Skip to content
MakerGain
← Back to cases

Granola Doesn’t Send a Bot to Meetings; It Bets on the Notes You’ll Still Open Afterward

Edited by MakerGainPublished: Article updated:

Granola does not send a meeting bot. Instead, it transcribes through the computer’s system audio and microphone, making an AI notepad for back-to-back meetings. Its co-founders say a bot is like an uncomfortable third party squatting in the corner of every call—and it doesn’t work in person either. The trade-off is that by default Granola does not make itself visible to other participants: online, automatic meeting chat notices and video watermarks can be set up; in person, the person in the room still has to speak up. Its preference for “useful but incomplete” over “wrong but confident” is an engineering trade-off, not a guarantee of accuracy.

A meeting bot has an awkward problem: it cannot get into in-person meetings. So Granola simply does not send a bot. Instead, it uses the computer’s system audio and microphone to transcribe in the background, which means Zoom, Google Meet, Teams, Slack Huddles—and even face-to-face or impromptu conversations—all go through the same capture method (Granola’s website[1]). It defines the product as an AI notepad for back-to-back meetings: notes, action items, and recall, available on macOS, Windows, iOS, and Android—all of them (same source).

Why no bot? Co-founders Chris Pedregal and Sam Stephenson put it plainly: a bot cannot be used in in-person meetings, behaves inconsistently across meeting platforms, and worst of all, it is “extremely awkward—like an uncomfortable third party squatting in the corner of every call” (Why Granola doesn't use a bot[2]). In their view, that is as much a fundamental flaw as “it doesn’t work in person,” not a minor annoyance you can put up with.

Granola processes meeting audio for transcription, but it does not record or retain it; the feature table on its pricing page also states “audio deleted after transcription” (same source[2], Pricing plans[3]). They want Granola to feel like a notepad, not a recorder. That choice has another side: without a bot, Granola is not automatically visible to other participants by default. The company offers automatic meeting chat notices and video watermarks; these take a few minutes to set up, and once enabled they run automatically during transcription, letting participants know Granola is on. But in person there is no one-size-fits-all move; the simplest approach is still to say it out loud. Granola also reminds users to get consent each time they start taking notes, and notes that some jurisdictions may require prior permission (same source). In other words, online, “letting people know” can be a one-time setting that runs automatically afterward; face to face, the person sitting there still has to speak up.

The product is a workflow that spans before, during, and after meetings, but not every tier gets the full version. Before a meeting, it syncs your calendar and generates a Brief for external meetings, listing participants, what was discussed last time, and current priorities; during the meeting, you can take notes yourself, as much or as little as you like, and the system combines the meeting context to write notes that are clear to you personally; after the meeting, it provides notes, action items, and follow-ups (Granola’s website[1]). The feature table on the pricing page marks Brief and follow-up emails as Business and above; Basic does not include either. Basic still has cross-meeting AI chat, shared folders, templates, and voice input (Pricing plans[3]). The main deliverable after a meeting is those notes, action items, and follow-ups; the transcript is the underlying material.

Image: Official Granola display[1], accessed 2026-11-10, original interface preserved.

This orientation toward “writing each person their own notes” also shows up in how they look at data. On May 10, 2025, founding engineer Jim Fisher wrote that the team largely replaced funnel and DAU analysis with a dot-plot data table sorted by sign-up and showing each user’s number of meetings per day. What they ask while staring at the table is “Ben churned… wait, did he switch to another Gmail?” “Ana seems to have invited colleagues?”—not “an average of 2.44 meetings per day.” Averages hide the distribution, and they also hide specific people (Start with a dot-plot[4]). The product writes notes for each person, and the team also starts with specific people when looking at data.

Granola Chat is the other side of this trade-off. The company’s new version of Chat, released in “late April,” has to find answers across several months or even years of notes and transcripts; the retrieval process is described as Search → shortlist → read → repeat, and Coverage Notes tells you which sources it looked at (How Granola thinks about designing agents[5]). One number in the engineering explanation: the token count of transcribed text is about 10 times that of structured meeting notes. So the new Chat chose the latter of “reading 10 meetings closely” and “getting an overview of 100 meetings” (same source). That is an explicit concession: better to read each meeting more shallowly in order to cover more meetings, and then lay out “what I looked at” for you to judge for yourself.

Granola describes this trade-off as four output states: useful and comprehensive, useful but incomplete, incomplete and visibly so, and wrong but confident. The first is ideal, the second is often acceptable, the third is still useful, and the fourth must be avoided—“an answer that sounds certain but is wrong will quickly destroy trust, and trust is much harder to rebuild than it is to lose” (same source). They would rather choose “useful but incomplete” than “wrong but confident.” For a tool that searches notes across months and years, that matters more than reading a few more meetings.

Who pays is clear on the pricing page, but it does not lock buyers into one kind of person. Basic is free, with unlimited AI notes and also cross-meeting AI chat, shared folders, templates, and voice input; individuals can opt out of model training at any time. But meeting history is kept for only 30 days, and there is no pre-meeting Brief or follow-up emails. Business is $14 per person per month, opens up history, and adds Brief, follow-up emails, advanced models, integrations, and centralized management; Enterprise is $35 per person per month, aimed at larger companies, and adds SSO, organization-level sharing and API management, HIPAA and other security and administrative controls, plus a workspace-wide model training opt-out, with model training off by default (Pricing plans[3]). Beyond price, this article is more concerned with three differences: how long history is kept, how far sharing and management extend, and whether compliance controls are sufficient. The pricing page positions Business as “Great for individuals or small teams,” and individuals can buy it too; only Enterprise brings in stronger compliance and administrative controls. The people operating it every day are the ones in meetings; the people paying may be individuals, or administrators buying on behalf of a team.

MCP Connector connects meeting notes to the AI apps users already use; it is not another meeting feature. Permissions are tiered: Basic can read only an individual’s notes from the last 30 days, Business extends to personal and public workspace notes, and Enterprise is configured by administrators (same source). This set of permissions explains a very practical problem—individuals want to search their own old notes, teams want to share, and companies want unified control over what can be connected and how long it can be viewed.

Brex’s case page provides a specific scenario: Lloyd Cilliers, Brex’s head of IT, decided to roll Granola out officially only after discovering that many employees were already using it; during implementation, some teams first tried the Business plan, and within several weeks they moved to Enterprise (Brex saves 427 hours a week with Granola[6]). The same page self-reports that knowledge-worker adoption was greater than 50% within 6 months, with 25k+ hours saved since adoption and 427 hours saved per week. These numbers come from a customer case page, with no independent measurement, and they cannot be generalized from this one customer to all companies. The cited official website, pricing page, and this case page contain no Granola revenue, gross profit, or customer count, so these materials alone do not show whether the business is profitable.

Among comparable tools, Fathom’s free plan lists unlimited recording and transcription, instant AI meeting summaries, clips and playlists, cross-call search, and offers two capture methods: bot-free (Mac beta) or a bot joining the meeting (Fathom pricing[7]). Both have a bot-free capture path, but they retain different material: Granola explicitly says it deletes audio after transcription, while Fathom’s free plan lists unlimited recording, transcription, and call storage. The difference is whether you want “no one can go back through old recordings after the meeting” or “keep the old recordings for later.”

Granola is betting on the page of notes you will still be willing to open after the meeting; to that end, the product by default does not make itself visible to participants on the user’s behalf, but instead offers automatic chat messages and watermarks so that online notices can be set once and then run automatically. But when you sit down face to face, the person who presses start and speaks up to say “I’m taking notes” is still the person sitting there.

Continue reading

AI productivity
He Lost Half His Customers and Was Happier: BoltAI Raised Its Price from $19 to $79

In September 2024, BoltAI founder Daniel Nguyen looked back at raising the price from $19 step by step to $79: customers were down to about half, monthly revenue was slightly below the $39 tier, but refunds and support inquiries fell, and he counted founder happiness in the pricing math. BoltAI is a native Mac AI workspace that brings together multiple providers and local models; what makes this trade-off notable is that it exchanges fewer customers for a lighter solo support burden.

AI search
He Treated Reddit as a Survey. 135,000 People Used It for Research. In 2025, It Stopped at a License

Fed built GummySearch into a tool for audience research using public Reddit discussions: it searched and tracked community conversations by categories such as pain points and people seeking solutions. After commercial licensing talks with Reddit failed, he chose to stop sales on November 30, 2025, and gradually shut down the site rather than continue operating without a license; paying users kept transitional access, and the free tier retained only login, export, and account deletion.

AI audio
MacWhisper Turns Recordings Into Text, Then Leaves Identifying the Speakers to You

Podfeet's Allison uses MacWhisper for interview captions and transcripts: she enters a 12.6-second offset, exports an SRT, then labels speakers sentence by sentence and merges paragraphs. Automatic recognition can separate voices, but it cannot decide identities for her; she uses ChatGPT to help with segmentation and runs into confusion between an API key and a Plus subscription. Developer Jordi Bruin tightens the terms because of license resale.

AI productivity
The day the accountant came asking for the bookkeeping records, he decided to let AI go through his own inbox

In March 2023, software company founder Romeo Bellon was overdue on bookkeeping and his accountant pressed him for records. He developed Receiptor AI, which scans inboxes, extracts receipts and invoices, and pushes them to accounting software; he launched it with Luigi Fernandez Ortega in August. It turns historical email into a retroactive scan priced by email volume, while subscriptions process only new email; its privacy policy says documents are retained indefinitely and automatic deletion is still in development.