Fireflies.ai

Remove the models and what is left is a bot that sits in a meeting and stores audio, which nobody buys. Every capability any tier is sold on is inference: transcription across more than a hundred languages, summaries, action item extraction, an assistant answering questions across every meeting held, search down to the sentence, topic and task detection, sentiment analysis, question detection, talk time analytics, topic trackers, drafted email replies and voice agents.

The free tier's headline is transcription, so even the unpaid product is model output. The recording, storage, sharing and integration layers around it are delivery mechanics for that output rather than independent value, and the vendor's own naming treats the product as an assistant rather than a recorder. This is consistent with the whole lane, where no record has ever cleared the removal test, and it is the first record graded in this batch that does not. A buyer here is buying inference and should evaluate accuracy accordingly.

The product acts unattended in other people's meetings and in a user's mailbox, and the controls over it are reserved for the top tier. Autonomy is real and several layered: the notetaker joins meetings automatically from a calendar without anyone admitting it each time, voice agents operate, a live assist layer runs during calls, and an email assistant triages an inbox and drafts replies.

Against that the administrative controls are enumerated but positioned high, with a rules engine, a super admin role and audit logs all sitting at the quoted tier, user groups at the tier below, and single sign on with directory synchronisation only at the top. Two lower level controls do exist and are worth crediting: the notetaker's display name can be changed, and a documented route exists to use the product without connecting a calendar at all. Nothing published bounds what a drafted reply may say or when an agent hands back. Ask what an administrator can restrict below the top tier, and what review sits between a drafted reply and sending.

A precisely documented machine surface over a completely undisclosed model layer. The published side is strong for a product of this shape: an application interface is available on every tier including the free one, a public reference documents it down to individual named queries with anchors, and a credit system meters the assistant layer with allowances stated per tier and its own published explanation of what a credit buys. On the undisclosed side there is nothing at all.

No provider, model family or version is named for transcription, summarisation, the question answering assistant, the voice agents or the drafted email replies, and nothing describes how accuracy is measured or what happens when a transcript is wrong. That gap matters more here than on most records, because the output is a record of what people said that will be relied on later as if it were fact. Ask which models perform transcription and summarisation, and how a customer corrects and audits an inaccurate transcript.

A very large adoption claim and no measurement anywhere near it. The pricing page states use across more than a million companies, which is the largest scale claim recorded in this batch, and nothing accompanies it: no definition of a company, no counting method, no date and no distinction between a free signup and a paying account.

Four testimonials carry a full name, a job title and a named employer, which is proper attribution, and all four are qualitative statements from small companies with no figure in any of them. Customer logos are displayed and a customer section exists which was not reached on this pass. Independent review platform badges are shown for category leadership.

For a product whose value proposition is transcription quality, the absent measure is conspicuous: no word error rate, no accuracy figure by language, and no benchmark against any alternative, despite accuracy being the thing a buyer is actually purchasing. Ask for measured transcription accuracy by language and the basis of the company count.

No outreach is conducted from here and one newer layer sits at the edge of the question. The core product records and summarises meetings that the customer is already holding, so there is no prospecting, no sending and no recipient list, and the sending obligations that this axis exists to examine do not arise. Graded in the middle of the band with that scope stated rather than at the floor, since the floor asserts an undisciplined sending product and this does not send.

The edge case is the assistant layer that triages an inbox and drafts replies, which operates inside the user's own mailbox on correspondence already addressed to them rather than initiating contact, so the obligation still sits with the user's own mail platform. Nothing published addresses whether drafted replies are marked as machine assisted to the recipient. Ask whether a recipient of a drafted reply can tell it was generated, and what review step precedes sending.

Real instruments exist, were not read, and the control that matters most for third parties is sold as an upgrade. A privacy policy, terms of service, a dedicated page for the health information regime and a trust centre are all published and none was retrieved on this pass, so retention periods, lawful bases and the rights exercise route are unestablished and nothing here asserts they are absent.

European data protection is stated as covered on every tier including the free one, which is better than gating it. What is established from the pricing matrix is the shape of the gap. This product holds recordings and transcripts of people who are not its customers, including whoever the customer happened to be meeting, and the three controls governing how long that material is kept, where it sits and who touched it, being custom data retention, private storage and audit logs, are all reserved to the quoted top tier. A bulk delete exists below that. Ask what the default retention period is on lower tiers and how a third party participant requests erasure.

No purchased corpus, and a provenance question that is about consent rather than licence. There is no contact database, no enrichment and nothing brokered: the material this product processes is the customer's own meetings, uploaded files and calendar, so the usual chain of sources, licences and resale permissions does not apply and most of the axis is out of scope. Graded in the middle of the band with that stated.

What does sit in scope is that a meeting transcript is not solely the customer's data. It contains the recorded speech of every other participant, frequently people at other companies who have no relationship with this vendor, and the lawful basis for capturing and storing their voice and words rests entirely on the customer having obtained consent, with nothing published describing how the platform supports, records or evidences that. Ask what record exists that participants consented to recording, and whether that record is retrievable per meeting.

The core mechanic runs inside three companies' meeting platforms and no standing with any of them is stated. The notetaker joins calls hosted on the major conferencing services, which is not the vendor's infrastructure and not the customer's, and nothing published names a partner programme, a marketplace listing, an approved application status or any conformance position with those platforms, each of which sets its own rules about automated participants and recording.

The exposure is real and undisclosed: if a platform restricts third party bots, the product's primary capture method stops working for every customer at once. Two things keep this off the floor. The platforms are named plainly rather than described through euphemism, and the vendor ships a documented alternative that captures through a browser extension without any bot joining, so a customer is not wholly dependent on bot admission. That extension is distributed through the official browser store. Ask what standing the notetaker holds with each conferencing platform and what happens if bot access is withdrawn.

The commitment this axis exists to find is made, and it reaches one step further than most vendors take it. The vendor states on its own pricing page that customer data is never used for artificial intelligence training or for any purpose outside the customer's direct business needs, and separately that a zero retention position is imposed on the model providers it uses for meeting content, which addresses the downstream question almost nobody in this index answers: not just whether the vendor trains on your calls, but whether the third party model it sends them to keeps them.

Data ownership is stated as explicitly set out in the terms of service, and a mode exists at the top tier that retains only transcript and summary rather than the recording. What holds it below the top band is where the commitment lives and how it is worded.

The training statement sits on a marketing panel rather than in a document a buyer can hold, no audited management standard backs it, and the vendor's own two formulations differ, one covering data and the narrower one covering personal data, which is the distinction that decides whether a whole transcript is in scope. Ask for the training commitment in the agreement, worded to cover transcripts and recordings rather than personal data alone.

Two published features cut directly against the one signal that tells other participants they are being recorded. By default a notetaker appears in the participant list, which is a real disclosure and more than a silent recorder gives.

Against that, the vendor documents that the bot's display name can be customised, so the participant list entry need not identify the product or announce a recording, and separately ships a capture mode through a browser extension that operates without any bot joining at all, removing the visible signal entirely.

The people affected are meeting participants at other companies who never bought this product, and in several jurisdictions their consent is a legal precondition rather than a courtesy. Nothing reached describes an in meeting announcement, a consent prompt, a jurisdictional consent mode or any record that consent was obtained. Graded in the middle of the band because the default behaviour does disclose. Ask what an external participant sees and is asked before recording begins, and whether the botless capture mode announces itself at all.

Programmatic access on every tier including the free one, which is rare enough to be the headline here. Most vendors reserve an interface for their most expensive customers; this one lists it against the free plan and publishes a reference at a documentation subdomain, evidenced by a deep link to a specific named query with its own anchor rather than a bare footer pointer.

Around it the integration surface is organised into six categories with their own pages, spanning conferencing, audio, customer record systems, dialers, collaboration and storage, with every paid tier marked as carrying all of them. A browser extension is distributed through the official store, desktop and both mobile platforms are supported, and the vendor runs its own store of packaged skills, which is a genuine extension point rather than a connector list. It stops below the top band because the documentation itself was not retrieved on this pass, so what the interface exposes is unverified, and no integration count or marketplace listing is stated.

One isolation option is sold and geography is untouched. Private storage is published as a top tier feature and described as dedicated cloud storage exclusively for a single organisation's data, which is a real tenancy choice a buyer can specify rather than a claim of privacy, and it sits alongside custom data retention at the same tier. That is more than most vendors in this category offer at all. Everything about location is absent.

No hosting provider, no region, no processing location, no residency election and no cross border transfer mechanism appears in any material retrieved, and the trust centre and security page were not reachable on this pass so nothing here asserts they are silent on it. European data protection is claimed across all tiers without a transfer position being described.

The question carries weight because the stored material is recorded speech of identifiable people across whatever countries a customer meets in. Ask in which regions recordings and transcripts are stored, and whether a region can be elected.

Real certifications, a live trust centre and an unusual decision about who gets them. The service organization control attestation of the period type and European data protection are both marked as available on every tier including the free one, where most vendors in this index gate compliance to their most expensive plan, and the health information regime is covered at the top tier with its own dedicated page.

A trust centre operates on its own subdomain, named in the footer for all three regimes, and a separate security page and a bug bounty route are published. In product controls are enumerated in the pricing matrix rather than described: single sign on, directory synchronisation, a super admin role, private storage, custom retention and audit logs reachable through a documented interface query.

Held below the top band because the trust centre and security page bodies were not retrievable on this pass, so no auditor, audit period or report request route is established, and no examination report or public artifact was located. Ask for the current report, its audit period and the auditor.

Every tier carries a number, including the enterprise one, which almost nothing else in this index manages. Four plans are published with per seat figures at both billing frequencies on the two middle tiers, a permanent free plan at zero, and a top tier priced at a stated per seat rate rather than behind a quote, marked annual only.

The usage economics that normally decide the real bill are published as figures rather than adjectives: storage allowances stated in minutes per team and per seat, recording length limits in hours, assistant credit allowances per tier, and a linked published explanation of what a credit buys. The first three tiers purchase through checkout without a conversation, invoicing is noted for the top tier, and volume discounting is acknowledged with the thresholds named.

Two caveats sit inside the grade rather than against it: the free plan advertises unlimited transcription against an asterisk with no footnote rendered on the page, and the volume discount percentages are not stated. A buyer can nonetheless budget their own purchase at any size, which is the test.

The routes out are real, plural and mostly available low in the ladder. Transcripts, summaries and recordings are all explicitly downloadable from the first paid tier, which is a direct answer to the question rather than an inference, and a bulk delete exists for removing material at scale. Programmatic access sits on every tier including the free one, so an extraction path exists for customers who never pay anything, and the interface reference documents named queries.

Ownership is stated as belonging to the customer and located in the terms of service rather than asserted loosely, and the top tier adds custom retention control. It stops below the top band on the departure question specifically. Nothing published states what happens to stored recordings and transcripts after cancellation, how long a departing customer has to retrieve them, or whether the storage allowance is lifted to allow a full extraction. Storage caps below the top tiers also mean older recordings are removed during normal use. Ask for the retention window after an account closes and whether export limits are lifted on cancellation.

No sending infrastructure exists here to examine. The product operates no mailbox of its own, no sending domain, no warmup, no rotation and no throttling, because it records and summarises meetings rather than contacting anyone, and its output travels as notes into whatever system the customer already uses. Graded in the middle of the band with that scope stated rather than at the floor, since the floor asserts a sending product without discipline and this does not send.

The one adjacent surface is the assistant that drafts replies inside a user's own inbox, where any message that results is sent by the user's own mail platform under that platform's authentication and reputation, not the vendor's, so the deliverability consequences sit with the customer's existing provider. Nothing published addresses volume, pacing or safeguards on drafted replies. Ask whether drafted replies can be sent automatically without review, and what limits apply if so.

Ten use cases addressed with their own pages, which is the broadest coverage documented in this batch, spanning sales, recruiting, marketing, product research, engineering, collaboration, venture capital, healthcare, real estate and podcasting, and two of those are backed by actual regulatory coverage rather than positioning, with the health information regime and an education records agreement both offered.

The commercial ladder runs from a permanent free plan through three priced per seat tiers with volume thresholds named at fifty, one hundred and two hundred seats, so every size has a described and mostly priced path. International reach is evidenced concretely rather than claimed: transcription in more than a hundred languages, a site published in five, and desktop, mobile and browser clients.

It stops below the top band because no geographic coverage statement exists anywhere reached, with no regions, no market list and no statement of where the vendor sells or supports, which is the same gap that holds several records in this batch at this grade.