Is Granola’s bot-free recording better for privacy? It avoids a visible bot in the meeting — which also removes the disclosure your counterparts would otherwise get, moving that obligation to you.
Correction to earlier profile draft
An earlier draft of this profile incorrectly described Granola as Mac-only and as having "raw audio that never leaves the user's device." Both statements are inaccurate. Granola supports macOS and Windows as of early 2026 (iOS uses temporarily cached audio for post-meeting processing). Audio is captured locally on the user's device, then transmitted to Granola's cloud for transcription, then deleted after transcription. The architecture is meaningfully better than Otter.ai or Fathom because no visible bot joins meetings and audio is not persistently stored — but it is not a fully local-first model. This corrected profile reflects the accurate architecture.
Plain-English risk rating: 4 of 5 (Free/Business, training on by default and no disclosure to other participants) / 2 of 5 (Enterprise Workspace, training off by default and admin-enforced)
Granola is a fast-growing AI meeting-notes tool (1M+ users by late 2025; $125M Series C at $1.5B valuation early 2026) that has carved out a privacy-positioning niche by avoiding the visible-bot architecture that has put Otter.ai, Fireflies.ai, and others into class-action litigation. The architecture is bot-free — Granola captures system audio directly from the user's device rather than joining the meeting as a participant — which structurally avoids the consent-disclosure question that drives the Otter/Fireflies litigation. Audio is transmitted to Granola's cloud for transcription and then discarded; no audio recordings persist. A previous version of this rating was labelled "pending litigation outcome". Granola is not a defendant in the Otter.ai or Fireflies proceedings, and its bot-free architecture structurally avoids the consent question driving them, so that label implied a pending outcome against Granola that does not exist and is withdrawn (14 August 2026). The 4 is unchanged and rests on band 4's weak-disclosure limb.
The risk picture is less straightforward than the marketing implies. Training on customer data is on by default on Individual and Business plans, with opt-out only available at the Enterprise tier. SOC 2 Type 2 certification (achieved July 2025) is meaningful baseline assurance. The bot-free architecture is genuinely a structural privacy improvement but it introduces a separate disclosure question: meeting participants on the other side of the call have no visible indicator that AI transcription is happening, which shifts the consent compliance burden onto the Granola user rather than the platform.
Recommended for
- Sole proprietor on Mac or Windows: Free tier (25 lifetime meetings) for evaluation, Individual at $18/month for regular use. Disclose to meeting participants that you are using AI transcription before recording starts.
- Small team (2-10 people): Business at $14/user/month for team features and shared notes. Recognise that training opt-out is not available at this tier — your team's transcripts contribute to Granola's model improvement.
- Regulated industry: Enterprise tier (starting $35/user/month) with model training opt-out, custom data residency negotiation, and explicit policy on which meeting types are appropriate for Granola transcription. The SOC 2 Type 2 certification is the floor; sector-specific regulatory review is still required.
- The honest answer for most 1-10 employee businesses: Granola is a genuine privacy improvement over Otter.ai for individual professionals who can disclose AI use to their meeting counterparts. The bot-free architecture is the differentiator. For teams handling client-sensitive content where training opt-out matters, the Enterprise tier at $35+/user/month is the right floor — the Business tier at $14/user/month is comparable in cost but does not include the training opt-out.
Critical pre-deployment warning (silent recording disclosure)
This is the structurally important fact about Granola's bot-free architecture: because no visible bot joins the meeting, your meeting counterparts have no indicator that AI transcription is happening. This is a privacy advantage for the Granola user (no awkward "recording bot has joined" notification, more natural conversation flow) but a disclosure compliance challenge.
Corrected 3 September 2026. This paragraph previously stated that all-party consent is required for recording in "New Zealand under the Privacy Act 2020". That was wrong twice over, and it had been published here since this profile went up. New Zealand is a one-party consent jurisdiction: under section 216B(2) of the Crimes Act 1961, a person who is a party to a private communication may lawfully record it without the other participants agreeing. The Privacy Act 2020 does not create a consent requirement for recording either.
What New Zealand law does require is different, and it still falls on you rather than on Granola. Recording a meeting collects personal information, so information privacy principle 3 of the Privacy Act 2020 requires you to tell people that you are collecting it, why, and who will receive it. That is a notification duty rather than a consent duty — but a recorder with no visible bot makes it unusually easy to breach, because nothing else in the meeting signals that collection is happening. Principle 4 additionally requires that collection not be unreasonably intrusive, and covertly recording colleagues can breach the employment duty of good faith.
Separately, if any participant is sitting in a genuine all-party consent jurisdiction — California, Washington, Florida, Illinois, Maryland, Massachusetts, Montana, Nevada, New Hampshire or Pennsylvania — their law may reach the call wherever you are. So the practice we recommend does not change; only the reason for it does. Granola's silent operation shifts the entire disclosure burden onto the user. For any external meeting: verbal disclosure at the start ("I'm using an AI tool to take notes during our call") plus a meeting-invite line item disclosing AI use.
For internal meetings within your own organisation, your employee handbook should cover AI transcription in meetings. For external client meetings, disclosure is both ethical practice and likely legally required.
Data retention default
- Audio recordings: Captured locally, transmitted to Granola's cloud for transcription, deleted after transcription. Granola does not persistently store meeting audio.
- Text transcripts and AI summaries: Stored in Granola's infrastructure under standard data retention
- LLM provider retention: Per Granola's contracts with model providers (Anthropic primary; specifics evolve)
- Account deletion: Removes notes and transcripts; given the no-persistent-audio architecture, there is no parallel audio deletion required (a structural advantage over cloud-transcription competitors)
- GDPR right to erasure: Supported via individual note deletion combined with the no-persistent-audio architecture
Training opt-out
Free, Individual, Business — TRAINING ON BY DEFAULT. This is a correction to commonly-stated marketing claims. Granola uses customer transcripts to improve its AI features on the non-Enterprise tiers. Opt-out is available only at the Enterprise tier.
Enterprise — NO TRAINING ON CUSTOMER DATA BY DEFAULT per the Enterprise contract terms. This is the structurally important upgrade for any team handling client-sensitive content.
The Phase 2 version of this profile incorrectly stated that no-training was the default across tiers. The corrected position: no-training is only the default at Enterprise.
Zero Data Retention availability
- Granola operates zero-retention by architecture for raw audio (deleted after transcription)
- For text transcripts and summaries, the no-training Enterprise contract is the relevant control
- LLM provider zero-retention contracts in place per Granola's published security documentation
Plan tiers and pricing (as of early 2026)
| Tier | Price (USD) | Training default | Suitable for |
|---|---|---|---|
| Free | $0 | On by default | Personal evaluation; 25 lifetime meetings |
| Individual | $18/month | On by default | Individual professionals (verify opt-out availability per current settings) |
| Business | $14/user/month | On by default | Small teams; training opt-out NOT available |
| Enterprise | $35+/user/month | Off by default | Larger orgs needing no-training contract |
Jurisdiction
- Primary processor: Granola (UK-based company, London headquarters)
- Cloud infrastructure: AWS
- Third-party AI subprocessor: Anthropic primary
- SOC 2 Type 2 certified (July 2025; audit completed in three months due to the no-persistent-audio architecture reducing scope)
- GDPR-compliant; UK Data Protection Act 2018 jurisdiction
- Subject to additional jurisdictional requirements based on user residency
Breach history (public incidents and active litigation)
30 July 2026 — Chamberlain v. Granola, Inc. and Granola Labs Ltd. (active proposed class action)
Case No. 3:26-cv-07926-EMC, US District Court for the Northern District of California, filed 30 July 2026 on behalf of a proposed nationwide class and a California subclass.
The complaint alleges that Granola captures both microphone and system audio without disclosing its presence to other meeting participants by default, and then uses those recordings to train its own AI models by default. It pleads seven claims, including intrusion upon seclusion, the federal Electronic Communications Privacy Act, and California’s Invasion of Privacy Act sections 631 and 632 — the same statutes at issue in the Otter.ai litigation. CIPA provides statutory damages of USD 5,000 per violation or treble actual damages.
Two features of the complaint matter for a small business deciding whether to use this tool. It quotes Granola’s own marketing, which presents the absence of a visible bot as a selling point, to argue that the lack of disclosure is a deliberate design choice rather than an oversight. And it notes that the people whose voices are captured are frequently not Granola account holders, and so have no setting to adjust even where an opt-out exists.
Nothing has been proven. No response had been filed at the time of writing and no court has ruled. An allegation is not a finding.
Why this moved the rating on 3 August 2026: Free/Business went from 3 of 5 to 4 of 5. The training default is confirmed by Granola’s own documentation independent of the lawsuit, and the no-disclosure-by-default design is confirmed by Granola’s own marketing copy. Enterprise Workspace stays at 2 of 5, because training is off by default and admin-enforced there.
13 August 2026 — the Otter ruling, and why it reaches beyond Otter
In In re Otter.AI Privacy Litigation, No. 25-cv-06911-EKL (N.D. Cal.), the court decided Otter’s motion to dismiss on 13 August 2026, granting it in part and denying it in part. The federal wiretap (ECPA), California Invasion of Privacy Act and Illinois biometric claims survived and proceed.
This is not a Granola case and nothing in it decides anything about Granola. It is recorded here for one reason. Otter had argued it was merely a tool acting for the meeting host; the court declined to accept that at the pleading stage, holding that Otter could be treated as a third-party eavesdropper because it retained the conversations and used them for its own commercial purposes, including training its models.
Chamberlain v. Granola pleads several of the same statutes and describes the same underlying practice — capture without disclosure to other participants, and training on what is captured by default on Free and Business. The reasoning in the Otter order is therefore available to plaintiffs against similarly situated products. Whether any court applies it to Granola is a question for a court, and none has.
No rating has moved and none should be inferred from this entry. Granola stays at 4 of 5 (Free/Business) and 2 of 5 (Enterprise Workspace). What changed on 13 August is not a fact about Granola, but the standing of a question this profile already answers under Training opt-out: whether the vendor uses what it records for its own purposes. Nothing has been proven on the merits in either case. Added 3 September 2026.
No major public breach of Granola infrastructure as of May 2026. The company is relatively young (founded 2023; rapidly growing) and the no-persistent-audio architecture genuinely reduces the breach blast radius compared with vendors that store raw recordings.
Note on the bot-free architecture and the litigation environment: Superseded 3 August 2026: Granola no longer sits outside this litigation wave. Chamberlain v. Granola was filed on 30 July 2026 — see above. Until then, Granola’s silent operation had avoided the class-action litigation affecting Otter.ai (In re Otter.AI Privacy Litigation, consolidated October 2025) and Fireflies.ai (Cruz v. Fireflies.AI Corp., filed December 2025, voluntarily dismissed 11 March 2026 without any responsive pleading), and the reasoning below is retained because the Chamberlain complaint attacks precisely the architectural assumption it rested on. The reason is structural: Granola's architecture does not generate biometric voiceprints of non-account-holder meeting participants (because the user-side audio capture is treated as belonging to the user, not as recording-of-third-parties). Updated 3 September 2026: that ruling has landed. On 13 August 2026 the court held that Otter could be treated as a third-party eavesdropper, on reasoning that turns on what the vendor did with the recordings rather than only on who consented — see the entry above. The exposure-by-association point stands and is no longer conditional: Granola's silent operation means the user, not the platform, is the consent-compliance actor, and that user could become a defendant.
Category-level risk: The no-persistent-audio architecture is a meaningful structural improvement over cloud-transcription competitors. The remaining risks are: (a) training defaults on non-Enterprise tiers, (b) the silent-recording disclosure compliance burden falling on users, and (c) the general category risk that AI notetakers face during the ongoing legal uncertainty about two-party consent.
What this means in plain English for SMB owners
Three honest takeaways:
- Granola's bot-free architecture is genuinely better than Otter.ai's for privacy-conscious individuals — but the disclosure compliance burden is real. No visible bot means no automatic disclosure to your meeting counterparts. Corrected 3 September 2026: we previously called New Zealand a two-party-consent jurisdiction and attributed that to the Privacy Act 2020. Both were wrong — under s216B(2) of the Crimes Act 1961 a participant to a conversation may lawfully record it. What does fall on you (not Granola) is the Privacy Act 2020 duty to tell people you are collecting their information and why — a notification duty, not a consent one, and one a silent recorder makes unusually easy to breach. Covertly recording colleagues can also breach the employment duty of good faith, and any participant sitting in a genuine all-party consent state brings that law to your call. Verbal disclosure at meeting start plus an invite line item is the baseline practice.
- Training is on by default below Enterprise, and turning it off is left to each individual user. Free, Individual and Business tiers train on your transcripts unless the user finds the setting and switches it off; there is no single admin switch below Enterprise, and Granola’s own policy states that data already incorporated into a model is not removed, so an opt-out is forward-looking only. Enterprise Workspace is opted out by default and the setting is admin-enforced. For sole proprietors using Granola for personal note-taking, this may be acceptable. For any team handling client-sensitive content where no-training matters, the Enterprise tier at $35+/user/month is the right floor — do not assume the cheaper Business tier has no-training defaults.
- The no-persistent-audio architecture is the strongest structural privacy claim in the meeting-AI category. Combined with the SOC 2 Type 2 certification (achieved July 2025), Granola is one of the most defensible meeting-AI choices for individual professionals on macOS or Windows. The migration path from Otter.ai during the Otter litigation pendency is straightforward; budget for the disclosure-discipline change as part of the migration.
Sources
- Granola privacy and security documentation (granola.ai/security, verified 2026-05-24)
- Granola blog: AI notetaker privacy compliance for product research SOC 2 and GDPR (March 31, 2026)
- Granola Series C funding announcements (early 2026)
- TL;DV: Granola AI Review (May 2026)
- Max Productive: Granola AI Review 2026 (February 2026)
- Textify Analytics: Granola AI Future of Meeting Notetaking (March 2026)
- BuildBetter: Best Granola Alternatives for Private Meeting Notes 2026 (March 2026)
- MeetJamie: Granola AI Pricing 2026 (December 2025)
- MeetingNotes: Fellow vs Granola AI comparison (January 2026)
- Feisworld: Granola AI Review 2026 (January 2026)
Related on AI Leakage
- Compare all 29 AI tools in the risk directory — see how Granola stacks up against the rest, tier by tier.
- Take the 5-minute “Am I Leaking?” check — a personalised view of your business’s AI exposure.
- Check a prompt before you paste it — our free Data-Safe Prompt Rewriter.
- Shadow AI vs AI leakage — why even approved tools like Granola can leak data.
- Get plain-English AI Leakage Alerts — we email you when an AI tool you use changes its data policy or has an incident.
- Get the free AI Acceptable Use Policy template — a plain-English policy with the tool-by-tool risk guide built in.