Source: raw/AI_Meeting_Note_Taker_-_How_to_Use_AI_Notes_Without_Compliance_Risks.md (AI Master, YouTube TsIU3aVS4KQ, fetched 2026-10-30). The video reviews one product, Fellow, and ends with “book a demo with the Fellow team” and links in the description. The transcript does not use the word “sponsored”, but treat it as a product promotion. Certification and compliance claims are the vendor’s, relayed by the presenter.
Most AI note takers were built for productivity first, with security added once legal and IT reviewed them. This review of Fellow tests the note quality briefly, then spends most of its time on governance: who can record, what is kept and for how long, what gets redacted, and how a compliance team reviews meetings. The checklist is the reusable part: it gives the questions to put to any meeting recorder before it joins calls that carry client data, health data or privileged advice.
Key Takeaways
- Note quality first. With a calendar connected, the bot read the invite and joined a test call on its own. The summary separated decisions from discussion and listed action items with owners; both follow-ups the presenter buried mid-call came back as owned tasks. Action items push into CRM and task tools.
- Botless capture. A native desktop app records system audio locally, so the other side sees no bot in the participant list. The presenter says a visible bot can derail calls with portfolio companies or opposing counsel. A mobile app captures in-person and hybrid meetings with speaker labels.
- Searchable archive with sources. “Ask Fellow” answers natural-language questions over meeting history and links each answer to the moment in the meeting, “which is what makes it checkable rather than just plausible”. Note templates per meeting type (client onboarding, investment committee) keep output consistent.
- Integrations: Outlook, Google Calendar, Zoom, Teams, Google Meet, Salesforce, HubSpot, Slack, a verified Coda connector, an API and an MCP server. A central library has role-based access; shared clips of one section inherit permissions and are logged.
- Governance controls tested:
- Recording policy at organization, team or individual level; meeting types marked always, never, or user’s choice; domain rules so calls with a given counterparty are never captured; information-barrier separation rules for firms that need them.
- Pause leaves a real gap. After pausing mid-call, the transcript had no text for that stretch (not hidden behind a permission), and the pause appeared as a timestamped logged event.
- Consent disclosures standardized by policy. In two-party-consent jurisdictions, “saying that you always mention it out loud isn’t evidence.” He would use one standard disclosure: “far easier to defend than 14 improvised versions.”
- Decoupled retention. Set source recordings and raw transcripts to a short schedule and keep AI summaries longer, which keeps decisions and commitments without a verbatim record. Zero-day retention deletes source recordings and transcripts after processing. The trade-off: once raw audio is gone you cannot check exact wording, so set it per meeting type. He does not claim it settles discoverability; that is a question for counsel.
- Redaction of names, account numbers, non-public personal information and flagged terms, automatically by keyword policy or by a reviewer, before sharing or syncing. He presents this as a stated capability, plan-dependent, that he did not test. Password-protected recaps for external sharing.
- Compliance portal: a compliance admin searches and reviews any recorded meeting regardless of individual sharing settings. Plain-language trackers (he wrote one for material non-public information and one for personal data) return matches with sub-labels. Review actions are logged, redactions stay visible, and a super-admin API exports audit and deletion logs.
- Vendor claims relayed: SSO, SCIM, role-based access; SOC 2 Type II; AES-256 encryption; GDPR and HIPAA referenced in compliance materials (confirm scope directly); “they don’t train AI models on your data.”
- His limit statement: “no tool makes your firm compliant.” It gives a compliance program controls, logs and certifications; responsibility stays with the firm. For internal stand-ups and marketing syncs it is more governance than you need.
- Who it fits: private equity, hedge funds, investment banking, fintech, legal and health tech, where one conversation can carry material non-public information, health data or privileged advice. His test: “If your security review of an AI note taker takes longer than the pilot itself”, a tool with recording rules, retention and audit logs built in is worth evaluating.
Try It
The presenter’s rollout order, usable for any AI note taker:
- Write the recording policy before choosing the tool. “A pilot without a policy just spreads recordings around faster.” Decide which meeting types are always, never or optionally recorded, and which counterparties are excluded.
- Set retention per meeting type. Keep summaries longer than raw transcripts where that makes sense; reserve zero-day retention for meetings where losing the verbatim record is worth it.
- Standardize one consent disclosure and apply it by policy.
- Give supervision two trackers, not 20. “A tracker that flags everything stops meaning anything.”
- Test the pause yourself: pause mid-call, say something, resume, and check the transcript has a gap and a logged event.
- Bring your own questions to the vendor demo: where recordings live, who can open them in six months, what you can hand a regulator, and the exact scope of GDPR/HIPAA claims.
Implementation
Tool/Service: Fellow (AI meeting notes: calendar bot, botless desktop capture, mobile app, Ask Fellow, compliance portal, API, MCP server). Setup: Connect a calendar; configure recording policy, domain rules, disclosures and retention in admin and workspace settings; grant compliance-portal access through a compliance admin. Cost: Not stated. Redaction depends on plan and configuration. Integration notes:
- The MCP server means meeting history can be queried from an agent; the same retention policy should govern what the agent can read.
- The disconnect-versus-delete distinction in What Personal Agents Keep applies here too: a short raw-transcript retention window does not remove summaries or anything already synced into a CRM.
- For sharing transcripts with frontier models outside the tool, see task-relative redaction.
Open Questions
- Pricing and which plans include zero-day retention, redaction and the compliance portal are not stated.
- How far consent-disclosure logging goes was something the presenter “pinned down in a demo”, but the answer is not given.
- Whether summaries kept under zero-day retention can be traced back to the source after deletion is not explained.
- GDPR and HIPAA scope is unconfirmed; the presenter says to confirm it directly with Fellow.
- No comparison with other note takers (Otter, Granola, Fathom, Zoom’s built-in notes) is made.
Related
- What Personal Agents Keep — Disconnect Is Not Delete — retention and deletion controls for agents
- Stripe’s Kai — a governance-first company agent
- Task-Relative Redaction for Sensitive Files — redacting before sending files to a model
- Claude Duet Safety, Privacy, and Security — consent tiers for personal data in an agent
- Connectors — how MCP servers like Fellow’s attach to Claude
- Personal Agents Compared (October 2026) — agents that read meetings and inboxes