Is AI safe for law firms? Connecting it to client files

It can be, if you control three things: what the AI can read, what it's allowed to do, and where your data ends up. So is AI safe for law firms and accounting practices? Yes, when a person approves everything it sends and client data stays out of public tools. In my experience most of the risk comes from how it's connected, and much less from the AI itself.

I wrote this for principals and partners in law, accounting and other practices that hold confidential files. There's a vendor checklist near the end you can take into any demo. It's general information, not legal advice. For your own obligations, speak to your lawyer or your professional body.

Where does your client data go when you use AI?

When someone in your office pastes a client letter into an AI tool, that text leaves the building. It goes to the company running the model, gets processed on their servers, and may be stored there.

Three things decide whether that's a problem:

1. Who the provider is, and what the contract lets them do with your data.
2. Where the servers are. In Australia or overseas.
3. Whether your inputs train their models. Some products learn from what users type. Others don't, or let you opt out.

The regulator is clear on public tools. The OAIC's guidance on privacy and commercially available AI products says best practice is to keep personal information, and especially sensitive information, out of publicly available generative AI tools.

For lawyers it's sharper still. A joint statement from the Law Society of NSW and the legal regulators in Victoria and WA says solicitors can't safely enter confidential client information into public chatbots like ChatGPT. If you use a commercial tool with client information, they say to review the contract terms carefully. The Law Society's January 2026 guide goes further and recommends a full contract review before you buy, including how prompts and data are stored.

Business versions of AI tools often come with different terms to the free ones. Don't assume either way. Ask, and get the answer in writing.

What does the Privacy Act say about AI client data in Australia?

The Australian Privacy Principles apply to organisations with an annual turnover over $3 million, and to some smaller businesses regardless of turnover, including health service providers. If your firm sits under the threshold, your professional duty of confidentiality still applies, and your clients expect the same care either way.

The OAIC says privacy obligations cover what you put into an AI system and anything it produces that contains personal information. Three principles do most of the work:

• APP 6, use and disclosure. You can generally only use or disclose personal information for the purpose you collected it, unless an exception applies, such as consent. The OAIC notes that entering client details into a chatbot can count as a disclosure to the company that owns it.
• APP 8, sending data overseas. Before personal information is disclosed overseas, you have to take steps to protect it. Many AI services run outside Australia, so where the data is processed matters.
• APP 11, security. You must take reasonable steps to protect personal information from misuse, interference and loss. Connecting a new system to your files is a security decision, and the OAIC suggests checking a product's security record before you sign.

One date worth knowing. From 10 December 2026, if a computer program makes or informs decisions that could significantly affect someone's rights or interests using their personal information, your privacy policy has to say so. An agent drafting an email for a person to check is a long way from that. If AI scores, screens or declines anyone in your firm, ask your adviser whether it applies.

Reading vs acting: what's the difference?

When you connect AI to your systems, it can do two kinds of things.

It can read. Look up a matter, summarise a file, tell you which enquiries nobody followed up.

It can act. Send the email, update the record, book the meeting, move the money.

A bad read costs you a second look. A bad action can send a client's details to the wrong person, and you can't take that back. So keep read access narrow, limited to the folders and fields the job needs, and treat every action as something a person approves.

Who approves anything that gets sent?

A named person, every time, until you decide otherwise in writing.

That's how we build. Our agents draft and hold. Nothing leaves until a person has read it and said yes. The Law Society's guide makes the same point for lawyers: practitioners should always oversee and review AI-generated work. I explain how this works inside a CRM in what an AI CRM costs in Australia.

If a vendor tells you their agent sends on its own to save you time, ask two questions. What happens when it gets one wrong? And who finds out?

What should AI never touch?

My list for any practice:

• Client money and trust accounts. Our agents never move money, and we never hold banking logins.
• Internet banking and card numbers. Under any circumstances.
• Files it doesn't need. An agent chasing new enquiries has no reason to open litigation files or tax returns.
• Public AI tools. Client information stays out of them, as above.
• Systems outside what you agreed. If it isn't on the list, it doesn't get access.

Under all of this sits one rule: give it the least it needs to do the job.

Can you see what it did, and can you switch it off?

Both answers should be yes, today, before you connect anything.

Every action should be logged with what it did, when, and in which system. You should be able to ask for that log at any time.

You should also have a written list of every system the AI can touch, and be able to cut off any of them. With us, every system lives in a signed access register you can revoke instantly. Every action is logged. Message "PAUSE" and everything stops within the hour. You can read the full list on our services page.

If a vendor can't explain how to switch it off in one sentence, that tells you something.

Is AI safe for law firms? The questions to ask a vendor

Take this list into any demo. Ask for the answers in writing.

1. Which AI provider processes our data, and is that named in the contract?
2. In which country is our data processed and stored?
3. Is anything we send used to train models? Can we opt out, in writing?
4. Who at your company, or the AI provider, can see our inputs and outputs?
5. What can the system read, and what can it do? List both.
6. Does anything go out without a person approving it?
7. Is every action logged, and can we see the log whenever we ask?
8. How do we revoke access or switch it off, and how quickly does it stop?
9. Has this product had a security incident? What happened?
10. What happens to our data when we stop using you?

A good vendor answers these without fuss. A vague answer to question 3 or 6 is enough reason to walk away.

What to do this week

1. Ask your team what AI they already use. Free tools on phones count. You can't manage what you don't know about.
2. Write a one-line rule. No client information in public AI tools. It matches what the OAIC and the Law Society both say.
3. Read your privacy policy. The OAIC suggests updating it with clear information about how you use AI. If it doesn't mention AI at all, that's your first job.

Weighing up AI for your practice and want a straight answer on what it would touch? Book a 20-minute call. You can also read how we handle your data in our privacy policy.