Why solicitors ask this before using ChatGPT on client work
Solicitors should check a tool's terms, data handling and the firm's client instructions before entering matter information. Where those checks have not been made, keep confidential material out, even if the prompt omits a client name.
For most firms, the safest default is simple: do not put identifiable client information, privileged material, special category data, matter documents or commercially sensitive details into a public AI tool unless the firm has approved that exact tool, account type and processing arrangement.
The safe answer for solicitors using ChatGPT with client data
For a proposed client-data use, record the firm's decision on confidentiality, processing terms, retention, access and review responsibilities. Until those checks are complete, keep client material out. Removing names or using a firm-managed account does not establish safety: assess indirect identification, the information revealed and any processing by the provider or connected services.
For drafting, summaries, research support or first-pass analysis, identify the information processed, who receives it and the terms covering that use. Check the result before it influences advice or client communication. Assess professional confidentiality and privilege separately from data protection; a processing contract alone does not settle all three.
Check where AI is already touching client data
What counts as client data in this context
Client data is broader than a name in a prompt. It can include matter facts, document extracts, chronology details, contract clauses, witness information, case strategy, financial records, personal data, special category data, commercially sensitive information and privileged communications.
Even a prompt that looks anonymous can still reveal enough context to identify a client, dispute, transaction or individual. The policy should define prohibited inputs, how staff check a redacted prompt for indirect identification or sensitive facts, and which client tasks require a separately assessed processing arrangement.
The checks a firm should complete before approval
Before solicitors use ChatGPT or a similar tool with client data, the firm should answer five questions in writing:
- Processing role: what role does the provider take for each processing activity, including the prompt, uploaded files, logs and connected features?
- Contract terms: assess the provider's role and any applicable Article 28 requirements against the actual processing, product tier and enabled features. Does the contract cover that arrangement? An enterprise label is not a substitute for checking the terms.
- Training and retention: can prompts, files and outputs be used to train models, retained for review, or accessed by provider staff?
- Location and access: processing and storage locations, plus the people able to access logs, files and outputs.
- Firm controls: who owns accounts and permissions, what activity can the firm review, and which user guidance, matter restrictions and incident routes are needed? Assess SSO and audit logs against the workflow rather than treating them as sufficient safeguards.
If any answer is unknown, the tool should not receive client data. Unknown is not a safe processing basis.
How to use AI safely when the firm still wants the efficiency
A workable operating model separates low-risk AI use from client-data use. For general drafting support, structure suggestions, training examples or policy templates, consider synthetic examples that reveal no client information. If real material is redacted, check what the remaining facts disclose before using it. Client-specific work needs an assessed tool and processing arrangement, controlled access and a reviewer who understands the matter.
A legal team can define its operating rules around the following checks, adjusting them to the matters and tools involved:
- approved AI tool list with permitted and prohibited uses;
- prompting rules for redaction, indirect identification and sensitive facts that remain after names are removed;
- matter-type rules for privilege, litigation, employment, family, immigration and other sensitive work;
- a proportionate audit record of the tool, purpose and review, with access and retention rules that avoid unnecessary copies of client material;
- named professional owner for output quality and client communication;
- incident route for accidental disclosure or use of an unapproved tool.
Put practical AI governance around legal work
What the human review needs to prove
Human review should be more than a quick read-through. The reviewer should check legal accuracy, missing context, invented facts, outdated law, confidentiality risks, tone, and whether the output is appropriate for the client and matter. Where AI materially influenced a document or recommendation, decide what record will explain the review and final decision without retaining unnecessary confidential or personal information.
That evidence does not need to be heavy. It can be a matter note, workflow log, checklist or approval record. Record the task, relevant data categories, review performed and person responsible for the final work. Set an appropriate retention period rather than keeping prompts and outputs by default.
Conclusion
Before approving client-data use, assess the exact workflow's terms, data handling, retention, access and professional review. Keep client material out while those questions remain unanswered. A firm-managed interface still needs scrutiny of the provider and connected services behind it.
Build an approved AI workflow for client-facing work
