Which client data may be entered into AI tools?
Classify the client material and approve it for the exact tool, purpose and account tier. If you cannot confirm terms and safeguards for that use, do not upload it.
A client email, an audit working paper and a valuation schedule need different checks. Give staff examples of permitted and restricted inputs for the task they are doing, rather than asking them to interpret a general instruction to use AI carefully.
The safest answer is to classify the data before anyone uploads it
Define three buckets for each approved use. Allow inputs that have been checked for that task, such as public material or tested examples. Require separate approval for client documents, financial schedules, claim files, contracts, matter notes and audit evidence. Exclude credentials, client-restricted material and any other data the service has not been approved to handle. Do not assume that an anonymised example or internal draft is suitable without checking what it still reveals.
This keeps staff from making the decision at the prompt box. It also gives partners, compliance teams and reviewers a clear record of why a tool was suitable for a task.
Map where client data touches AI in your workflows
What firms should check before approving client data use
Start with the tool, not the model brand. The question is whether the configured service protects the data in the way the firm needs. Check whether prompts, uploaded files, outputs and logs are retained, whether they can be used for training or evaluation, where processing happens, who can access the content and how deletion works. If the vendor cannot answer clearly, treat the tool as unsuitable for sensitive client work.
Have the engagement's relevant reviewers determine which professional and data requirements apply to the actual material and decision. Record the sources, conditions and unresolved points rather than infer obligations from the firm's sector or treat all confidential material as personal data.
Set rules for matter, audit, valuation and advisory data
Matter and advisory files often contain client strategy, negotiations, commercial terms and privileged context. Audit and valuation material can include evidence, assumptions, estimates, judgements and working papers. These should not be pasted into public AI tools or unapproved assistants. If assessing AI support, review the exact environment, permissions and named reviewer. Define proportionate logging with confidentiality, minimisation, access and retention controls; use controlled source references where appropriate and do not retain complete prompts or client files by default.
The policy should also cover partial data. A spreadsheet excerpt, a meeting note or a pasted email can still identify a client or reveal a live issue. Staff need examples of what counts as client data, not only a legal definition.
Put an AI data-use policy on a monthly governance rhythm
How to make the rule usable in daily work
A useful control is a simple decision checklist attached to each approved tool:
- Is this tool approved for this category of client data?
- Is the client, matter or engagement type restricted?
- Will prompts, files, outputs or logs be retained or used for training?
- Is a human reviewer responsible for the output before it is relied on?
- Can the reviewer trace the material use and decision using proportionate records or controlled source references, without retaining full prompts or client files by default?
If the answer is unclear, do not upload the material. Consider a separately approved public or synthetic example instead. Summarising a client file or removing names may leave sensitive facts in the input, so check the replacement material too.
Conclusion
Before an upload, staff should be able to check the data category, approved task and service settings. Record the decision and reviewer without copying more confidential material into the register than the review needs. Recheck approval when the input or task changes.
Build the controls into day-to-day AI workflows
Need a safer AI route for your firm?
If staff are unsure what they may upload, map the actual client inputs against approved tasks and services. Use those examples to write rules they can check before entering data.