Skip to main content
    Back to Blog
    26 min readLast reviewed:
    HIPAA
    GDPR
    Privacy
    AI Compliance
    Notetakers
    Meetings

    AI Meeting Notetakers: HIPAA, GDPR, and Privacy Compliance in 2026

    T

    Team Sonomos · Sonomos

    The Sonomos research team writes about AI privacy, data protection, and how to use generative AI safely at work.

    AI meeting notetakers have become standard infrastructure for distributed teams in 2026. Otter.ai, Fireflies.ai, Grain, Read AI, Zoom AI Companion, Microsoft Copilot in Teams, and a growing list of competitors all promise automatic transcription, meeting summaries, action-item extraction, and searchable archives of every conversation. The productivity case is real. The privacy and compliance case requires careful thought — especially for organisations in healthcare, legal, financial services, and any sector operating under European privacy law.

    This guide addresses the four questions that regulated organisations most often get wrong: consent to record, HIPAA implications, GDPR implications, and the specific risks created by AI notetakers that generic call-recording tools do not. Each statute, case and vendor position below links to its source, and the full list, with the date it was checked, is at the end. This is a description of what the laws and vendors say, not legal advice.

    What AI notetakers do that regular call recording does not

    Traditional call recording captures audio (and sometimes video). AI notetakers do substantially more:

    • Transcription. The audio is converted to searchable text, making spoken content indexable by the notetaker's systems and any integrated tools.
    • Participant identification. Most notetakers attempt to identify who said what, using voice profiles, calendar integration, or participant name inputs.
    • Summarisation and analysis. AI models generate summaries, action items, sentiment indicators, and topic extractions from the transcript.
    • Integration with CRM, project management, and other systems. Meeting content is often pushed to Salesforce, HubSpot, Notion, or similar platforms, multiplying the locations where the content lives.
    • Searchable archives. The notetaker builds a searchable database of meeting content across the organisation.

    Each of these features creates additional privacy surface area. A recording can be deleted. A summary that has been pushed to a CRM, used to improve a vendor's models, or indexed in a third-party archive is harder to erase.

    Recording consent is not an AI issue — it is a statutory requirement that predates AI notetakers, and it is now the basis of class actions against notetaker vendors.

    In the United States, federal and state wiretap laws apply side by side:

    • Federal law is one-party. The federal Wiretap Act allows a person to record a communication they are a party to, or one where a party has given prior consent, unless the recording is made to commit a criminal or tortious act (18 U.S.C. § 2511(2)(d)). State statutes can be stricter.
    • Several states require all parties to consent, or at least to be told. The statutes differ in what they cover:
      • California prohibits recording a confidential communication without the consent of all parties, with fines of up to $2,500 per violation for a first offence and possible jail time (Cal. Penal Code § 632).
      • Florida makes interception lawful when all parties have given prior consent; a violation is generally a third-degree felony (Fla. Stat. § 934.03).
      • Illinois prohibits a party from surreptitiously recording a private conversation unless all other parties consent (720 ILCS 5/14-2).
      • Maryland allows a party to record only where all parties have given prior consent (Md. Code, Cts. & Jud. Proc. § 10-402).
      • Massachusetts prohibits secretly recording a conversation without prior authority from all parties (M.G.L. c. 272, § 99).
      • Pennsylvania permits a person to intercept a communication where all parties have given prior consent (18 Pa.C.S. § 5704(4)).
      • Washington requires the consent of all participants, and treats consent as obtained when one party announces the recording in a reasonably effective way, with the announcement itself recorded (RCW 9.73.030).
    • Two states often put on the all-party list are more nuanced. Oregon requires all participants to be specifically informed before an in-person conversation is recorded, but only one participant's consent for telephone and radio communications (ORS 165.540). Michigan's statute prohibits eavesdropping without the consent of all parties (MCL 750.539c) but defines eavesdropping as recording "the private discourse of others" (MCL 750.539a), and the Michigan Court of Appeals read that wording as not reaching a participant's own recording in Sullivan v. Gray, 117 Mich. App. 476 (1982) (Varnum LLP's summary).

    The list above covers the states this guide checked against the statute text; it is not exhaustive. Violations can be criminal: California provides for jail time, Florida's offence is a felony, and Michigan's is a felony punishable by up to two years.

    The Otter.ai class action. On August 15, 2025, Justin Brewer filed a proposed class action against Otter.ai in the U.S. District Court for the Northern District of California (complaint, No. 5:25-cv-06911). The case is now consolidated as In re Otter.AI Privacy Litigation. The plaintiffs allege that the Otter Notetaker joins Zoom, Microsoft Teams and other meetings as a "silent participant", records and transcribes conversations without the consent of all participants, keeps voiceprints to identify the same speakers in later meetings, and uses conversation content to train its speech-recognition models. Otter argues it is an invited participant acting as the host's recording tool. On August 13, 2026, Judge Eumi K. Lee granted Otter's motion to dismiss in part and denied it in part: claims under the federal Wiretap Act (ECPA), CIPA § 631, and Illinois BIPA §§ 15(a) and 15(b) survived, while claims including the CFAA, California's CDAFA and the Washington Privacy Act were dismissed with leave to amend (order on motion to dismiss). That is a ruling on the pleadings, not a finding that Otter broke the law; the case continues.

    The Fireflies.ai BIPA claims. On December 18, 2025, Katelin Cruz, an Illinois resident whose voice was recorded at a nonprofit group's meeting, filed a proposed class action, Cruz v. Fireflies.AI Corp., in the U.S. District Court for the Central District of Illinois. The complaint alleges that Fireflies' speaker-recognition feature creates and stores voiceprints of every meeting participant, including people who never created an account, without the written notice, written release and public retention policy that the Illinois Biometric Information Privacy Act requires (Epstein Becker Green's summary of the complaint). These are allegations; no ruling on them was found as of this review.

    BIPA defines "biometric identifier" to include a voiceprint (740 ILCS 14/10). Before obtaining one, a private entity must tell the person in writing what is being collected, for what purpose and for how long, and receive a written release; it must also publish a retention schedule that destroys the data when the purpose is met or within three years of the person's last interaction, whichever comes first (740 ILCS 14/15). Liquidated damages are $1,000 per negligent violation and $5,000 per intentional or reckless one, and repeated collection of the same identifier from the same person by the same method now counts as a single violation (740 ILCS 14/20). In the Otter case, the court found the Illinois plaintiffs' voiceprint allegations sufficient to proceed.

    The practical baseline US organisations commonly adopt:

    • Disclose at the start of every recorded meeting that the meeting will be recorded and transcribed by an AI tool. An announcement does not by itself satisfy every all-party statute — some require consent rather than notice — but Washington's statute, for example, treats a recorded announcement as consent.
    • Review voice-fingerprinting or biometric speaker identification against BIPA and similar state laws before turning it on.
    • For external meetings with participants in unknown locations, apply the strictest all-party standard that could apply.

    In the European Union and UK, the GDPR and UK GDPR apply to the recording and processing of meeting participants' personal data (voice, name, statements). The analysis differs from US wiretap law. Consent under the GDPR must be "freely given" (Article 4(11)), and the European Data Protection Board considers it problematic for employers to rely on employees' consent because of the imbalance of power, saying that for most processing at work the lawful basis "cannot and should not" be consent, though it does not rule consent out entirely (EDPB Guidelines 05/2020 on consent, paras. 21-22). The other bases are listed in Article 6(1); for recording business meetings, legitimate interests under Article 6(1)(f) is the one most often discussed, and it is the basis that requires weighing the organisation's interests against the participants' rights.

    HIPAA and AI notetakers

    Any meeting that includes protected health information (PHI) — patient names, diagnoses, treatment plans, scheduling details, insurance information — is subject to HIPAA's privacy and security requirements when a covered entity or business associate creates or handles it. PHI is individually identifiable health information in any form or medium (45 CFR 160.103). This covers:

    • Telehealth appointments recorded for clinical documentation purposes.
    • Clinical team meetings that discuss specific patients.
    • Administrative meetings that include patient names or identifiers.
    • Any meeting where a healthcare worker describes a patient's case.

    The BAA requirement. A vendor that creates, receives, maintains, or transmits PHI on behalf of a covered entity or business associate is a business associate (45 CFR 160.103). A covered entity may let it handle PHI only after obtaining satisfactory assurance, documented in a written contract (45 CFR 164.502(e)) that meets the requirements of 45 CFR 164.504(e), including returning or destroying PHI at the end of the contract. That contract is the Business Associate Agreement (BAA). A notetaker that transcribes a meeting containing PHI fits that definition, so sending PHI to one without a BAA in place puts the covered entity outside the rule.

    ToolBAA position (vendor's own documentation, November 2026)Notes
    Zoom (including AI features)Zoom signs BAAs with healthcare customers: accepted at checkout on Pro, through sales on Business, Business Plus and Enterprise, or enabled by an admin on healthcare plansAI features are available to BAA customers, but Zoom says certain AI features may not be available under a BAA
    Microsoft Teams and CopilotMicrosoft's BAA is available through its Online Services Data Protection Addendum by default to covered entities and business associates; Microsoft Teams and Microsoft Copilot are listed as in-scope commercial servicesCopilot in Teams meetings needs an add-on Copilot license
    Otter.aiOtter announced HIPAA compliance on July 10, 2025; its pricing page lists "HIPAA compliance (add-on)" on the Enterprise planAsk Otter's sales team for the BAA and confirm which features it covers
    Fireflies.aiBAA available on Enterprise with Private Storage; Fireflies' setup guide says a plan downgrade or storage change immediately disables HIPAA modeEnterprise plan only
    GrainGrain says it "offers a BAA on enterprise plans"Grain's security page does not address HIPAA; confirm with Grain
    Read AIRead AI's privacy page says a "BAA available on request"Confirm the plan and configuration Read requires

    These are the vendors' own statements as of the review date at the end of this post. Product offerings and BAA terms change, so confirm with the vendor before relying on them for a HIPAA program. For how one of these BAAs applies in practice, see our Microsoft Copilot HIPAA analysis; for a broader vendor checklist, see the HIPAA-compliant AI checklist.

    Minimum necessary. Even with a BAA, covered entities and business associates must make reasonable efforts to limit PHI to the minimum necessary for the purpose, with exceptions such as disclosures to a provider for treatment (45 CFR 164.502(b)). Recording every meeting and archiving it indefinitely in a searchable AI database is hard to reconcile with that standard unless there is a documented clinical or administrative purpose for it.

    Keeping PHI out of the notetaker. Some organisations use AI notetakers for administrative meetings but route clinical meeting content through a different, more controlled path. Where AI notetaking is used for clinical meetings, configuring the tool to delete transcripts after a defined retention period and to not integrate with general CRM or productivity systems reduces the PHI footprint. Retention controls are a plan feature at some vendors: Otter's pricing page lists "Custom data retention" in its plan comparison, and Fireflies lists custom data retention as an Enterprise feature (security).

    GDPR and AI notetakers

    For meetings involving people in the EEA — employees, clients, counterparties, or any natural person — the GDPR applies to the AI notetaker's processing.

    The core GDPR requirements:

    • A processing agreement with the notetaker vendor. An AI notetaker that records and transcribes meetings for your organisation processes personal data on your behalf. Article 28(3) requires that processing to be governed by a contract or other legal act covering, among other things, documented instructions, security, sub-processors, help with data subject requests, and deletion or return of the data at the end. This is the data processing agreement (DPA).
    • Lawful basis. See the consent discussion above: legitimate interests is the basis most often discussed for recording business meetings and requires a documented balancing test, and the EDPB considers employee consent problematic in most workplace processing.
    • Transparency. Article 13 requires participants to be told, at the time their data is obtained, who the controller is, the purposes and legal basis, any recipients (including the notetaker vendor), any transfers outside the EEA, how long the recording is kept, and their rights. A short spoken announcement is unlikely to cover that whole list on its own, which is why organisations pair it with a link to a privacy notice.
    • Data subject rights. Participants can request access to, and a copy of, transcripts in which they appear (Article 15), and erasure where one of the grounds in Article 17 applies. The controller can only honour these requests if it knows where every transcript is stored, including copies the notetaker has pushed to third-party tools. Our guide to GDPR data subject rights and AI covers how these requests work in AI systems.
    • International transfers. Transfers outside the EEA need an adequacy decision or appropriate safeguards such as the Commission's standard contractual clauses (Article 46). Since July 10, 2023, personal data can flow from the EU to US companies that participate in the EU-US Data Privacy Framework (European Commission). Zoom, Fireflies and Read AI say they participate in the framework; Otter's DPA relies on the standard contractual clauses (sources in the table below).
    • Biometric data. Voice profiles used to recognise who is speaking can be biometric data, which the GDPR defines as data from specific technical processing of physical or behavioural characteristics that allows unique identification (Article 4(14)). When processed to uniquely identify a person, biometric data is a special category under Article 9(1), and processing is prohibited unless an Article 9(2) exception, such as explicit consent, applies.
    • DPIA. Article 35 requires a data protection impact assessment for processing likely to result in a high risk. Article 35(3)(c) names "systematic monitoring of a publicly accessible area on a large scale", which does not describe a meeting room. The relevant guidance is the Article 29 Working Party's DPIA guidelines, which list systematic monitoring and data about vulnerable data subjects (naming employees as an example) among nine criteria, say that processing meeting two criteria will in most cases require a DPIA, and give "a company systematically monitoring its employees' activities" as an example where one is likely required (WP248 rev.01). An organisation-wide rollout that records and AI-summarises employee meetings is a strong candidate under those criteria.

    DPA availability by tool (vendor's own documentation, November 2026):

    ToolGDPR DPA
    Otter.aiThe Terms of Service (effective September 19, 2025) include a Data Processing Attachment (Appendix 1) for data within the scope of the GDPR or UK GDPR, with the EU standard contractual clauses (module two) and the UK Addendum
    Fireflies.aiA public Data Processing Addendum (last updated March 6, 2026); Fireflies says it is self-certified under the Data Privacy Framework, with standard contractual clauses as a fallback
    ZoomThe Global Data Processing Addendum is incorporated into Zoom's terms; Zoom says it uses SCCs and participates in the Data Privacy Framework
    Microsoft Teams and CopilotThe Microsoft Products and Services Data Protection Addendum
    GrainNo public DPA found on Grain's site as of this review; Grain's security page says it complies with the GDPR. Ask Grain for its DPA
    Read AIRead AI's privacy page says it acts as a processor under Article 28 and offers a DPA to eligible Workspace customers in the EU and EEA

    As with the HIPAA table, confirm directly with the vendor. These are summaries of vendor documents, not legal opinions. For a wider view of the GDPR questions AI tools raise, see GDPR compliance for ChatGPT, Claude and Gemini.

    The specific privacy risks of AI notetakers

    Beyond consent and regulatory compliance, AI notetakers create privacy risks that are different in kind from passive recording:

    1. Verbatim transcripts of privileged conversations. Attorney-client privileged discussions, healthcare consultations, M&A negotiations, and whistleblower reports all typically occur in meetings. If those meetings are being transcribed and stored in a third-party AI platform, the privileged or confidential content is now in a third-party system, potentially with retention settings beyond the organisation's control. How courts treat AI tools and privilege is still developing; see AI and the work product doctrine for the current cases.

    2. The integration proliferation problem. AI notetakers push summaries and action items to Slack, email, CRM, and project management tools. A single meeting involving a client's personal data may generate several downstream records across several different systems. Erasure requests become very difficult to honour when the personal data has propagated this broadly.

    3. Long-lived searchable archives. Unless someone sets a retention rule, a notetaker builds a persistent, searchable database of recorded meetings. Retention controls are often an admin feature: Grain lets admins set workspace-wide auto-deletion (security), and Fireflies lists custom data retention under Enterprise (security). For regulated organisations, a large archive can bring discovery and litigation-hold obligations for content that may not have been intended as a business record.

    4. Participant identification across meetings. Some tools build voice profiles that link statements across different meetings over time. The Otter and Fireflies complaints both centre on this. Under BIPA a voiceprint is a biometric identifier, and under the GDPR voice data used to uniquely identify someone is special-category biometric data, each with its own consent and retention rules.

    5. Training data questions. Vendors' terms differ on whether meeting content may be used to improve their models. Otter's Terms of Service permit machine learning on user content to improve the service and the use of de-identified data for training; the Otter plaintiffs allege their conversations were used to train its models. Fireflies' DPA permits deidentified data to be used to improve its products. Zoom says it does not use customers' audio, video, chat or other communications-like content to train its own or third-party AI models (Zoom BAA article), and Read AI says model-improvement contribution is opt-in and off by default (privacy). Read the current terms for the plan you actually use.

    A practical compliance checklist for AI notetakers

    Before deployment:

    • Identify which meetings will be recorded (all? only certain types?) and which will be excluded (clinical consultations? legal calls? M&A discussions?).
    • Select a tool and plan with BAA coverage (for HIPAA workloads) and a GDPR DPA (for EU personal data).
    • Sign the BAA and DPA before first use.
    • Assess whether a DPIA is needed against the WP248 criteria (its own employee-monitoring example meets at least two).
    • Review the tool's data retention defaults and configure them to match your retention policy rather than the vendor's default.
    • Review and restrict integrations — disable push to CRM, project management, and other tools unless there is a documented business purpose and the integration is covered by appropriate data-sharing agreements.

    For each meeting:

    • Disclose at the start that the meeting is being recorded and transcribed by an AI tool, naming the tool.
    • Where an all-party consent statute applies, obtain affirmative confirmation from all participants before starting the recording.
    • In meetings with EU data subjects, provide the Article 13 transparency information (by reference to a linked privacy notice, if the notice is current).
    • For meetings involving PHI: confirm the tool operates under your BAA for this meeting type.
    • Exclude highly sensitive meetings (privilege, M&A, clinical consultations beyond the scope of the BAA) from AI notetaking unless specifically appropriate.

    Ongoing:

    • Establish a process for honouring data subject access and erasure requests against meeting transcripts.
    • Review vendor terms annually for changes to training-data use, retention defaults, and integration partners.
    • Add the notetaker vendor to your vendor risk management program, including annual SOC 2 report review.
    • Maintain litigation-hold procedures that cover meeting transcripts in addition to email and documents.

    Frequently asked questions

    Does Zoom AI Companion require a separate BAA for HIPAA?

    Yes, in the sense that a BAA is not part of Zoom's standard terms: it has to be executed. Zoom's support article says Pro customers can accept the BAA at checkout, Business, Business Plus and Enterprise customers go through sales, and admins on healthcare plans can enable it in plan management. A separate "Zoom for Healthcare" subscription is not the only route. Zoom says its AI features are available to customers with a BAA in place, but that certain AI features may not be available to them, so confirm the specific AI Companion features you plan to use (Zoom).

    Not without disclosure. California Penal Code § 632 prohibits recording a confidential communication without the consent of all parties (§ 632), and § 631 bars willfully reading or trying to learn the contents of a communication in transit "without the consent of all parties to the communication" (§ 631). In the Otter.ai case, the court held that the California plaintiffs had adequately alleged a § 631 violation against Otter, the vendor whose bot joined their meetings (order). A recording that relies only on the host's consent leaves that question open.

    Is an AI meeting transcript PHI under HIPAA?

    It depends on what it contains and who holds it. HIPAA does not define "medical record"; it defines PHI as individually identifiable health information held or transmitted in any form or medium (45 CFR 160.103). A transcript, summary or audio file in which a clinician discusses an identifiable patient's diagnosis or treatment, created or maintained by a covered entity or business associate, is PHI whatever its format. If it is used to make decisions about the patient, it can also fall within the "designated record set", which includes records used to make decisions about individuals (45 CFR 164.501).

    Does HIPAA set a retention period for AI notetaker transcripts?

    Not directly. The six-year rule often quoted for HIPAA, in 45 CFR 164.530(j), applies to the documentation the Privacy Rule requires (policies and procedures, required written communications, and documented actions), not to transcripts or medical records as such. Retention periods for clinical records, financial-services records and general business records come from other rules, and the notetaker's retention settings can be configured to match whichever schedule applies. Indefinite retention makes data-subject rights and litigation holds substantially harder to manage.

    Our employees use Otter.ai on personal devices. Is that a problem?

    Potentially a significant one. An employee using a personal Otter account sits outside the organisation's contract with Otter: the organisation has no admin control over retention, integrations or sharing, and Otter's HIPAA offering is tied to the Enterprise plan (pricing), so a personal account would not be covered by a BAA. If the meetings involve EU personal data or PHI, the organisation may still be responsible for that processing as the controller or covered entity, without the agreements and settings it would normally require.

    The bottom line

    AI meeting notetakers are a legitimate and useful productivity tool — and they create real privacy compliance obligations that are easy to overlook when a tool is adopted one user at a time. The baseline for regulated use is straightforward: a BAA for any meeting that could include PHI, a GDPR DPA and a documented lawful basis for any meeting involving EU data subjects, disclosure that meets the strictest all-party consent law that applies, and retention settings configured to match policy rather than vendor defaults.

    The deeper risks — integration proliferation, long-lived archives, biometric voice profiles, and the difficulty of honouring erasure requests across multiple downstream systems — require more deliberate configuration choices at deployment time. The organisations that get this right treat the notetaker's settings and integrations as part of a data-minimisation strategy: capture what is needed, route it only where it belongs, and retain it only as long as required.

    Sources and primary references

    Reviewed November 5, 2026. Statutes and regulations are cited to official legislature sites, Cornell LII, legislation.gov.uk's copy of the GDPR as adopted by the EU, or FindLaw's statute text where the official site could not be read. Vendor statements are cited to the vendor's own documentation as it read on that date; vendors revise these pages, so confirm before relying on them.

    Sonomos publishes this blog and makes Locke, which detects and masks sensitive information on your device before it reaches an AI model. How Locke works

    This article is general information about privacy and AI, not legal advice, and reading it does not create a solicitor–client or attorney–client relationship. Regulatory requirements turn on your specific facts and jurisdiction and change over time. Consult qualified counsel before relying on anything here, and check the publication and last-reviewed dates above. We may revise posts, but we do not undertake to keep any post current as the law changes.

    Reference pages for this post