. Pay, performance notes, leave, medical certificates, anything from an HR file.
- Financial records. Bank details, invoices tied to a named client, payroll.
- Health information. Anything about a person's health. Health businesses should keep it out entirely, and use AI only for general front-desk and admin tasks, never anything clinical.
- Confidential business material. Contracts, unreleased plans, pricing you haven't published, supplier terms.
- Other people's copyright material. Government guidance tells staff not to put "third-party copyright-protected information into public generative AI tools".2
- Logins and secrets. Passwords, access codes, security keys.
The list of examples above is a starting point, not a legal checklist. What counts as personal or sensitive information is set by privacy law where you operate. Check your national regulator, such as the OAIC in Australia, the Office of the Privacy Commissioner in New Zealand, the ICO in the UK, or your lawyer.
What happens when it goes wrong
The cyber security agency describes a case from early 2025 where a contractor for an Australian organisation put personal details of people in a government program into an AI system. It led to a serious data spill that was treated as a notifiable data breach.1
The same page says where such breaches get reported: "Report a notifiable data breach to:" and then names the privacy regulators in Australia and New Zealand.1 If you think one has happened, don't wait for the policy meeting. Contact your regulator or lawyer.
When staff really need AI help with a real case
Sometimes a task is built around a real person's details. There are two honest options.
- Anonymise first. The agency's advice is to "remove, anonymise or change personal details when using an AI application". Its stated aim is that the text can't be used to identify or link to an individual.1 Swap "Mrs Collins at the Hobart branch" for "a customer". Strip account numbers.
- Do it without AI. If the task can't be made anonymous, a person does it the old way. That's a legitimate answer.
Anonymising takes more care than it looks. A job title, a suburb and a date together can point to one person. If staff aren't sure, they should ask the policy owner.
Wording it so it sticks
Short, concrete and with examples beats long and legal. Put the "never" list on one page staff can keep open. Add a line saying who to ask when unsure and that asking is always fine.
The whole policy is laid out in the section-by-section outline. You'll find sentences to copy in the sample wording.
Notes and sources
- Artificial intelligence for small business, Australian Cyber Security Centre (ASD)
- Staff guidance on public generative AI, Digital Transformation Agency (Australian Government)