The setting says no training. The salesperson says enterprise. A colleague says the document is anonymized. Before relying on any of those statements, trace where the information goes and what the proposed legal task requires. Client confidentiality is not a feature label, and attorney-client privilege is not a promise a software vendor can make for every matter.
A useful review distinguishes the professional duty concerning client information, the fact-specific legal assessment of privilege, contractual restrictions, technical access, and the accuracy of any resulting work. Those questions interact, but they are not interchangeable. Removing a name does not settle all of them. Neither does keeping a tool inside the firm's account.
This guide supplies a data-flow worksheet and a hypothetical vendor review. It offers operating questions, not a guarantee of confidentiality, privilege, or compliance. ABA Formal Opinion 512 interprets the ABA Model Rules and addresses confidentiality and consent in particular generative AI uses. The responsible lawyer must assess the governing duties and facts of the proposed representation.
Make the permission decision inspectable
The legal workbook has a separate data-permission sheet because output quality and permission are different questions. A tool can produce an accurate summary from information it should not have received. A permitted tool can also produce an inaccurate summary.
Record the product, tier, contract version and configuration you actually intend to use. Check data retention, model-training use, sharing, access, storage and deletion against the permitted purpose. Preserve the evidence and its date, so a later configuration change can be evaluated rather than assumed harmless.
Separate the questions that need different evidence
“Where is the information stored?” requires provider and configuration evidence. “May we send it?” requires the organization's instructions, client relationship and applicable duties. “Does this affect privilege?” depends on the actual facts and legal context. Do not make one affirmative answer substitute for the others.
The Florida Bar's advisory Opinion 24-1 discusses retention, sharing and self-learning policies. It is expressly nonbinding guidance in its Florida context; it does not prove that every product tier behaves the same way. Source and scope.
Do not make consent carry the entire decision
Meeting capture adds questions about recording law, ethics, purpose, access and retention. New York City Bar Opinion 2025-6 addresses those issues under New York ethics rules and discusses checking transcripts. It is not a nationwide consent law or a ruling that AI use automatically waives privilege. Read the opinion.
Use the synthetic exercise before a matter
The supplied agreement permits model improvement, while the fictional customer policy prohibits training on matter documents. The exercise shows how to preserve that conflict in an exception log. It does not decide an actual client's permissions or legal position. Keep the real matter out of an unapproved trial.
Separate confidentiality from privilege before reviewing the product
Confidentiality asks what duties apply to information relating to the representation and what safeguards the proposed use requires. Privilege concerns a legal protection for qualifying communications, assessed under applicable law and facts. A document can require confidential treatment without being privileged. A communication may raise privilege questions that cannot be resolved by a vendor's security certification.
Write down which question you are assessing. Does the firm have permission to disclose particular information to this service? Who can access it? Could using the service affect a claimed privilege? Does an agreement restrict processing? What must the client understand? These questions need evidence and, where appropriate, advice from someone who can evaluate the relevant legal context.
Avoid binary product labels such as privilege-safe. They obscure differences among tasks and information. An approved tool used to edit a public statement is not the same use as sending a client's confidential strategic analysis. The same vendor can offer different features, account tiers, integrations, and retention behavior. Approval should describe the tested use and environment, rather than certify a brand name in the abstract.
In the hypothetical example, a litigation team wants to summarize a client's internal investigation materials. The vendor advertises a no-training account. Counsel identifies additional questions about attachments, diagnostic logs, support access, connected storage, and the purpose of disclosure. The team has not yet approved the use. A favorable answer to the training question is useful evidence, but it is one part of the review.
Keep the privilege assessment qualified. Do not state that any AI use waives privilege or that no-training settings preserve it automatically. Identify the communication, parties, purpose, access, and governing law for the specific assessment. If the team lacks the information needed to answer, record that uncertainty and choose another method until the relevant question can be resolved. A tool demonstration does not replace that work.
Draw the complete information path
Trace the proposed path from source to deletion. Include the device where a document starts, the upload interface, preprocessing, model provider, storage, retrieval index, output location, integrations, support systems, logs, backups, and any export. Ask who operates each component and who can access it. A single product interface can conceal several services with different terms.
Distinguish input, derived data, and output. A document may be broken into text fragments, converted through optical character recognition, indexed for retrieval, or summarized. These derived records can still contain sensitive information. Deleting the original file does not necessarily remove an index, log, summary, or backup. Ask about each relevant copy instead of assuming all representations follow one deletion control.
Record destinations in ordinary language. The document leaves the firm's storage and enters the vendor account; a connector can read selected folders; a model processes submitted text; outputs remain in the workspace. Each statement should have supporting terms, documentation, or a verified technical test. Mark unknown steps. Do not fill them with assumptions based on how another product works.
The hypothetical vendor review finds that chat content is covered by one retention setting while a file-analysis feature follows a different policy. The team also discovers that a support ticket could include an uploaded diagnostic bundle. Counsel and technical staff narrow the proposed use and document the unresolved feature. The issue is not whether the vendor is trustworthy in general; it is whether this workflow's information path is adequately understood and appropriate.
Review the output path as seriously as the upload. A summary may be exported into a broadly shared folder or pasted into a status report. The model provider might not train on the input while the firm itself distributes the output too widely. The confidentiality review therefore needs internal permissions and recipient decisions, not only vendor assurances. A restricted source can become an unrestricted summary if nobody owns that handoff.
Ask vendor questions that require inspectable answers
Ask which contractual terms govern the proposed account and feature. Obtain the relevant privacy policy, service terms, processing agreement where applicable, and product documentation. Keep versions and dates. Marketing language can orient the review, but it should not substitute for the documents or technical facts needed to assess access, use, retention, and responsibility.
Ask whether inputs or outputs are used to train, improve, evaluate, or troubleshoot systems, and what each term means in that service. Ask what controls apply by default and what changes require action by an administrator. A setting called off may affect one category while leaving another unchanged. Have the vendor identify the boundaries and exceptions instead of treating a short answer as complete.
Ask about personnel access, subprocessors, region, incident processes, deletion, backup treatment, export, and account termination. The relevant questions depend on the task and information; this list is not a universal legal standard. Involve technical specialists where counsel cannot verify an architectural assertion. Counsel can assess the legal implications of facts without pretending to know an undocumented backend.
The hypothetical team asks whether no retention includes attachment processing and troubleshooting records. The vendor supplies a feature-specific table showing an exception. The review records that exception and excludes that feature from the proposed trial. This is more useful than a yes-or-no survey scored by total affirmative answers. One consequential unresolved exception can matter more than many reassuring responses unrelated to the use.
Ask what happens when terms or configurations change. Can the firm identify the change and reassess affected uses? Who will receive notices? Which administrator can expand a connector's access? Approval should have an owner and a renewal trigger. A vendor review without change management can become stale while users continue treating it as a permanent clearance for all client information.
Preserve the vendor's supporting answer in a location the reviewer can inspect. A note saying retention checked does not reveal which feature or exception was examined. Record the document version, relevant section, date, and factual conclusion. If a technical specialist verified behavior through a test, preserve the approved test design and result without retaining unnecessary client content. Another person should be able to understand the basis of the approval after the original evaluator leaves. This is especially useful when a feature appears under a familiar product name but follows different processing terms. The reference record lets administrators identify which uses need reassessment when the feature changes. It remains evidence supporting a bounded decision, rather than a representation that the vendor's future behavior or every legal consequence is guaranteed.
Classify the information before uploading it
Describe the proposed input precisely. Is it public authority, a published agreement, a client's draft contract, personal information, investigation material, litigation strategy, or mixed content? A broad label such as legal documents tells you too little. Identify the source, purpose, recipients, and restrictions relevant to the actual package.
Minimize the information needed for the task. If the task is to improve the structure of a generic letter, a fictional example may be sufficient. If the task is to compare a particular clause against an approved playbook, the relevant text and definitions may be necessary. Do not strip context so aggressively that the output becomes misleading while claiming privacy has been solved. Data minimization and task validity require a deliberate balance.
Treat de-identification as a claim to examine. A document without names can still reveal an individual, company, transaction, or strategy through context. Unique dates, facts, counterparties, and quotations may identify the matter. Record what was removed and what information remains. An assistant can propose redactions, but a reviewer needs to inspect whether the resulting text contains identifying or otherwise protected material.
In the hypothetical investigation package, the team replaces employee names with labels but retains detailed event descriptions and job roles. Counsel finds that colleagues familiar with the organization could identify the people. The package is not treated as harmless anonymous text. The team either changes the method or completes the review appropriate to the remaining information, rather than relying on the labels as proof of protection.
Classify outputs as well. A generated timeline may combine otherwise innocuous facts into a revealing narrative. A summary may include legal advice or sensitive allegations. Assign access and storage according to the content and purpose of the output, not the fact that software wrote it. Preserve source identity so later reviewers can distinguish an allegation from a verified fact.
Assess client communication and consent for the actual use
Communication should help the client understand the proposed method and its material implications. A broad sentence in an engagement letter may not explain a particular tool, information path, or risk. The appropriate discussion depends on the use and governing duties. Avoid giving clients an abstract technology disclosure that does not tell them what happens to their information.
ABA Formal Opinion 512 addresses consent in certain uses of self-learning tools and discusses understanding terms and risks. It does not establish an identical disclosure requirement for every AI-assisted task. Evaluate the actual proposed disclosure, the client's circumstances, and applicable rules. Do not translate one opinion into a nationwide consent script.
A practical conversation can identify the task, information proposed for use, vendor or environment where relevant, material risks, review method, and available alternatives. Ask whether the client has instructions or restrictions the firm must consider. Confirm the resulting decision in the appropriate record. Consent, when needed, is not a replacement for competence, suitable safeguards, or assessment of other legal obligations.
In the hypothetical investigation, counsel explains that the proposed service would process specified materials through an external environment and that a feature-specific retention question remains unresolved. The client asks to wait and use another method. Counsel records that instruction and continues the representation accordingly. The firm's desire for a fast pilot does not convert an unresolved data flow into an acceptable client decision.
Also consider information supplied by a client through their own tools. A client may send an AI summary, record a conversation, or upload documents independently. Ask how the material was prepared and what information was disclosed, without assuming privilege was lost or all legal advice is unusable. The lawyer should assess the actual facts and any required response. An assistant's polished summary should not obscure that inquiry.
Address recording and transcription as a separate workflow
Recording a client conversation creates a different information package from summarizing an existing document. It can capture voices, incidental statements, third-party information, and context participants did not intend to distribute. The workflow may involve a meeting platform, transcription service, summarization model, calendar integration, storage, and automatic sharing. Review each part rather than importing approval from a contract-review pilot. Include who can activate the recording and who can change automatic sharing. An individual user's calendar preference can otherwise change the information path without the matter team noticing.
The New York City Bar's Formal Opinion 2025-6 discusses AI recording, transcription, client tools, confidentiality, privilege, consent, retention, and accuracy in a New York professional ethics context. It is advisory guidance, not federal recording law or a rule for every state. Recording requirements must be evaluated for the actual participants and jurisdictions.
Decide whether recording is needed. A lawyer's contemporaneous notes may serve the task without creating a complete recording. If recording is proposed, clarify participants' understanding, allowed uses, access, storage, retention, and sharing. Do not rely on a bot's visible name as proof that every relevant consent or communication issue is settled. Have the responsible person evaluate the actual requirements.
In the hypothetical client call, an unapproved assistant joins through the client's calendar settings and sends a summary to a distribution list. Counsel pauses to understand the tool and recipients and discusses the appropriate next steps. The recording question is not solved by deleting the lawyer's local copy. Other accounts, transcripts, and outputs may still exist, and the response depends on the facts.
Check transcript accuracy before using it as a factual record. Negations, names, dates, attribution, and conditional statements can change meaning. A summary saying the client approved a strategy may omit a condition or misattribute the statement. Preserve uncertainty, inspect the source where permitted, and obtain clarification when necessary. Do not allow an automatically generated action item to become authority for a legal decision.
Keep internal access narrower than the workspace
An enterprise account can contain many users who should not see the same matters. Map matter access, ethical screens where relevant, administrators, support roles, and shared folders. A workspace's existence does not answer whether a particular employee should read a client's information. Review how the assistant retrieves material across matters and whether its permissions match the user's actual authorization.
Consider retrieval features carefully. A tool may search a broad document collection to answer a narrow question. If that collection contains restricted information, the output could disclose material the user should not have accessed. Ask what permission checks occur before retrieval and at output. Test with approved synthetic examples representing different access groups, rather than exposing actual restricted matter content to discover whether controls work.
The hypothetical firm creates two artificial matter folders with different user permissions. A reviewer asks the assistant about one folder while logged in as a user lacking access to the other. The test reveals that a shared index can surface text from both. The firm pauses the retrieval feature, involves technical staff, and changes the configuration before proposing a client-data trial. A favorable vendor training policy would not address that defect.
Separate administrator capability from ordinary use. Administrators may be able to export logs, change access, enable connectors, or broaden retention. Identify which powers are necessary and how their use is reviewed. Do not describe role-based controls in a policy that nobody verifies against the actual account. Configuration evidence should support the access map.
Review output sharing. A draft may inherit a public-link setting, synchronize to a team channel, or appear in an automatically generated report. Set a release decision appropriate to the task and recipients. A human review of legal accuracy does not necessarily inspect distribution permissions, so assign both checks. The responsible people should know which one they are completing.
Preserve records and deletion limitations honestly
Decide what the firm must retain for the matter and what temporary processing copies should be removed under the approved plan. Do not equate delete chat with destruction of every copy. Ask about attachments, indexes, derived text, logs, backups, exports, support records, and retention exceptions. Record the limits the vendor can substantiate.
Coordinate deletion with applicable recordkeeping, preservation, and client obligations. A general article cannot tell you to delete a matter record simply because the AI trial ended. Nor should a firm retain every diagnostic copy indefinitely without examining the need. Legal, records, and technical staff should assess the relevant requirements and assign an owner for implementing the plan.
For the hypothetical investigation trial, the approved plan retains the lawyer's accepted analysis and identified supporting sources in the matter system. Temporary synthetic test files are removed from the trial environment according to the chosen procedure. The vendor's backup treatment is documented separately. The record does not state that all traces vanish immediately if that has not been established.
Test the procedure using permitted material. Confirm what users can remove, what administrators can remove, and what remains visible afterward. Inspect exports and connected storage locations. If deletion cannot be verified for a component, record the limitation and decide whether the proposed information should enter it. A promise written in the firm's policy does not change a service's technical behavior.
Make termination practical. The firm should know how to export necessary records, revoke connectors, remove accounts, address retention, and preserve matter continuity. Assign the tasks rather than assume the vendor automatically closes every path. A pilot can be reversible only if the team understands what ending it actually entails. Some information decisions cannot be fully undone after disclosure, which is why approval belongs before the upload.
Prepare an incident response for the information path
Describe the events that require immediate internal escalation. Examples include an unintended recipient, a cross-matter retrieval, an unapproved feature, a missing source restriction, or a vendor notice affecting data. The list should fit the actual workflow. Staff need a person to contact and a way to pause further processing without destroying records needed to understand what occurred.
Collect facts before making legal conclusions. Identify the material, account, feature, time, recipients, settings, and relevant logs. Preserve evidence through the authorized process and involve appropriate legal and technical personnel. Do not declare that privilege was waived, no harm occurred, or notification is unnecessary solely because a tool produced the event. Those assessments depend on the governing duties and circumstances.
In the hypothetical investigation workflow, a support export includes text from a client attachment. The staff member stops additional exports and reports the event. The team identifies the support recipient, reviews the terms and access, and assesses the response. The incident record distinguishes confirmed facts from open questions and names who will make the required legal and client communication decisions.
Avoid quiet workarounds. If a restricted feature is unavailable, staff should not paste the same material into a personal account to finish the assignment. Provide a manual route and a realistic service plan. A policy that prohibits unapproved tools while leaving urgent work impossible invites behavior the team will struggle to supervise.
After the response, update the data-flow map and acceptance criteria. The purpose is to prevent recurrence and understand whether the workflow remains appropriate. Do not use a single contained incident as proof that the system is safe, or one failure as proof that every AI service is unusable. Make the next decision from the particular cause, consequences, controls, and remaining uncertainty.
Evaluate an approved trial using synthetic and public material
Start with information that the firm is authorized to use and that does not require an unresolved disclosure decision. Synthetic documents can test access boundaries, source attribution, deletion, and output sharing. Public sources can test a draft summarization task. These trials cannot establish that later client-data use is appropriate, but they can expose technical and review defects before that decision.
Build test cases around the questions you need answered. One file can contain artificial sensitive fields to check whether outputs repeat them. Another can contain two conflicting dates to test uncertainty. A third can sit behind a different access group. Record the expected behavior and actual result for each. Do not create a vague success criterion such as the tool felt secure.
The hypothetical team tests a document summary with an inserted instruction telling the assistant to copy unrelated workspace content into the answer. The team checks whether the system treats that instruction as source text or follows it. The result informs its technical review and permissions choices. It does not prove that all possible malicious instructions will be resisted.
Have the intended reviewer inspect outputs using the proposed production method. If a summary cannot be traced to source locations, the review may be too burdensome even when data controls are acceptable. Accuracy, confidentiality, and usability are distinct acceptance questions. All should be examined for the task, without assuming that success on one settles the others.
Write the approval as a bounded decision: which task, information categories, feature configuration, reviewers, storage, and recipients are allowed. State exclusions and reassessment triggers. Record unresolved questions as unresolved. The proposed trial should not begin merely because a checklist contains many completed fields if the missing answer concerns a necessary condition for using the actual information.
Include a usable description of failure. If the tool returns information from an unselected matter, a user needs to know that the result should not be circulated or accepted as normal. If a transcript attributes an approval to the wrong speaker, the user needs a correction route. If the vendor announces a change to an enabled feature, an administrator needs an instruction about reassessment. These examples tie policy language to observable events. Reviewers can then identify a problem without having to decide its full legal consequences themselves. The responsible lawyer and technical team assess those consequences from the preserved facts. This division also helps avoid unsupported assurances in routine status messages. A staff member can say the feature is paused pending review without declaring that a breach occurred or that no confidentiality concern exists. The operating record should preserve the uncertainty while the appropriate people determine the response.
Use 30 days to make one informed information decision
During the first week, describe the task and information category, draw the data flow, and identify applicable instructions and restrictions. Obtain the primary vendor documents and assign legal and technical reviewers. Select an alternative method for the matter while questions remain open. The client should not have to wait for a software evaluation unless that delay is appropriate and agreed in context.
In the second week, resolve factual questions with vendor and technical evidence. Test permissions, output sharing, and deletion using approved synthetic material. Keep a record of feature-specific exceptions. Review the source package's confidentiality and any privilege questions separately from the tool's accuracy. A favorable contract term does not tell you that the generated analysis is correct.
In the third week, prepare the proposed client communication and obtain any required decision or consent through the appropriate process. Run only the trial the firm has authorized, with defined review and release steps. Record errors, unplanned copies, and questions raised by users. Pause if the use expands beyond the approved information or path.
In the fourth week, choose whether to retain the narrow workflow, revise it, or use another method. Assign an owner for terms changes, access, retention, and incident escalation. A completed vendor questionnaire is evidence for a decision, not a permanent guarantee. Record the reasons for the actual choice so another reviewer can understand it later.
If your hesitation concerns a missing vendor answer, pursue that answer. If the review is complete but you keep avoiding the client conversation or enforcing a boundary, name the behavior and schedule the action. AI and the Andes Retreat is a five-day seminar in Peru's Sacred Valley for examining such behavior and changing action. It is not legal advice, privilege certification, or an AI confidentiality training promise.
Compare your options
| Work item | Evidence to inspect | Decision or escalation |
|---|---|---|
| Input | Information category, source, restrictions, and purpose | Minimize content without making the task misleading |
| Processing | Feature-specific terms, providers, and derived records | Resolve necessary unknowns before client-data use |
| Access | Users, administrators, retrieval, and integrations | Test boundaries with approved artificial records |
| Output | Content, recipients, and storage permissions | Review both accuracy and distribution |
| Retention | Matter needs, temporary copies, logs, and backups | Document actual deletion limits and assign an owner |
| Client decision | Instructions, material risks, alternatives, governing duties | Assess communication or consent for the specific use |
Work through one ai client confidentiality and privilege decision
Complete the worksheet with legal and technical input appropriate to your proposed use. These questions help organize an assessment; they do not establish a guarantee or replace the rules and facts governing the representation.
A reflection tool for planning your next step, rather than a predictive score.
- Describe the information and task without uploading the source package.
- Draw each processing and sharing step, including derived records and logs.
- Obtain the terms and technical evidence for the exact account and feature.
- Assess confidentiality, privilege, contracts, and client communication separately.
- Test permissions and deletion using authorized synthetic material.
- Record a bounded approval or choose another method while necessary questions remain open.
A hypothetical litigation team proposes summarizing investigation materials in an enterprise account. The review finds that a file feature has retention exceptions and that a shared retrieval index can expose artificial test material across access groups. The team excludes the feature, pauses retrieval, and uses another method for the client matter while questions are resolved. A synthetic test later verifies the revised access boundary. The approval remains limited to specified information and processing steps, with a reviewer, storage destination, client decision where required, and change owner. It makes no claim that every future communication retains privilege.
Your working notes
Use this page while you think. Your notes stay in this browser tab; download them before you close it.
What does this question change for you?
A practical answer can change how you work. It can also leave a personal decision unresolved: what to learn, what to stop postponing, or which responsibility to take next. Name that decision before collecting another prediction.
Work through your next decision. Your notes stay in the browser tab, and the exercise gives you no predictive score.
If a recurring pattern keeps preventing action, AI and the Andes Retreat offers five days of practical inquiry in Peru's Sacred Valley, followed by support and accountability. Read about the seminar and assess whether that work fits your question.
Frequently asked questions
Search research and audience discussions informed these questions. Some wording has been adapted to the task covered by this guide.
Does using AI violate attorney-client privilege?
There is no product-independent answer for every communication. Assess the parties, purpose, information path, access, and governing law for the actual use. Do not assume automatic waiver or automatic preservation. The professional confidentiality review is related but distinct. This guide organizes the necessary facts; it does not determine privilege for your matter.
Are AI documents protected by attorney-client privilege?
The fact that software created a document does not establish its legal protection. Identify what the document communicates, why it was created, who receives it, and the applicable legal context. Counsel must assess the actual circumstances. A vendor label or restricted folder alone does not resolve that assessment.
Does attorney-client privilege survive when AI listens?
Assess the actual recording and processing arrangement, participants, access, purpose, and applicable law. The NYC Bar's Opinion 2025-6 discusses these questions in a New York ethics context. It is advisory guidance, not federal recording law or a universal privilege guarantee.
How does the use of AI tools affect client confidentiality?
The workflow can introduce new recipients, processing services, copies, and sharing paths. Draw the actual flow and inspect terms, feature-specific behavior, internal permissions, retention, and output distribution. Evaluate the proposed information under governing duties and client instructions. A no-training setting addresses one question without answering every access or disclosure question.
What are a lawyer's ethical obligations regarding client confidentiality?
Identify the governing professional duties for the representation and assess the information, proposed disclosure, safeguards, and necessary communication. ABA Formal Opinion 512 explains generative AI issues under the Model Rules, which are not binding nationwide law. The review should use current applicable rules rather than treating a general checklist as a compliance certification.
Can I ensure confidentiality simply by choosing an enterprise AI account?
No account label by itself establishes that the specific use is appropriate. Inspect the exact feature, contractual terms, processing path, access, logs, retention, and output recipients. Test controls with authorized material. Record a bounded decision and limitations rather than promise that the account eliminates every confidentiality concern.
How should I prepare a proposed use of confidential client information?
Describe the task and information before uploading it. Determine the minimum useful input, governing restrictions, approved environment, and required legal and technical review. Use artificial or public material for early tests when the disclosure decision remains unresolved. Preparation should preserve enough context for the task without treating minimized content as automatically unrestricted.
How should I evaluate cloud access beyond an AI training setting?
Trace storage, attachments, diagnostic records, support access, retrieval indexes, outputs, and backups. Read feature-specific terms and verify important technical facts with suitable help. The source discussion raises concerns; it does not prove that every cloud service is inappropriate. Make the decision from the actual workflow and evidence.
Can document terms remain confidential after identifiers are removed?
They can still require protected treatment. Commercial details, unique facts, and context may remain sensitive or restricted. Identify what remains and assess it under the representation, client instructions, agreements, and proposed tool path. Replacing names does not supply blanket authorization to process the document.
How should I account for de-identification work in an AI trial?
Record the inspection needed to identify and remove information, the risk of leaving context that reveals the matter, and the effect of removals on task accuracy. Compare the whole workflow with an alternative method. If useful context cannot appropriately enter the tool, choose another approach instead of uploading an inadequately reviewed package.
Should I inspect metadata as well as visible text before using AI?
Yes, assess the actual file and processing path, including relevant metadata, hidden content, attachments, and derived text. Do not assume that a visible name replacement changes every copy or identifier. The relevant technical behavior should be verified with authorized material and appropriate expertise, then included in the information decision.
How should I assess data residency for a legal AI use?
Identify the actual locations and processing providers, applicable legal or contractual restrictions, and client instructions. Obtain feature-specific evidence rather than infer location from a vendor's headquarters. This guide does not announce a universal residency rule for law firms. Record unresolved locations and ask the appropriate legal or technical reviewer what they mean for the proposed use.
What if the matter has a protective order addressing AI?
Read the actual operative order and assess its scope for the proposed information and feature. Do not infer its contents from a forum title or another matter's order. Associate the approved workflow with the relevant restrictions and source version. Escalate any uncertainty before processing material that may be covered.
What should I do if a client records a consultation and sends it to AI?
Ask what was recorded, which service received it, who can access the outputs, and what settings or copies exist. Assess appropriate advice and response under the actual facts and governing requirements. Do not assume privilege disappeared or that deleting your local copy removes every transcript. Discuss approved alternatives for future conversations.
Do lawyers have to disclose the use of AI?
The appropriate communication depends on the use, client, governing duties, and sometimes the actual tribunal. ABA Formal Opinion 512 discusses communication and consent in context; it does not create one nationwide script for every background feature. Explain the material information path and method where required or useful, and assess the actual proposed use.
Should a lawyer quietly use an unapproved AI account for work?
Do not assume a personal preference supplies authorization for client information or overrides firm and client instructions. Clarify the approved environment, proposed data, supervision, and required decisions. If the workflow blocks necessary work, escalate and use the authorized alternative. The legal assessment depends on the actual rules and facts; a general product name does not settle it.
Which information-handling errors should an AI policy make observable?
Examples include unapproved access, cross-matter retrieval, unintended recipients, undocumented feature exceptions, and unplanned copies. Define the events relevant to the approved workflow and the person to contact. Staff should preserve facts and pause appropriately without having to decide the entire legal consequence themselves. Counsel and technical staff assess the response.
Can I test legal document review before using client data?
Use authorized artificial or public material to test source attribution, access boundaries, output sharing, and reviewer effort. Such a test can reveal defects without resolving later client-data permission. Record its limits. A useful result on public material does not authorize processing a confidential matter in the same account.
Why should I obtain the primary order before relying on an AI-use summary?
A forum description or generated summary can omit scope, conditions, or current status. Retrieve the authoritative operative source and inspect its relationship to the matter. Keep its version and the question counsel assessed. The audience question supports the need for the source; it does not prove what any particular order requires.
How should I record missing information in a vendor review?
Name the unanswered feature, processing step, access, retention exception, or restriction, and identify who can supply the evidence. Decide whether that answer is necessary before the proposed use. Do not mark the vendor approved merely because many other fields are complete. A precise unresolved item lets counsel choose a narrower method or another environment.
Sources and evidence
These guides were drafted with AI assistance and checked against the linked sources. Examples labeled hypothetical illustrate a decision; they are not participant stories or measured results.
Studies answer specific questions about particular people, tasks, and tools. Their dates and limits matter. A result about task assistance is not a forecast for an entire profession.
- ABA Formal Opinion 512: Generative Artificial Intelligence Tools
- Formal Opinion 2025-6: Ethical Issues Affecting Use of AI to Record, Transcribe, and Summarize Conversations with Clients
- Florida Bar Ethics Opinion 24-1 (2024-01-19)
Where the audience questions came from
These linked discussions and comments supplied questions and wording. They do not establish the truth of the replies, verified professional identities, or how common a concern is.
- Public audience question: How should I evaluate cloud access beyond an AI training setting?
- Public audience question: Can document terms remain confidential after identifiers are removed?
- Public audience question: How should I account for de-identification work in an AI trial?
- Public audience question: How should I assess data residency for a legal AI use?
- Public audience question: What if the matter has a protective order addressing AI?
- Public audience question: What should I do if a client records a consultation and sends it to AI?
- Public audience question: Should a lawyer quietly use an unapproved AI account for work?
- Public audience question: Which information-handling errors should an AI policy make observable?
- Public audience question: Why should I obtain the primary order before relying on an AI-use summary?
- Public audience question: How should I record missing information in a vendor review?
