best 14 ai meeting notetakers

meeting content is not stored by any third-party vendor after processing and not accessed by one once the service completes, with the security page stating a zero-day retention policy outright. account data is deleted within 30 days of closure. no other vendor here documents retention that tightly.

on training the answer is equally flat: 'your meeting data is never used for ai model training', with vendors contractually prohibited from using it for their own. that is a blanket commitment rather than a setting to find, which distinguishes it from granola and fathom's opt-outs and from otter's absence of one.

it also carries the only named hipaa business associate agreement in this ranking, on enterprise, alongside soc 2 and gdpr, and covers zoom, meet, teams and ten-plus other platforms in 100-plus languages. the one gap: its own policy still doesn't say whether the recorder is visible to other participants or how consent is meant to be handled.

bluedot captures through a chrome extension and desktop and mobile apps rather than joining the call: 'unlike most note-taking apps, bluedot never invites a bot to your meeting or asks for calendar access'. it works on any platform including in-person meetings, and only charges for people who actually record.

the training commitment is unscoped and on the homepage: 'your data is never used to train or improve ai models', with google workspace api data separately excluded. that's stronger than avoma's and jamie's workspace-only carve-outs.

two gaps. it names processing locations — us-based analytics and german aws — but commits to no data-residency region for customer content, and retention is only a generic four-year contract clause with 'custom data retention' sold on business. and because there's no bot, nobody else in the meeting has any signal it's running, which is the trade this architecture makes. note the domain is bluedothq.com; bluedot.so is an unrelated company.

granola reads microphone and system audio on your own machine and never joins the meeting, so there's no bot in the participant list. more unusually, it doesn't keep the recording at all: 'we do not retain or store such recordings once the transcription is created'. of fourteen tools, that's the strongest structural limit on what can later leak.

the default undoes some of it. granola trains on de-identified aggregated data unless you opt out in settings, and that opt-out is only on by default for enterprise accounts. it does contractually bar openai and anthropic from training on your personal data, which is a real and separate commitment — but granola itself is included until you say otherwise.

notes are stored in a us aws vpc with no eu option, gdpr compliance handled by dpa on request. and as with bluedot, no bot means no signal: nothing in the meeting tells anyone else it's running beyond whatever you say aloud.

read notifies participants that the meeting is being measured and gives them an opportunity to opt out, which they do by typing 'opt out' in the chat. across fourteen tools that is the only product-level consent mechanism for people who didn't choose the software, and it's the reason this ranks fourth despite weaker answers elsewhere.

the design implies presence in the call rather than silent device capture, so unlike granola or bluedot there's something to notice. the mechanism does fail where meeting chat is disabled or unsupported, and the vendor says so.

the defaults are less generous. collection is automatic unless someone actively opts out, audio and derived data are kept for up to two years with custom retention sold as an enterprise+ feature, and the training position is nuanced rather than a no: google workspace data can't feed generalised models, but personalised model improvement is permitted for consenting account holders. its security page returned 404 on both attempts, so certifications are unverified.

'spinach ai never uses customer data to train ai models' is stated flatly on its security page, unscoped by data source. combined with soc 2 type 2 audited by ey, gdpr compliance and hipaa with business associate agreements available, that's a stronger posture than most tools twice its price.

the product is aimed at engineering rhythm rather than sales calls — zoom, meet, teams and slack huddles, with integrations into jira, linear, clickup, asana, confluence, notion and more — and the free starter tier includes unlimited recording, transcription and basic summaries.

the documentation gap is real and awkward given the compliance positioning. the dedicated privacy policy urls returned 404 twice, so capture method, consent handling and data residency are all unverified, and there's no default retention figure — only 'custom data retention' as a paid add-on. the zero-retention language that does appear describes its ai subprocessors' handling, not spinach's own storage of your transcripts.

sembly offers local data storage in eu data centres with content held exclusively within the eea, available as a standard option rather than gated behind enterprise. for teams with residency requirements and a normal budget, that's the differentiator, and only meetgeek's ireland-only hosting comes close.

retention is configurable in a useful way: users can set rules to automatically delete meeting content after a chosen period, with deleted content persisting in backups for up to 14 days. that's a real control, short of fireflies' zero-day default.

the training position is the weak point and it's tier-gated: 'we do not use audio, video, or text data from enterprise plan customers for model training. other plans can manage this with opt-out setting'. so below enterprise, training is on until you turn it off — the same structural gap as otter and granola. capture method is undocumented; the how-it-works page 404'd.

retention is published per tier and split by data type, which nobody else here does: free keeps transcripts three months and audio one; pro keeps transcripts a year and audio six months; business keeps transcripts indefinitely and video twelve months. that lets you choose a plan on the retention question rather than discovering it later.

on consent it does more than most: 'we alert participants when they join meetings if you are recording a meeting with our service. participants have the option to leave the meeting.' that's weaker than read ai's opt-out — leaving the meeting is a blunt instrument — but it's a real notice, and meetgeek is explicit that collecting consent remains your responsibility.

the gap is training. the closest statement is an access limitation — meetgeek says it doesn't access your recordings and transcripts unless you share them for support — which is not the same as saying it doesn't train on them, and no clause addresses training directly. hosting is ireland-only, which serves as its eu story but leaves no us option.

grain captures on the desktop without a bot, and can also run alongside zoom's own cloud or local recording. every call recorded through it includes a personalised recording disclaimer, and it leans on zoom's native recording disclosures and attendee consent reporting rather than building a parallel mechanism. reusing the platform's consent infrastructure is a sensible design most rivals ignore.

the billing model is the other good idea: you pay only for people who record, upload or import, while everyone else joins as a free viewer. for a sales team where a handful record and many review, that changes the real cost substantially.

the silence is the problem. across its privacy and security pages there is no statement about whether meeting data trains grain's models — a 'feed your ai' phrase appears with ambiguous meaning — and no retention window for audio or transcripts. hosting is us-only. soc 2 type ii is in place, but on the two questions this ranking exists to answer, grain's documentation doesn't answer either.

jamie claims eu hosting and gdpr compliance, with its vertex ai processing held in eea data centres primarily in ireland. it also claims the broadest capture scope in this ranking, covering all online platforms plus in-person meetings, and it's the only tool here exposing mcp so claude, chatgpt or cursor can query your meeting history.

the training language looks absolute — 'we do not use personal data or other user data to develop, improve, or train artificial intelligence or machine learning models' — but the surrounding context ties it specifically to google workspace api data. that makes it a scoped commitment rather than the unqualified one bluedot and spinach offer, and worth confirming directly before relying on it.

eu containment is also partial: while vertex processing sits in the eea, other named subprocessors including google, meta and cloudflare process in the us under adequacy frameworks. retention has no company-wide default — only third-party windows of twelve months at vertex and thirty days after use at elevenlabs. pricing is in euros, unlike everything else here.

fathom's privacy policy says the thing almost nobody else does: 'some states or countries require all parties to consent before recording a video or audio meeting — not just the consent of one party', and tells users to obtain that consent themselves. it doesn't provide a mechanism, but naming the obligation is more than thirteen other vendors manage.

the free tier is the most generous here — unlimited recordings and transcriptions with ai summaries, clips and search — which makes it easy to evaluate properly before paying.

training is on by default with an opt-out in account settings, and third parties including openai, anthropic and google are contractually barred from training on the data. retention is 'as long as necessary' with custom policies sold on enterprise. the security and compliance page returned 404 on both attempts, so certifications and any self-hosting option are unverified — and the domain has moved from fathom.video to fathom.ai, which older links won't reflect.

avoma reaches further than the others by capturing dialer calls through aircall, dialpad and ringcentral alongside zoom, teams, meet, gotomeeting and highfive, feeding a conversation and revenue intelligence layer with salesforce, hubspot and pipedrive integrations. free viewer seats mean only active users are billed.

the training language is narrow in a way that's easy to misread: data obtained through google workspace apis is not used to develop, improve or train generalised ai or machine learning models. that is a commitment about google's data, made because google requires it. no blanket statement covers your zoom recordings or dialer calls, and no opt-out control is mentioned.

everything is hosted in aws facilities in the usa with no eu option, though cross-border transfer can be opted out of by emailing support. retention isn't quantified — 'as long as you use the service' — with custom retention on enterprise. soc 2 type ii and annual pentests are documented, but the capture mechanism isn't stated anywhere.

the claim is the most emphatic here: 'we'll never, ever use them to train ai', stated as absolute policy rather than a default with an opt-out. alongside it, tl;dv advertises eu data hosting and storage, end-to-end encryption, gdpr compliance and soc 2 certification.

all of that comes from the homepage, which was the only page we could reach. the pricing pages returned 404 on every attempt, and no privacy or terms page could be crawled, so retention periods, consent handling, capture method and the scope of the eu hosting are all unverified.

that leaves it in an awkward position for a ranking built on reading policies: the marketing says what you'd want a vendor to say, and there's no document behind it we could open. we've ranked it on what's verifiable, which is very little, rather than on the strength of the claims.

otter's privacy policy describes using your data to improve and monitor the services, 'including training our proprietary ai technology on de-identified audio recordings and on transcriptions (which may contain personal information)'. that covers both the audio and the text, and the only opt-out documented on that page concerns the sale of personal information and targeted advertising — not training.

retention matches: information is kept 'for as long as necessary to fulfill the purposes set out in this policy, or for as long as we are required to do so by law', with no fixed period and no zero-retention option. hosting is us-only.

as a transcription product it remains strong — live captioning across zoom, teams and meet, 300 free minutes a month, and the steepest annual discount here. it is also the least transparent of the fourteen about how it actually joins a call: neither its privacy nor security pages describe the capture mechanism at all.

circleback's privacy policy says communications data is shared with 'ai model and transcription providers, strictly on a transactional basis for providing the services'. that describes a data flow, not a training position: it neither confirms that your meetings train models nor rules it out, and no opt-out is described anywhere. of fourteen vendors it's the only one that leaves the question genuinely open.

retention is similarly open-ended — data kept 'for as long as you have an open account with us or as otherwise necessary' — with no period and no zero-retention option, and everything hosted and processed in the united states.

the marketing homepage claims soc 2, gdpr, eu-us data privacy framework and hipaa coverage via a security portal, but that portal is javascript-rendered and returned no readable content, so those claims are unverified against any actual trust documentation. the product itself covers slack huddles, in-person meetings and imports of past recordings.