Do law firms need a DPA before using AI tools?
A law firm needs Article 28 processor terms when an AI provider processes personal data on the firm's behalf. First check the provider's role and terms for the feature being used; confidentiality and professional duties require separate controls.
For the proposed legal task, ask the relevant professionals to identify the confidentiality, privilege, supervision and file requirements that apply. Record the approved tool, data boundary, output reviewer and reasons, with any unresolved legal question kept open rather than described as safe use.
The direct answer for client data
Where the provider acts as a processor for the firm, the DPA question concerns the processing it performs on the firm's behalf. Review the supplier's terms for that role before approving client-data use. Separately check whether prompts, files or outputs can be used for model training or improvement, what transfer arrangements apply, and what records the firm can keep of the approved use.
Keep internal experiments within the firm's approved AI policy and use material cleared for that purpose. Anonymous or synthetic test material avoids putting real matter details into an unapproved tool; material being public does not by itself settle the firm's data-use obligations. Procurement, compliance and supervision should agree the boundary before fee-earners start testing.
Map the AI risk controls before client data enters a tool
What the DPA should prove before legal work starts
For processing carried out on the firm's behalf, review the agreement against the actual service and data involved. Check the recorded processor role, categories of personal data, processing and storage locations, sub-processor arrangements, retention and deletion terms, and breach notification provisions. Refer gaps to the firm's responsible data or legal reviewer.
The wider supplier review should also check model-training use, human access to client content, access controls, logging, export and deletion rights. Check whether any security certification offered as evidence covers the product tier in use. Record where each commitment appears in the contract or settings, and keep client personal data out of the tool while material gaps remain unresolved.
How this connects to SRA duties and client confidentiality
Even a well-written DPA does not remove the firm's professional duties. The firm still needs to decide which legal tasks AI may support, which tasks are out of scope, who reviews outputs, and how errors are caught before advice reaches the client. AI should assist controlled workflow steps such as first-draft research notes, summarisation or internal checklists only where the matter risk and data controls are suitable.
Higher-risk work needs tighter rules. Privileged information, vulnerable-client matters, regulated advice, litigation strategy, negotiation positions and final client communications should not be routed through an unapproved or weakly governed AI tool. Where AI is allowed, the file should show what tool was used, what data was entered, who reviewed the output and why the final legal judgement is defensible.
Keep AI policy, vendor review and audit evidence current
A practical implementation checklist
- Classify the use case: separate research, drafting, review, client communication and decision support.
- Classify the data: mark whether the workflow touches personal data, special-category data, privileged material or commercially sensitive matter information.
- Check the contract: establish the provider's role, review processing terms where it acts on the firm's behalf, and separately check training use, sub-processors, retention, deletion and transfers.
- Check the product tier: consumer, team, enterprise and API terms can differ materially.
- Set review rules: name the professional owner and require human review before the output affects a client or matter file.
- Keep evidence: log the approved tool, task, data boundary, reviewer and exception handling route.
Carry the approval decision into procurement, permissions, staff training, matter-opening guidance and supervision. Staff need to know which data and tasks are permitted, who reviews outputs and where to raise an exception. A signed agreement does not replace those working instructions.
When the answer should be no for now
Pause the proposed client-data use if the provider's role or applicable terms are unresolved, its data locations or training rights are unclear, or the required deletion, security or logging evidence is missing. Also pause if staff cannot explain what data they would send or who would review the output. Resolve those gaps against the intended matter workflow before approving it.
Start with a limited workflow and approved data boundary, then check whether the review and evidence steps work in practice. Use that record to decide whether the proposed next workflow needs different controls before expanding.
Turn approved AI use cases into controlled implementation steps
Conclusion
Before approving client personal data for an AI tool, establish the provider's role and applicable terms, training-use position, transfer arrangements, review process and evidence record. The firm should be able to explain what data was used, why the use was approved and who checked the output. Keep unresolved contractual or professional-duty questions with the responsible reviewer.
Need a safer AI route for your firm?
For the intended matter workflow, record the provider's role, permitted data, unresolved contractual questions and named reviewer before approving tool use.