Client data is safe with AI when you decide what goes in, not when the tool promises it is. The provider’s settings matter, but the rules are yours to write, and most businesses haven’t written them.
Most owners we talk to sit in one of two camps. In the first, AI gets treated like a search box, so privacy is someone else’s problem. In the second, a breach feels so likely that nobody in the business is allowed to touch it. Both cost you. The first is how a client’s payroll ends up in a free chat window. The second is how the business down the road gets a two-year head start.
What “safe” actually means
Four questions cover it. Who owns what you type in? Does the provider use it to train their models? Where is it stored? Who inside the system can see it?
Answer those for the tool you’re using and you’re most of the way there. The rest is deciding what your business will and won’t put in, and writing that down.
Read the data terms, not just the terms of service
Every AI product has terms of service and, separately, data processing terms. The first says what you can do with the tool. The second says what the provider can do with your information. Owners read the first and skip the second.
Look for the training clause. Some providers use your inputs to improve their models by default. Some let you opt out. Business plans usually rule it out in writing. Know which one you’re on before a client’s file goes anywhere near it. The free tier and the business tier of the same product often sit in different buckets.
Know where it lives and who shares the room
Where the data is stored matters twice. The Privacy Act asks you to handle personal information with reasonable security. Sending it offshore to a provider with weak protections is a risk you carry, even if you didn’t mean to. And the laws of the country where it sits decide who else can get at it.
Ask the provider for Australian or at least Asia-Pacific data residency. A clear answer is good. A vague one tells you something too.
Then ask about the access model. Consumer tools run thousands of businesses through shared infrastructure. That’s fine for drafting the newsletter. Client financials, health records and legal files are a different conversation, and that’s where business-grade accounts or a separate environment come in. You don’t need them for every task. You need to know which tasks they’re for.
Use the client-in-the-room test
Before pasting anything in, ask one question: if this client could see my screen right now, would they be fine with it?
Yes means go ahead. A pause means something. Either strip the identifying details first, use a tool with better terms, or keep that task manual for now.
Stripping details is easier than it sounds. Names, addresses, account numbers and dates of birth come out. The pattern stays in. An AI can tell you what’s unusual about a set of loan balances without knowing whose they are. Most of the useful work survives de-identification, because the AI is looking at the shape of the problem, not the person. Ordinary drafting passes the test easily. Feeding a client’s revenue figures into a free tool to tidy up a summary usually doesn’t.
Write the rule down so nobody improvises
One page is enough. It says what client data can go into AI and what gets de-identified first. It says which tools are approved and who can use them. And it says who to ask when someone isn’t sure.
Without it, the enthusiastic new hire is making privacy decisions on the fly, and the cautious one is doing everything by hand. Neither is what you’d choose. The other common gap is shared logins. One account, one password on a sticky note, and no way to know who put what in. Give people their own seats. It costs a little more and it’s the only way the rule can be checked.
This page is also part of the setup you own. The three documents that make AI useful in a business are who you are, how you work and what’s happening now. The data rule belongs in how you work. Put it there and every tool you use inherits it, because it’s your document, not the vendor’s setting. Swap tools next year and the rule comes with you.
The Australian angle
At the time of writing, the Privacy Act covers businesses turning over more than $3 million a year, plus health providers of any size. The Office of the Australian Information Commissioner has said it’s watching AI use closely. If you’re above the line, weak storage or training on client data without consent can put you in breach.
Below the line, your clients still care. A data incident doesn’t check your revenue before it hits your reputation.
The businesses handling this well haven’t avoided AI. They’ve decided what goes in, written it down, and moved on to the useful work. If you want that decided properly, set up your own AI Operating System starts with those documents. Work with Brightwaters covers doing it with your whole team in the room.