Compare the architecture. The summaries already converged.
By Fazit
Comparing AI notetakers by their features is comparing them by the thing they no longer differ on. In 2026 every serious tool writes a clean summary, extracts action items, and drafts a follow-up. The comparison that still means something is architectural: where in the pipeline does a durable artifact of your conversation get created, and who holds it afterward.
The features are a tie
Fathom, Otter, Fireflies, tl;dv, and Granola all produce good notes. Model quality is now a commodity input; the note quality gap between the leaders is small and shrinking. If a comparison ranks these tools on summary quality, it is ranking noise. So set the output aside and look at the pipeline that produces it.
Compare the pipeline instead
Every notetaker is a data pipeline with three questions attached to it. Does your audio leave the device? Where is it turned into text? And what durable artifact of the conversation survives the process. That last column is the whole liability, because a durable artifact is what gets subpoenaed, breached, trained on, or inherited in an acquisition.
AUDIO OFF-DEVICE? TRANSCRIBED PERSISTENT ARTIFACT Fathom yes, upload vendor cloud recording kept in cloud Otter yes, upload vendor cloud recording + trained model Fireflies yes, bot vendor cloud recording + voiceprint tl;dv yes, bot vendor cloud video recording Granola yes, stream 3rd-party APIs note; raw audio deleted after Fazit no on device note; audio never written
Five of the six rows create a persistent artifact on someone else's infrastructure. One does not. Everything below is why each row reads the way it does.
Fathom
Strength. A genuinely generous free tier, a polished experience across Zoom, Meet, and Teams, and the enterprise trappings buyers ask for: SOC 2 Type II, HIPAA, SSO. It is cross-platform where Fazit is Apple Silicon only. If cost and coverage are the constraint, this is the honest pick.
Fatal weakness. The recording lands in Fathom's cloud, and its policy permits using de-identified customer data to improve its own models unless you find the opt-out. In April 2026 Fathom was acquired by QuestionPro, which means the corpus of everything it ever recorded changed corporate ownership in an M&A transaction. That is not a hypothetical about a vendor someday changing hands. It already happened.
Where Fazit wins. There is no server-side recording to inherit. Nothing was stored, so nothing changes hands when the cap table does.
Otter
Strength. Fast real-time transcription, a searchable company-wide archive, a long track record, and aggressive pricing. If meeting notes are a shared knowledge base for a whole org, Otter has scale on its side.
Fatal weakness. Otter is the defendant in a federal class action (Brewer v. Otter.ai) alleging it recorded conversations without all-party consent and trained its speech models on user recordings; its policy describes training on de-identified recordings. A motion to dismiss was argued in May 2026 and is under submission. Otter is simply the litigated version of the risk the entire cloud category carries.
Where Fazit wins. You cannot train a model on a recording that was never created. Fazit's audio lives in a RAM ring buffer and is consumed by an on-device model, so there is no recording to subpoena, leak, or feed into a training set.
Fireflies
Strength. The deepest integration ecosystem here: CRM field sync, Slack, unlimited-transcription tiers, conversation analytics. That makes it sticky for revenue teams that want pipeline automation, not just notes.
Fatal weakness. Fireflies faces biometric-privacy litigation in Illinois (including Cruz v. Fireflies.AI Corp., filed December 2025) alleging voiceprint collection without consent, and its recording bot is now flagged by Google Meet as a third-party security risk that can be blocked from joining. A centralized store of voiceprints is precisely the asset BIPA statutes were written to govern.
Where Fazit wins. No voiceprint is extracted, no bot needs to be admitted or gets flagged, and there is no central store of meeting data to breach, because there is no central store.
tl;dv
Strength. Strong recording and playback UX, coaching and conversation analytics, and EU-headquartered positioning that reads well to European buyers who want a video artifact of every call.
Fatal weakness. A video recording is more data than an audio one, not less, and it is still a visible-bot cloud recorder. Free-tier recordings auto-delete after three months (a retention limit, not a privacy guarantee), and despite the EU branding tl;dv is described as still working toward full GDPR compliance. Cloud recording under an EU flag is still cloud recording.
Where Fazit wins. For a European regulated professional, a pipeline that transcribes in volatile memory and never writes an audio file to a sound carrier is a materially stronger position than a promise of compliant storage. Under German law (§201 StGB), “never created” and “created then deleted” are different legal facts. That argument is in Why “Never Records” Is Not Marketing.
Granola
Strength. Granola is the one tool that shares Fazit's instinct: no bot in the room, capture the audio your own device already hears, and do not retain it. It is SOC 2 Type II, the notes are excellent, and on privacy it is genuinely better than every bot-based tool above. If Fazit did not exist, Granola would be the right answer.
Fatal weakness. Granola is bot-free, but it is not on-device transcription. To turn audio into text it streams your call to third-party services (Deepgram, AssemblyAI, OpenAI, Anthropic) and deletes the raw audio after transcription. Your conversation still transits the public internet to four subprocessors. Deletion is a promise about what happens after the audio arrives on someone else's server, not a guarantee that it never arrived.
Where Fazit wins. Fazit's speech model runs on the device. Audio is captured into a fixed-size RAM ring buffer, a local model consumes a snapshot, and the note is written to your vault as Markdown. The audio never leaves the machine, so there is no subprocessor to list, no transfer mechanism to disclose, and no “deleted after transcription” step to trust, because there is no transcription server to delete it from.
audio_retained: false in its frontmatter, not as a compliance checkbox but as a description of the execution path.Why this is hard to copy
The cloud vendors built their economics on storing data and training models. Re-architecting to RAM-only would destroy the value proposition they sell today. Granola is closer in instinct but still depends on external transcription for reliability and model quality. Moving that on-device is not a settings toggle; it is a different product. The pipeline column is the differentiator precisely because it is the expensive one to change.
So compare on the column that is hard to fake. The summary is a tie. The recording is the decision.