How to Write an AI Policy for a Queensland Law Firm
Published 25 August 2026
Short answer: QLS Guidance Statement No. 37 doesn't give firms a template, but it's specific enough to build one from. At minimum, a firm AI policy needs to name which tools are approved, what client information can and can't go into them, who verifies AI-assisted work, how client consent is captured where a third party is involved, who owns the policy, and how breaches are handled.
A recurring theme across professional AI guidance (TPB(GS) 55/2026 for accountants, AHPRA's guidance for clinicians, QLS GS37 for solicitors) is that "we told staff to be careful" isn't a policy, and regulators increasingly expect an actual one. For law firms, GS37 points to a documented, communicated AI policy as the practical starting point for reducing exposure. Here's what that document needs to contain.
Why informal guidance isn't enough
A firm-wide message or a comment in a staff meeting might genuinely change behaviour, but it doesn't create anything a supervisor or legal practitioner director can point to later. If GS37's firm-level responsibility provisions are ever tested, after an incident, a complaint, or a privilege dispute, "we told everyone" is a much weaker position than a dated, distributed, acknowledged policy. The gap between the two is mostly paperwork, which makes it one of the easier compliance gaps to close.
What the policy needs to cover
- 1. Which tools are approved, and which aren't. Name them specifically. "AI tools" as a category is too vague to enforce. Staff need to know whether a specific tool is approved, and what to do if they want to use one that isn't.
- 2. What information can and can't go into an approved tool. This is where the policy earns its keep. General research assistance carries different risk to summarising a client's discovery documents. The policy should draw that line explicitly rather than leaving it to individual judgment matter by matter.
- 3. Who verifies AI-assisted output, and how. GS37 is explicit that AI output must be checked by someone competent to know if it's wrong before it reaches a client or a court. The policy should name who that is for different types of work, rather than leaving verification as an assumed but undefined step.
- 4. How client consent is captured, if a tool involves a third party. Where a tool sends data anywhere outside the firm's own systems, the policy should specify how and when client consent is obtained, and where it's recorded.
- 5. Who owns the policy. GS37 points to supervisors and legal practitioner directors as accountable for firm-wide systems. The policy should name an owner, responsible for updating it as tools and guidance change.
- 6. How breaches or near-misses are reported and handled. Even a good policy will occasionally be worked around. A policy that doesn't say what happens next isn't complete.
A structural decision that makes the rest simpler
Every element above gets easier to write, and to follow, when the answer to "where can client data go" is simple. A policy that has to list approved external vendors, explain what each does with data, and set consent requirements tool by tool is harder to get right than one that can say plainly: approved AI tools run on infrastructure the firm owns, client data doesn't leave the firm's systems, and the verification and supervision requirements apply regardless.
That doesn't remove the need for a policy. GS37's supervision and verification requirements apply wherever the AI runs. But it removes the hardest section to write, which is usually the one governing external data flows.
Where to start
A policy doesn't need to be exhaustive on day one. The highest-leverage first step is naming what's currently in use, including tools embedded in existing software that nobody formally approved, and getting that list in front of whoever will own the policy. Everything else follows from an accurate starting inventory.