Granola

The Terms establish that content and information Authorized Users submit ('Customer Data') is owned by the Customer (e.g., the employer or workspace creator), not by Granola. Individual users store, access, and control their own notes, which are private by default. Granola does not claim ownership of user-submitted inputs.

The Privacy Policy describes AI-generated summaries, transcripts, and notes as outputs of the service and treats them as part of the user's/Customer's data, but the documents do not contain an explicit clause assigning ownership of AI outputs to the user. Outputs are labelled AI-generated and users are told to review them for accuracy. The silence on formal output ownership assignment lowers confidence.

Granola may use de-identified, aggregated user inputs and outputs for its own AI model improvement, and on Free and Business plans this is on by default (opt-out available). Third parties such as OpenAI and Anthropic are contractually prohibited from training on user data. A significant caveat: de-identified data already incorporated into models cannot be removed, and historical data used before opt-out cannot be guaranteed excluded.

Notes and transcripts are retained indefinitely by default unless a retention policy is configured (configurable auto-deletion is an Enterprise feature). Users can delete their account and data in-app or by email, and audio recordings are not retained after transcription. However, no deletion SLA is specified, and de-identified data in models is retained indefinitely and cannot be extracted.

Granola states it does not sell or share Personal Data for cross-context behavioral advertising and does not disclose data for third parties' independent marketing without consent. It shares data with Service Providers (e.g., AWS, Stripe), parties the user authorizes, and in corporate transactions or legal requests. The presence of targeted advertising language and broad Service Provider sharing, combined with the transfer to third parties in mergers/acquisitions, warrants moderate caution, though core sharing is disclosed and purpose-linked.

Users have explicit, in-app opt-out mechanisms for model training (Settings → Preferences → Data & sharing), marketing communications (unsubscribe/STOP), and can withdraw consent. Enterprise plans allow admin org-wide training opt-out and are opted out by default. These are concrete, disclosed opt-out controls that go beyond mere policy language.

Granola holds SOC 2 Type II certification (with a report available via its Trust Center under access request) and claims GDPR/UK GDPR compliance backed by a DPA incorporating SCCs. It explicitly denies ISO 27001 and FERPA, and offers HIPAA/BAA only to Enterprise customers. ISO 42001, NIST CSF, CCPA/CPRA are referenced only as aspirational or partially addressed (CCPA modeled), and EU AI Act compliance is claimed without third-party attestation. Because the strongest certification (SOC 2 Type II) is genuinely attested but other universal-baseline frameworks are claimed or absent, this rates YELLOW.

Granola discloses that its AI is assistive and does not make binding automated decisions, provides AI-generation notices where required, and offers usage analytics and a SOC 2 report for enterprise auditing. However, the documents provide limited transparency into specific model behavior or model-level explainability, and enterprise auditing focuses on security controls rather than model interpretability.

Granola discloses encryption at rest and in transit (via AWS), firewalls, VPC setups, anti-virus/malware protection, a Trust Center, dedicated security page, and a vulnerability disclosure program with monetary rewards (though no formal bug bounty). Local device processing and no meeting bot reduce attack surface. No breach history is disclosed in the supplied text. There is no built-in 2FA/MFA (delegated to identity providers), a minor weakness.

There are material differences between tiers. Enterprise users are opted out of model training by default and get admin-enforced controls, configurable retention/auto-deletion, SSO, and HIPAA/BAA availability. Free and Business tiers have model training on by default (with individual opt-out), no configurable auto-deletion, and are not eligible for HIPAA/BAA. This creates meaningfully higher risk on non-Enterprise tiers for sensitive data.

The documents indicate 'Materials flagged for safety, security, or policy review' may be among data sources, and Granola may 'directly step in' where there is a credible risk of harm, but there is no explicit blanket statement that staff routinely read user prompts/notes. Admins cannot view individual private notes. The policy is largely silent on the specific circumstances of human access to note content, lowering confidence.

The Privacy Policy discloses that Granola may share data with law enforcement, governmental entities, or regulatory bodies in response to valid legal requests or judicial process, and in bankruptcy/merger scenarios. No pending litigation or specific government-request transparency reporting is disclosed. Governing law is California with venue tied to the Agreement.

Granola collects significant PII (name, email, IP address, device IDs, location, professional data, social profiles) plus recordings and transcriptions of communications — a highly sensitive category given meeting content can contain SPI. Payment data (last 4 digits) is processed via Stripe. The categories are clearly inventoried with disclosed sharing recipients. Because communications content and potentially special-category data may be captured (with disclosure and user responsibility placed on the user), YELLOW is appropriate.

The Privacy Policy is effective 8th September 2026 and the User Terms were last updated 19th December 2025 — both well within 12 months of the 2026-09-22 analysis date. The policy demonstrably covers the AI capabilities the product ships: AI processing, model training, third-party model providers (OpenAI/Anthropic), integrations (Zoom, MCP), calendar sync, and automated processing rights. Coverage is strong and current.

Three consumer documents were supplied (User Terms, Security/Privacy FAQ, and Privacy Policy). They are broadly consistent on training, third-party bans, and audio non-retention. A minor tension exists regarding model-training default state: the FAQ says Business plan users are on by default and must opt out individually, while the Privacy Policy summary emphasizes opt-out and Enterprise default-off — reconcilable but requiring careful reading. No CRITICAL contradictions were found.