Why client-data boundaries come before AI adoption
Set data boundaries for each approved AI tool and task. Public research may be suitable for a low-risk tool; identifiable or confidential client material needs a separate decision about permissions, provider terms and human review.
Before staff enter client financials, legal facts, employee records or draft advice, record whether the firm has approved that material for the tool and task. Escalate any use that falls outside the approval or client instructions.
The safest approach is a four-zone data rule
Use four zones: prohibited data, restricted data, approved data, and anonymised or synthetic data. Keep prohibited material out of unapproved AI tools and honour any client prohibition on external processing. Assess restricted material separately for the proposed task, including permissions, contract, retention, training use and access controls. Match approved data to the tool and use covered by the decision. Use synthetic examples or checked anonymised material for prompt tests, templates and staff training.
This structure gives staff a fast decision path. It also gives compliance and practice leaders something auditable: what data category was used, which tool was approved, who reviewed the output, and what evidence sits behind the decision.
What belongs in the prohibited and restricted zones
Keep privileged legal material, confidential client facts, non-public financial data, live personal data, special category data, credentials, security information and trade secrets out of unapproved tools. Where client instructions bar external processing, keep the material out of that processing even if the tool has otherwise been approved.
Assess restricted material for the sector, engagement and client instructions before permitting it in a controlled environment. Examples include client drafts, matter summaries, support tickets, claim narratives, board packs, financial models, due diligence notes and regulated customer records. Ask the responsible reviewers to assess applicable legal requirements, client permissions and contractual controls, and assign an accountable output reviewer.
What the approved zone needs before use
Record the named owner, permitted uses, allowed data classes, retention and training/reuse arrangements, access controls, logging, review steps and incident route for each approved tool. Maintain a tools inventory and acceptable use policy. Check whether staff use personal or unmanaged accounts, and ask what led to use outside the approved route.
Set the governance model for approved AI use
Approval should also be role-specific. A marketing assistant summarising public research, a solicitor drafting client advice, and a finance team testing a cash-flow model do not carry the same data risk. The policy should say who may use which tool, for which purpose, with which data, and with what level of review.
How to turn the boundary into daily behaviour
Start with an inventory of tools already in use, including browser extensions, embedded AI features, meeting assistants, document add-ons, research tools, and public chat interfaces. Then create a short acceptable use policy that staff can actually follow. It should cover approved tools, prohibited data, verification requirements, disclosure triggers, supervision responsibilities, and incident reporting.
Require a qualified person to check AI-assisted client output for accuracy, context, confidentiality and fit with the engagement, including privilege questions where relevant. Record who checked it, what they checked and the changes made before delivery.
What evidence should be kept
Keep a light record for each meaningful AI use case: business purpose, tool name, data class, client or matter boundary, reviewer, output destination, decision made, and any incident or exception. For higher-risk use cases, add vendor due diligence, DPIA or privacy assessment notes, prompt/output samples where permissions and retention rules allow, model or tool version, and review sign-off.
Keep the record clear enough to explain the tool, data boundary, review and accountable person to a client, regulator, insurer or partner. Link it to the approval decision and any exceptions.
Conclusion
Give staff a data classification rule, a tools inventory and a clear route to approved uses. Keep the review record with the decision, and revisit the boundary when the task, tool or client instructions change.
Build a practical AI implementation plan
Need a safer AI route for your firm?
If this article reflects a live decision in your practice, the useful next step is to map the workflows, confidentiality risks, supplier controls and governance gaps before tools spread informally.