Australian lawyers keep asking a version of the same question: can we use Claude on client matters without waiving privilege? It is the right question. Legal professional privilege is one of the few protections that can keep a client's confidential legal communications out of the hands of an opponent or a regulator, and losing it can reshape a whole matter. The reassuring part is that privilege and careful AI use are compatible. The rules all come back to confidentiality, and confidentiality is something a firm can control.
What legal professional privilege actually protects
Legal professional privilege, also called client legal privilege in the Evidence Act 1995 (Cth), protects confidential communications made for the dominant purpose of giving or obtaining legal advice, or for use in litigation. It belongs to the client, not the lawyer. Where it applies, the client can refuse to produce those communications in court, in discovery, or in response to many regulator notices.
The protection rests on one foundation: the communication has to stay confidential. Privilege can be lost through waiver, and the most common way that happens is disclosure to a third party in a way that is inconsistent with keeping the material confidential. That is the exact point where AI tools enter the picture. If sending a privileged document to a tool counts as a disclosure that breaks confidentiality, privilege is at risk. If it does not, privilege holds.
Where AI tools create privilege risk
The privilege question with any AI tool comes down to who can see the material and what they are allowed to do with it. A consumer chatbot on a free plan and a contracted business deployment are very different animals, even when the underlying model is the same. Three factors decide which side of the line you are on:
Retention and training. If a provider keeps your inputs and uses them to train models, privileged content could travel beyond your control. Anthropic's commercial terms for Claude do not train on business or API inputs by default, which is a real difference from consumer settings.
Access and confidentiality terms. Privilege depends on the material staying confidential. A written agreement that binds the provider to confidentiality, with defined access limits, supports the argument that confidentiality was preserved.
Data handling and residency. Where the data is processed, who administers it, and how long it is kept all feed into whether a court would see the arrangement as consistent with maintaining privilege.
None of these are exotic. They are the same questions a firm already asks before putting client files in a cloud document system or an e-discovery platform. AI tools deserve that same diligence, not a separate category of fear.
The handling rules that keep privilege intact
A firm that wants to use Claude on privileged work should put a small number of rules in place and follow them every time. Consistency matters, because waiver arguments often turn on whether a firm treated the material as confidential in practice, not just on paper.
Use a commercial or enterprise tier, never a personal consumer account, for anything touching client matters. The terms and the data handling are different.
Have a written data processing agreement in place before privileged material goes near the tool, covering confidentiality, no-training, retention limits, and sub-processors.
Limit access to the deployment to people already inside the circle of confidentiality for the matter, the same way you would with a physical file.
Keep a short record at the matter level of what was processed and under which terms, so you can later show the material was handled as confidential.
Treat Claude's output as a draft for a supervising lawyer to review, not as advice in itself. The lawyer stays responsible under the Australian Solicitors' Conduct Rules.
Never paste privileged material into any tool whose terms you have not read, including browser plugins and meeting note-takers that quietly send text to a third party.
Firms sometimes worry that these steps make AI too slow to be worth it. In practice the setup is a one-time cost. Standing up a compliant Claude deployment for a mid-size firm, including terms review and access controls, usually runs a few thousand dollars of internal or advisory time. Set that against a downside where a single contested privilege claim in litigation can add $80,000 or more to a matter. The asymmetry favours doing it properly once.
Does using a contracted AI tool waive privilege?
The cautious view, and the one most Australian practitioners are settling on, is that using a properly contracted AI tool does not automatically waive privilege, because waiver requires conduct inconsistent with maintaining confidentiality. Disclosing privileged material to a service provider who is bound to keep it confidential, under controlled access, is generally consistent with keeping it confidential, in the same way that using an external typing service or a cloud host has long been accepted. The risk lives in the details: an unread free-tier agreement, one open account shared across unrelated matters, or privileged text pasted into a tool with training switched on.
Courts and regulators have started to give guidance. The Supreme Court of New South Wales and other Australian courts have issued practice notes on generative AI in proceedings, and the Law Council and state law societies have published guidance for practitioners. The Privacy Act 1988 and OAIC guidance also apply whenever the material contains personal information, which most client files do. Reading the current version of these before you write a policy is worth the hour.
What good looks like at an Australian firm
A firm handling privilege well with Claude usually looks calm rather than clever. There is a written policy that names which tiers and tools are approved. There is a standard set of terms that has already been through the firm's own risk review. Access is scoped to matters. Staff know that privileged material only goes into approved tools, and they know why. When a client or a court asks how AI was used, the firm can answer in a sentence and point to a record.
That posture does more than protect privilege. It lets a firm actually use Claude on the work where it helps most: summarising long document sets, drafting first-pass correspondence, building chronologies, and checking a brief for gaps, all under lawyer supervision. The confidence to use the tool comes from settling the handling rules first.
This is general information for Australian legal teams, not legal advice on your circumstances, and privilege questions can turn on the facts. If you want help setting up a Claude deployment that respects legal professional privilege and your conduct obligations, book a working session with Automata AI and we will map it to how your firm actually handles matters.
