What Client Data Rules Apply to AI Agents in a Law Firm
AI agents in a law firm touch client data the moment they open a document, so the useful question is not what the tool can do — it is what you are allowed to put in front of it. This is the boundary we draw before an agent goes anywhere near a matter file, the three places where it is genuinely arguable, and the one question most practices cannot answer today.
The duty did not change. The volume did.
Confidentiality obligations over a client’s information apply to how you handle that information. They do not loosen because a model is doing the reading. A solicitor who emails a client’s file to a stranger has breached something; the fact that the stranger is software does not help.
What has changed is scale. An agent reading four hundred documents in a morning does work a paralegal used to do — at a volume where nobody reads every output. The duty is the same. The opportunities to breach it are not.
The Law Society of NSW has published guidance mapping existing conduct obligations onto AI use. Read the current version yourself rather than relying on a vendor’s summary, and confirm the position with your professional body and insurer before you change a workflow. That is not a disclaimer. It is the actual first step.
The boundary, drawn plainly
Stays inside the firm
- Reading and extracting fields from documents already in your practice management system
- Building a chronology from material you already hold
- Comparing a draft against your own precedent bank
- Drafting an internal summary for a solicitor to check before it goes anywhere
- Filing, naming and finding documents consistently
Does not cross
- Pasting client material into a free public chatbot to see what happens
- Any flow where output reaches a client, an opponent or a court without a solicitor reading it
- Sending privileged advice to a provider whose terms permit training on it
- An agent deciding whether a conflict exists
- Automated advice on prospects, even hedged
The left column is most of the value and almost none of the risk. Firms that start here rarely need the difficult conversation, because the agent never leaves the building.
The grey zone is where firms actually get stuck
Genuinely arguable
Transcribing client interviews. The recording is client information. Whether transcription is a disclosure depends on where the model runs and whether the provider is a processor acting on your instructions.
Summarising a client’s own emails back to them. Usually fine. Unless the summary is being treated as advice, in which case a solicitor has to have read it.
Cross-matter retrieval. “Have we seen a clause like this before?” is one of the most valuable things an agent can do. It is also the one that can quietly move information between matters that should never touch.
None of these are settled by a policy document. They are settled by a data map: what you hold, where it sits, who can reach it, and what leaves the building when a button is pressed.
Privilege does not care that it was convenient
If privileged material is processed by a third-party model under terms nobody in the firm has read, you have created an argument you will have to run later, at the worst possible moment. Some providers now offer no-training processing in an Australian region. Some offer it only on a higher tier. Verify the specific clause rather than the marketing page.
Local deployment — the model running inside infrastructure the firm controls — removes most of this question. It costs more to set up and does less out of the box, but for a firm holding privileged material it is frequently the cheaper decision once insurance is priced in.
Where this breaks down
A firm that cannot say where a document lives today cannot govern an agent that reads documents. Most practices hold three copies of every matter and no map. The AI policy is not the first job; the inventory is.
A boundary written as principles — “we use AI responsibly” — is unusable. It has to name processes, systems and approvers, or nobody can follow it under pressure.
And a compliant workflow in March is not automatically compliant in September, because someone added a step in June. The boundary needs an owner, not a filing cabinet.
What to write down before you switch it on
- The named process. Not “document review”. “Intake of disclosure from the other side in family matters”.
- The systems it touches. Practice management, document store, email, anything else.
- Who approves the output, and what they are approving. A name, not a role.
- Where the model runs. On your infrastructure, in an Australian region, or somewhere you would have to explain to a client.
- Retention and deletion. What the provider keeps, for how long, and how you prove it.
- The failure path. Who notices when it is wrong, and what happens to the matter in the meantime.
Two pages. If it takes longer than that to write, the process is not ready to automate — which is useful information in itself.
What we build
What we build
Agents that read, file and assemble inside the firm’s own infrastructure, with a solicitor approving anything that leaves the building and an audit trail of what was read. We do not put client material through a public model to demonstrate a demo.