An AI notepad for meetings. It listens through your own computer instead of sending a bot into the call, then turns the few lines you typed into complete notes you can search later.
Research Sep–Oct 2026 · figures as of the dates shown, not updated for this issueYou type a few fragments. When the call ends, Granola writes the notes you would have written if you had an hour to spare.
Granola is a desktop app for Mac and Windows, now also on iPhone, Android and Apple Watch. It reads your calendar, so it knows when a meeting starts. It listens to the audio on your own computer instead of sending a recording bot into the call, so it works the same on Zoom, Meet, Teams, a phone call or a conversation across a table.1
One misconception to drop early: this is not a transcription app. Otter, the category’s older leader, sends a visible bot and gives you something like a court stenographer’s record. Granola tries to give you the one-page memo a sharp chief of staff would write. To Granola, the transcript is raw material. The note is the product.
The time heavy users report saving each day at four to five meetings.2 Below about three meetings a day, the product is nice to have rather than essential. Judgment
1Product description from Granola’s site and help docs as summarised in the research notes (Sep–Oct 2026). Mobile apps: Android Jul 1, Apple Watch Jul 28, 2026. Sources: S13, S14.
2User reports gathered in a review roundup cited in the research (Sep 2026), not a company figure. The three-a-day threshold is the author’s judgment.
Two early decisions define Granola: your own typed lines anchor the AI’s write-up, and the audio is never kept.
The hard part of summarising a meeting is not shortening it. It is knowing what mattered. The few words you type are that signal. Granola turns a hard judgement problem into an easier expansion problem: the write-up is built around your lines, and the detail comes from the transcript, with a link back to the passage it came from.1
The second decision was about trust. Granola transcribes as you go and throws the audio away. The trade-off is real: there is no recording to replay, and on desktop the speaker labels are mostly just “Me” and “Them”, the gap rivals cite most.2
When we first created Granola, we made a very conscious decision not to record and store any audio … And everyone kind of laughed at us for that.
Chris Pedregal, CEO · Invest Like the Best, Feb 25, 20251Me/Them comes from two physical audio streams, not from AI guessing who spoke. Source: research notes on Granola’s architecture (Oct 16, 2026), based on Granola docs and founder interviews; S14, S31.
2Product gaps from critical reviews and a competitor’s comparison page cited in the research (Sep–Oct 2026); S1, S24.
Granola started with investors and founders: the people with the most meetings and the loudest networks. Often, the champion is also the boss.
Business software usually involves three people: the user, the champion who pushes for it internally, and the buyer who signs off on the budget. At Granola those roles can be the same person. In the 2024 cohort, 57% of users held leadership roles.1 When the champion is the boss, adoption moves down through the company, and it moves quickly.
“When you’re in a meeting we can expect to have 2% of your attention and … the other 98% is on the person you’re talking to.” Sam Stephenson’s design brief, Apr 2026.4
The number one piece of positive feedback we get … is that it helps me feel more present in the meeting.
Sam Stephenson, co-founder · The Cognitive Revolution, Apr 8, 2026, 1:31:561Sacra, 2024 cohort Reported. Retention from Series A coverage (TechCrunch, Oct 2024) and Okara (2026) Reported. 2Seat: one paid user licence. 3Upstarts Media. 4Cognitive Revolution, 1:02:50. S5, S12, S28, S29.
The two founders had never worked together. They held off launching for nine months against investor advice, then grew with almost no marketing.
The founders say the delay was deliberate. “Typical VC actually been pushing us to launch for about nine months before that. We held them off for nine months,” Stephenson said in May 2025.1 The rival Pedregal chose to worry about was “the one that hasn’t launched yet”.2
The growth came from people sharing it. “Most of our virality is … either just people recommending it to each other or people sharing notes with each other,” Stephenson said in April 2026. Every shared note is an invitation. That loop has to offset a cost he names himself: without a bot, there is no “huge glowing orb with a logo on it” in other people’s meetings.3
We were seven years late when we founded Granola. We didn’t launch for a year.
Chris Pedregal · Invest Like the Best, Feb 25, 20251Generative Now, May 15, 2025, 17:54 (beta of ~150: 17:26). 2Invest Like the Best, Feb 25, 2025. 3Cognitive Revolution, Apr 8, 2026, 1:10:30 and 52:28. Four-person team: Hallway Chat, Jan 24, 2025. Seed and Series A are confirmed by SEC Form D. The B and C rely on company and press statements; no Form D was found as of Oct 2, 2026. S1–S2, S5, S8, S12, S30–S31.
In 26 months Granola’s stated scope widened, then narrowed. The open question was who does the work with the context.
Through 2025 the ambition widened. In May, Pedregal said “meetings are really just to start like we’ll have to add emails, we have to add Slack”. By August, Granola was “a tool for thought to help you do work”.1 In early 2026 the record opened: an MCP server,2 then, after an a16z partner publicly quit over newly encrypted local files on March 16, a public API within the week.3 By July the plan was narrower: do meeting-adjacent jobs, and feed context to whatever agent the customer uses.
The seam still shows. In June the VP of Sales said “the product will start using the context of your conversations to help you get work done.”4 Email and Slack as inputs, first floated in May 2025, had not visibly shipped by October 2026. Gap
Be best in the world in anything that’s like meeting adjacent and then … be the best way to capture context to power any agents you have.
Chris Pedregal · AI & I (Every), Jul 15, 2026, 39:561Generative Now, May 15, 2025, 34:24; MAD Podcast, Aug 21, 2025, 36:40. 2MCP: an open standard that lets AI assistants such as Claude or ChatGPT read data from other tools. 3Public posts by Guido Appenzeller (Mar 16) and Pedregal (Mar 17, 2026), as cited in the research. 4The Crew, Jun 23, 2026, 18:09. S2–S3, S6–S7, S20, S31.
Among businesses paying for a dedicated AI notetaker, Granola was the default by autumn 2026.
Two independent spending datasets point the same way. On Ramp’s card data, 64% of businesses buying AI meeting notes were paying Granola, up 34 points in a year, and 74% of its buyers there had switched from a competitor.1 In YipitData’s mid-market panel, churn was near zero, spend tripled in six months, and switching ran mostly one way: toward Granola.2
| Rival | What the record shows | As of |
|---|---|---|
| Fathom | Headcount down 6% in six months; then sold to Superhuman | Apr / Sep 14, 2026 |
| Otter | ~$100M ARR (company); mid-market spend flat (Yipit) | Mar 2025 / early 2026 |
| Fireflies | ~40% more panel customers; spend falling | Feb 2026 |
The category is too small to appear in Ramp’s ranking of top software vendors. Winning 64% of a line item is not winning a budget. Judgment
1Ramp vendor page, updated Oct 2, 2026. The research notes the figure partly covers the months after Google switched its own notes on by default in July. 2YipitData, Apr 2026, panel data to Feb 2026. Rival rows: TechCrunch, Otter, YipitData. S10–S11, S16, S25.
The platforms give notes away. General AI apps treat them as one feature of a much bigger daily habit. Open source gives the whole design away for free.
It hasn’t changed anything from our growth rate or what not, but again, we’re still a drop in the bucket.
Pedregal, asked whether lookalikes showed up in the numbers · AI & I, Jul 15, 2026, 07:511Google Workspace Updates, Jul and Sep 2026. 2Windows Latest, Jul 7, 2026. 3Reviews through Apr 2026 described ChatGPT Record as Mac-only with no speaker identification; its status was not re-checked after Oct 2. GitHub star count as of Oct 2026. S15, S23–S24.
On Hamilton Helmer’s seven powers, Granola holds one today and is building three.
Counter-positioning means a rival cannot copy you without hurting itself. Otter’s growth depends on its visible bot, which advertises the product in every call it joins. Dropping it would break that loop. Google’s value is the surface it owns, so it cannot be neutral across Zoom and Teams. Through 2026, both moved further into their own products instead. That explains the switching data better than any feature.
But it only works against incumbents whose models punish imitation. Notion, ChatGPT Record and Anarlog are all bot-free. Against them, Granola’s idea is a well-made feature with a strong brand, not a structural power. Judgment
| Building | Evidence | Why it is not yet a moat |
|---|---|---|
| Switching costs | Meeting history, MCP and CRM connections | Notes export freely; the free plan keeps 30 days |
| Network effects | Shared Spaces, shared links | Still mostly single-player |
| Brand | A verb among founders; a rival named itself after it | Three years old, and mostly known in tech |
Even if somebody comes along and they build a better granola, you’re not going to switch because it has all of your context.
Michael Mignano, Lightspeed (seed investor) · Sourcery, Dec 27, 2025, 09:07. This has only been tested on long-time users.Framework: Hamilton Helmer, 7 Powers. Scores reflect the author’s reading of the competition research (Oct 2, 2026), not a third-party rating. S1, S4.
Revenue is undisclosed. What is visible: a price ladder built to recruit teams, and costs that rise with the most-loved features.
The company has said annual recurring revenue grew more than 400% in the year to March 2026, and 2.5 times between January and the Series C announcement.1 That jump coincided with removing the individual plan, so it is not a steady rate. Sales works bottom-up: reps take adoption that already exists in an account and “formalize it in a company contract.”2 On Oct 2, 2026, eight of 18 open roles were in go-to-market.3
Costs are less settled. “There was a time where half of our burn rate as a company was going on transcription,” Stephenson said.4 That is now cheap; the new cost is cross-meeting AI work: about half of weekly users run agentic queries, and Granola prepares meeting briefs even though only about 10% are opened.5
Agentic features are really expensive um and so as we build more and more of them at some point that’ll … break the math.
Chris Pedregal · AI & I, Jul 15, 2026, 56:45Gross margin is probably 50–70%, nearer the low end, below typical software. The CEO said only that it was not negative (Aug 2025). Estimate
1Reported via Forbes and TechCrunch, Jan–Mar 2026. 2The Crew, Jun 23, 2026, 36:25. 3granola.ai/jobs. 4Cognitive Revolution, Apr 8, 2026, 42:23. 5AI & I, 57:27 and 56:27. S3, S5–S7, S12–S13, S26.
The usual case against an app built on someone else’s AI is that the next model release makes it pointless. In Granola’s case that is the wrong test.
Granola does not train a frontier model and does not try to. “Our philosophy is to use the best model that is on the market as quickly as … possible,” Pedregal said in 2025.1 What it owns is the workflow around the model: when capture starts, what the user typed, which past meetings are relevant, and where the result goes next. So far, three things have meant each model generation helped Granola more than it hurt.
The catch is that the same models are available to Google, OpenAI and the open-source copies. The real test is relative: does Granola improve faster than the free version? Stephenson has said he has “worried about this since day one,” and, when asked why users still pay, “I don’t have great answers as to why.”4 Judgment
If Granola doesn’t get better when … the next version of Claude or GPT comes out, then I think over … time we’re cooked.
Sam Stephenson, co-founder · Cognitive Revolution, Apr 8, 2026, 1:30:331MAD Podcast, Aug 21, 2025, 38:31. 2Generative Now, May 15, 2025, 10:22; MAD 42:45. 3MAD 43:32 and 44:36. 4Cognitive Revolution, 1:29:05 and 1:29:33. Context window: how much text a model can consider at once. S2–S3, S5.
Capturing a call without a visible bot is what users love. It is also what a pending class action says is illegal.
Chamberlain v. Granola was filed in federal court in Northern California on July 30, 2026. It alleges that recording other people in a meeting without their consent, and training models on those recordings, breaks California’s wiretap law and federal law. These are allegations, not findings.1 Two weeks later, the same court refused to dismiss similar claims against Otter. It held that a notetaker that trains on recordings for its own purposes can be treated as a third-party eavesdropper.2
Granola’s defence was in place before the suit was filed: “granola from the get-go never stored the audio. It only stores the transcripts.”3 But reporting in April and July 2026 described defaults that cut the other way: share links anyone could open, training switched on by default for non-enterprise users, and folders shared across a whole organisation.4 As of Oct 2, 2026, coverage showed no answer or motion to dismiss; the case may have moved since. Status not current
The author’s estimate of the chance of some forced change, such as opt-in training or stricter sharing defaults. The worst case, mandatory notice to every participant, is put at 20–30%. Judgment
| Risk | What to look for | Read |
|---|---|---|
| Legal | Docket filings; changes to defaults | Highest |
| Platform clock | Spending data after Google’s Sep 29 auto-on | High |
| Margins | Usage-based pricing; briefs generated only when needed | Medium |
| Disclosure | No Form D for the B or C; UK FY26 accounts due Dec 31, 2026 | Medium |
| Cash | Runway est. 3.5–5 years Estimate | Low |
1N.D. Cal. 3:26-cv-07926. 2In re Otter.AI Privacy Litigation, order of Aug 13, 2026, as reported by Recording Law. 3Pedregal, Generative Now, May 15, 2025, 36:11. 4The Verge (Apr 2026) via Hacker News; a Hacker News user (Jul 2026). Granola has since added a user-settable “Private” sharing default and training opt-outs. S1–S2, S8–S9, S14, S17–S19.
Granola has won the paid notetaker market. The $1.5B price depends on a second act that it does not yet own.
The best retention in the category and the strongest switching data. A neutral record of every meeting, available to any agent through MCP and an API. A founding team that delayed launch, kept a clear identity, and cuts features to keep the product calm.
A small category whose price is heading towards zero. Every repeatable move has been copied. Margins that agent features will squeeze. A lawsuit aimed at the defining feature. No disclosed revenue, users or renewals.
Our reading: both are right, on different timescales. The first act was won on product quality, not on a structural moat. The second act, supplying context to agents, is the right direction, but the agents, buyers and budgets are not yet in place. In October 2026 the research put the window for turning brand and meeting history into organisation-level lock-in at about 9 to 18 months, before free, automatic notes become the default box ticked for small teams. Judgment
The 9–18-month window is the research’s judgment as of Oct 2, 2026, and is not updated here. S1.
Method. The research was carried out between Sep 20 and Oct 16, 2026, and no figure is updated after those dates. Every number carries the date it refers to. Filing marks public filings, Reported marks company statements and press, Estimate marks the author’s estimates, and Judgment marks the author’s opinion. Quotes come from transcripts of the episodes cited. Some were machine-transcribed, so filler words are cut and marked “…”. Timestamps are approximate.
Gaps. Revenue (never disclosed; estimates left out). Current user numbers (last disclosed: about 5,000 weekly users, Oct 2024). Net revenue retention. Enterprise renewals. Agents using the API in production. Form D filings for the Series B and C (absent as of Oct 2, 2026). Granola’s response in court. The size of the stake held by Nat Friedman and Daniel Gross’s fund.
Glossary. ARR: annual recurring revenue. MCP: Model Context Protocol, a standard for AI assistants to read other tools. Form D: SEC notice of a private fundraise. Counter-positioning: a model that rivals can’t copy without hurting themselves.