AI Meeting Notes for Lawyers: Consent, Confidentiality, Privilege, and Accuracy

AI meeting notes can save a lawyer from reconstructing a client call from memory. They can also create a verbatim record where none existed, send confidential facts to a vendor, misstate advice, or preserve material that later becomes relevant to discovery.

The right question is not “May lawyers use an AI note taker?” in the abstract. It is whether a lawyer has evaluated the particular tool, matter, client, recording law, data flow, output, and retention policy—and then supervised the workflow.

Recent bar guidance gives lawyers a concrete framework. This article focuses on United States practice and is general information, not legal or ethics advice. Duties and recording laws vary by jurisdiction.

What the newest ethics guidance says

In Formal Opinion 2025-6, the New York City Bar Association addressed AI tools that record, transcribe, and summarise conversations with clients.

Its core themes are practical:

obtain the client’s consent before using the tool;
consider whether recording is tactically wise, not merely technically convenient;
understand how the provider stores, uses, trains on, shares, and deletes data;
take reasonable steps to protect confidentiality and privilege;
independently review transcripts and summaries for accuracy;
acquire enough technical understanding to use the tool competently;
supervise lawyers and nonlawyers using it; and
address the risks if a client brings their own AI tool into the conversation.

The ABA’s Formal Opinion 512 provides the broader generative-AI framework: competence, confidentiality, communication with clients, supervision, candour, and reasonable fees. The State Bar of California’s AI resources likewise emphasise existing professional duties rather than treating AI as an ethics-free category.

These opinions are persuasive guidance, not a uniform national rule. Check the rules and opinions governing the lawyer, firm, client, and proceeding.

Client permission and recording-law consent overlap, but they are not identical.

Professional communication with the client

A lawyer may need to explain why the tool is being used and enough of its risks and benefits for the client to make an informed decision. Relevant facts can include whether audio is retained, whether a third party receives it, whether data may train a model, who can access notes, and whether the lawyer will place the result in the client file.

The discussion should happen before capture starts. A line in a long engagement letter may not adequately address a new tool or unusually sensitive conversation.

State recording and privacy law

US jurisdictions differ on consent to intercept or record communications. The applicable law may depend on where each participant is located, the technology used, whether there is a reasonable expectation of privacy, and other facts. Multi-state and international calls deserve a specific analysis.

A visible bot or platform icon is not necessarily informed client consent, and a client’s ethics consent does not automatically resolve statutory recording rules. A conservative firm policy is to give explicit notice, obtain affirmative agreement where appropriate, and provide an unrecorded alternative.

Confidentiality: map every recipient, not just the app name

Model Rule 1.6 requires lawyers to protect information relating to a representation. For an AI meeting workflow, “the vendor” may actually include several organisations and systems:

the meeting platform;
the recording or audio-capture layer;
speech-to-text and speaker-identification providers;
an AI model provider;
cloud storage, analytics, support, and backup vendors;
integrations that copy notes into a CRM, case system, email, or chat; and
firm administrators and device-management personnel.

Before use, review the contract and the actual configuration. Ask whether customer content is used for training, whether a human can access it, where it is processed, how long backups persist, what subprocessors are involved, what security controls exist, and whether deletion is verifiable.

A public consumer account and a negotiated enterprise deployment of the same brand may have materially different terms. Firm policy should identify approved configurations, not merely approved product names.

Privilege: local storage helps, but it is not a privilege guarantee

Attorney-client privilege is fact-specific. It generally depends on a qualifying confidential communication for the purpose of obtaining or providing legal advice, and it can be affected by disclosure, waiver, exceptions, and jurisdiction-specific doctrine.

Using a third-party service does not automatically destroy privilege, nor can a vendor promise that privilege is preserved. The defensible approach is to minimise unnecessary disclosure, select providers under suitable confidentiality terms, restrict access, and analyse unusually sensitive matters separately.

Local processing can reduce the number of third parties that receive the communication. That supports confidentiality, but it does not establish every element of privilege or prevent waiver through later sharing.

Lawyers should also distinguish privilege from work-product protection and from the broader ethical duty of confidentiality. They overlap, but they are not interchangeable.

Recording creates a new record—and possibly new risk

The New York City Bar opinion urges lawyers to consider the tactical wisdom of recording. That warning deserves its own decision point.

A full recording may capture:

speculation, strategy, or emotional reactions absent from ordinary notes;
statements by non-clients or about third parties;
inconsistencies that later need explanation;
confidential business, health, financial, or family details;
side conversations before or after the intended meeting; and
material subject to a preservation duty or discovery request.

Do not keep audio by default merely because storage is available. Decide whether the matter needs a recording, a transcript, a reviewed summary, or only a short human note. Align retention with the engagement, records policy, legal holds, client instructions, insurance requirements, and applicable law.

Deleting material after a preservation duty arises can create serious problems. Retention automation therefore needs a legal-hold override and clear ownership.

Accuracy: treat AI notes as a draft, not the record

Speech recognition and generative summaries fail in different ways. A transcript may confuse speakers, names, citations, dates, numbers, or accented speech. A summary can remove uncertainty, merge separate issues, invent consensus, or assign advice to the wrong person.

A lawyer should independently review output before relying on it. The standard should rise with the consequence:

Internal memory aid: spot-check against recollection and source audio where retained.
Client follow-up: confirm advice, deadlines, and responsibilities before sending.
File memorandum: review the complete context and identify it as an AI-assisted draft where the firm’s policy requires.
Court, regulator, transaction, or discovery use: verify against authoritative sources and applicable evidence rules.

Do not let a summary silently replace a signed agreement, court order, official transcript, or lawyer-approved advice. If notes will drive deadlines, enter those deadlines into the firm’s controlled docketing process.

A law-firm approval checklist

1. Classify the matter and conversation

Exclude or escalate categories such as criminal defence, investigations, trade secrets, health information, minors, family matters, highly sensitive transactions, or conversations subject to special protective orders.

Identify where participants are located, assess recording and privacy law, explain the tool to the client, document agreement, and offer a realistic no-recording path.

3. Complete vendor diligence

Review training, retention, subprocessors, data location, security, incidents, support access, deletion, export, account termination, and contractual confidentiality. Reassess material product or policy changes.

4. Configure for minimum exposure

Turn off unneeded recording retention, model training, public links, sentiment analysis, voiceprints, broad integrations, and automatic sharing. Use the smallest attendee and administrator access groups.

5. Define human review

Assign a responsible lawyer, specify which facts must be checked, and prohibit use of unreviewed output for advice, deadlines, filings, evaluations, billing narratives, or client commitments.

6. Govern the record

Decide whether audio, transcript, and summary enter the client file. Set retention and legal-hold rules for each. Make correction possible without erasing the history of a material amendment.

7. Train and supervise the team

Train lawyers, paralegals, assistants, contractors, and intake staff. Cover approved tools, prohibited matters, consent scripts, incident escalation, and what to do when a client or opposing party activates their own assistant.

When the client brings an AI bot

Lawyers need a policy for assistants they did not select. Before discussing substance:

identify who enabled the tool and for what purpose;
ask what it records, retains, and shares;
decide whether the lawyer and client understand the confidentiality implications;
consider the presence of other clients, third parties, or jointly represented people; and
pause or remove the tool if the risk cannot be resolved.

Silence should not become accidental acceptance. Meeting invitations can state the firm’s policy in advance, while the lawyer still confirms the actual setup at the call.

A narrower data path is easier to defend

Summit records system audio without joining the call as a bot and can perform transcription and summarisation locally on a Mac. For a law firm, that can reduce disclosure of raw client conversations to meeting bots and AI vendors. Summit also supports optional cloud and self-hosted model providers, so firms must configure and approve the pathway they intend to use.

No architecture guarantees confidentiality, privilege, accuracy, or compliance. Local files still require device encryption, access control, backups, malware protection, retention, legal holds, and secure deletion. The advantage is narrower: on-device processing can reduce third-party exposure and give the firm more direct control over where meeting content goes.

For a legal-specific workflow overview, see Summit for legal teams.

The practical rule for lawyers

Use an AI meeting assistant only when the benefit justifies creating and processing the record. Obtain appropriate consent, map every data recipient, preserve confidentiality, avoid categorical privilege claims, minimise retention, and have a lawyer verify the result.

The most important control is not the disclaimer attached to the summary. It is a supervised process that decides when the tool should be off.

Primary guidance

Try Summit on macOS to evaluate a local-first, bot-free meeting workflow for your firm.