Shadow AI in a law firm: where do client documents go?
A practical review of AI use outside a law firm’s agreed workflow: accounts, client documents, extensions, records and responsibility.
Article language: English

In this article
Shadow AI means using AI tools outside an organisation's agreed workflow. In a law firm, that might mean a personal account, a browser extension or a meeting bot. Start the review with specific tasks and documents, not just the question, “Do you use AI?” A tool working in the shadows rarely introduces itself that way. More often it is simply a convenient shortcut in everyday work.
People may understand the question differently. Some think only of chat tools; others include every suggestion in an editor. A task-based review makes the distinction clearer: summarising case files, translating correspondence, editing an agreement or taking meeting notes. This helps reveal how information actually moves.
What should the review establish?
Define its scope first. Are you looking at the whole firm, one team or a particular process? Assign an owner and identify the people who will assess technical, professional and legal questions. Record who may see the findings. The review report itself may contain information that needs restricted access.
For each use, establish four things: who uses it, on which account, for what task and with which materials. Add the product name, plan, integrations and date checked. A label such as “Copilot” or “writing assistant” does not identify a service and its terms well enough.
Record established facts and gaps separately. “We have not confirmed whether the extension reads the whole page” is a specific question to resolve. A general “low risk” rating without an explanation gives no sound basis for the next decision.
How can you ask the team useful questions?
Give examples of tasks and ask people to walk through a normal working day. Who pastes part of a document into a tool? Who attaches a whole file? Does a system retrieve content from email or storage automatically? Does a meeting produce a recording or transcript? Start with uses and move to product names afterwards. If accounts do not explain the process, choose a disclosed scope for observing normal work, without repeating operations on client files. The guide to observing work (in Polish) discusses agreeing the scope and separating observations from interpretation.
Explain what will happen to the answers. If every mention of a tool leads to an unpredictable reaction, the review may describe the rulebook rather than practice. You need conditions for a factual conversation and a clear procedure for an event that calls for urgent action.
Use a short form, but leave room for explanation. A closed list of popular brands can miss tools embedded in other services. Ask about add-ons, trial accounts and tools used for only one kind of matter as well.
How do you trace one document?
Choose approved test material and recreate the task without sending more client documents. Record where the file is stored, which tools take part and who receives the result. Check exports, conversation history and automatic link sharing.
A useful flow record covers:
- Start: storage location and who is authorised to read it.
- Transfer: manual paste, attachment, integration or automatic retrieval.
- Processing: product, account type and settings that have been checked.
- Output: document, message, transcript or another record.
- Remnants: history, copies, logs and temporary files.
- End: deletion or retention under the agreed rule.
The same application can appear in several flows. Record them separately when their sources or access differ. Only then can you see which setting affects a particular task.
What should you check in service terms and settings?
Separate use of data for model training from history retention, administrator permissions and integration access. These are different questions. For example, Microsoft documents interaction retention and administrative handling separately from model-training rules. See its official description of data and privacy for Microsoft 365 Copilot.
Read the documentation for the plan and sign-in method actually used. Check providers involved in delivering the service, history settings, export options and how access is withdrawn. Record the version or date on which you read the terms. A general marketing page does not establish which agreement governs an account.
When material contains personal data, assess its scope and purpose too. The EDPS describes guidance for EU institutions updated on 28 October 2025. Its announcement addresses, among other things, the purpose of processing and the division of roles and responsibilities. It provides context for questions, not a ready-made assessment of a law firm's obligations; those depend on its own professional and legal circumstances.
How should you treat anonymisation and document extracts?
Removing a name may leave an address, transaction description, distinctive facts or a combination of details that identifies someone. In a trial, check the entire material submitted, including the file name, comments and metadata. Do not assume that shortening a document resolves every confidentiality question. If a tool prepares the copy, test its automatic detection and removal of data on artificial material, distinguishing the operation's result from an assessment of whether further use is permissible. The article on the limits of automatic AI anonymisation (in Polish) examines what such a test can miss.
Prepare examples of material allowed in particular tools. A person doing the task should know what they may submit and where to raise a question. “Be careful” leaves too much to individual interpretation.
If an unapproved flow is found, preserve the information needed to assess it. Decide next steps under the firm's procedure. Mass deletion of history before understanding the situation can make it harder to establish what happened.
What might one register entry look like?
Exercise example: a team summarises extensive correspondence in a browser-based tool. Its register names the product, organisational or personal account, type of material and person performing the task. It separately records whether that person attaches a file or copies an extract. This describes a flow; it is not an assessment of an actual firm.
The “to confirm” field might include the extension's reading scope, access to history and storage of the result. The review owner assigns someone to check each question. Once checked, record the source and date of the answer. If the flow is approved, the instructions should state the conditions of use and when it needs reassessment. This provides a record to revisit if the plan or way of working changes.
How do you set priorities without giving a whole application one label?
Assess the particular flow, type of material and scope of access. As a working distinction, separate drafting with fictional data, work with an approved document and an unauthorised transfer of case material. These situations call for different questions and different decision-makers. A popular product's name alone does not say what was read or saved.
Exercise example: an employee uses an extension while editing a document. First establish whether they supplied an extract or the extension had broader reading access. Then check the account, integration and confirmed storage locations. Do not infer from an extension icon that the entire document was sent. Equally, do not treat the absence of a visible attachment as proof that nothing was transferred.

Record the established action, material category, period, confirmed sources and urgent open questions on an initial record. The people qualified for the firm's situation assess confidentiality, professional obligations and next steps. This article offers a way to collect facts; it does not set a reporting deadline or automatically classify an event.
What if the team cannot remember exactly how a tool was used?
Do not ask people to repeat the action by sending case files again. Preserve available information about product, account and time, and establish what evidence can be obtained without spreading the material further. Restrict access to the register itself. A screenshot of conversation history may contain more than the person managing the review needs.
Distinguish an operator's recollection from confirmed records. An entry such as “probably a summary of one document” should remain labelled as an account to check. Stating uncertainty helps plan the right action. Adding detail just to complete a table makes later assessment harder.
If sources disagree, preserve the discrepancy and assign an owner. Do not close the matter after finding the first convenient answer. An AI data-flow register helps separate input, connection, output and remaining records without copying all confidential content into the report. The guide to recording AI data flows (in Polish) discusses what to include.
How do you prepare an approved alternative for the actual task?
The employee still needs to do the work that led them to the tool. Once you understand why they used it, propose a concrete route: approved input, account, checks and storage location. A list of prohibited brands alone does not explain how to prepare notes before an urgent meeting.
Choose one bounded case and run your pilot against agreed AI evaluation criteria. The guide to setting criteria for an AI pilot (in Polish) discusses how to prepare them. Check usefulness of the output, correction effort and data-flow conditions separately. If the permitted scope prevents the task from being done, record that. The answer may be another way of working, not an automatic expansion of access.
Add examples of permitted input and situations requiring a question to the instructions. Mark them as exercise material. A new team member should be able to distinguish missing information from information they may independently provide to the service.
How do you accept a fix and check the work after an operator leaves?
Check an account or configuration change using test material and the appropriate role. Preserve confirmation of access before and after the change where permitted. Do not end the task with “the administrator was informed”. State what has been done and what still awaits a decision or confirmation.
When the responsible person changes, establish who owns the connections, history and outputs. Prepare the access and confidentiality materials needed for the handover. Disabling an employee's account and ending every flow need not be the same action; confirm the scope in the services used.
For the handover of code, accounts and confidential material, see the AI agent project handover guide (in Polish).
How do you maintain a simple procedure for reporting new uses?
Ask for the task, data category and expected result without attaching case material to the request. Identify the person who assesses it and how an answer will be returned. If the team cannot find current instructions, organise the firm's knowledge. The guide to organising company knowledge (in Polish) discusses sources, owners and review dates.
Preserve the reason for approval and the conditions for a new review in a project decision register. A change to the plan, integration or material category can then trigger a specific assessment rather than automatically relying on an old approval.
The project decision register guide (in Polish) also distinguishes a proposal from an approved decision.
When reviewing the instructions, ask a new person to explain where they would raise a question and what material they may prepare for a trial. If they need unwritten team knowledge, improve the document. A short exercise tests whether the procedure can be followed without involving more client documents.
After identifying actual practice, prepare rules for specific tasks rather than product names alone. The guide to a company AI-use policy (in Polish) shows how to record permitted inputs, oversight and exceptions. Reinforce the rules through your team’s training, using safe material to practise recognising errors and raising questions. The guide to training a team to check AI answers (in Polish) discusses how to prepare that practice.
Include AI-generated meeting notes in the review. This flow can sit outside a normal chat list and include recording, transcription and later sharing. Each stage needs an understood scope and responsibility. For the distinction between a transcript, a summary and an approved decision, see the AI meeting notes article (in Polish).
How do you turn findings into an action plan?
Divide the outcome into concrete decisions. One tool may need an account change, another restricted integration access, and a third replacement with an approved solution. Give each item an owner, deadline and way to confirm completion. Distinguish a recommendation from a change already made.
Add a simple route for reporting new uses. The team needs to know how to request approval for a tool and how long the assessment will take. Without that, the next urgent task may lead back to the old shortcut.
Start with an AI tool register, then establish the boundaries for agent work. The responsible agent workflow guide (in Polish) discusses task boundaries, access and approval. The Aeneia Audit offer describes its general scope for reviewing AI use and data flows. Discuss the firm’s needs and possible work before agreeing the scope.
To plan that inventory, the AI tools article (in Polish) covers owners, usage and tools outside central records.
Before starting the review, gather materials for an AI audit. Map each significant flow in your data-flow record, including storage locations and the evidence for confirmed settings. For an approach to documenting a flow, see the AI data-flow record guide (in Polish).
For organising evidence and controlling access to it, consult the AI audit preparation article (in Polish).










