Granola Added More Customers Than Almost Anyone by Being Surprisingly Unambitious

Nathan Labenz is a paying Granola user, and he opens with a ranking that still startled the founders. In January 2026, the spend platform Ramp reported that Granola added more new customers than any company it tracks except Anthropic. The AI meeting-notes app had just raised $125 million at a $1.5 billion valuation, with a team of about sixty. Labenz wants the growth story. Stephenson was taken aback to see Granola in that company, then names a mechanism that is almost boring: it is almost entirely word of mouth. People recommend it, or they share notes.

They had hoped, from the start, to build something good enough that friends would talk about it, with some product virality baked in. Unlocking those loops took a while. Now, he says, the thing is snowballing month over month. The later detail that completes the picture is a single moment users keep repeating. After a meeting, beautiful notes appear in Slack in about thirty seconds. A colleague asks how anyone could have written that so fast, digs, finds Granola, and downloads it. Recipes travel on social media. The snowball is still that Slack beat.

The character they hold in their heads, Stephenson admits, is more extreme than the average user. Meeting notes were never the destination. The team wanted to invent interfaces for ordinary people who use computers as a means to an end, not as a thing to geek out on. Notes were a foot in the door: a habit that might let them do more later. Early interviews were open-ended, with no agenda, just work lives. The archetype that emerged was someone in back-to-back meetings all day — sales, account management, recruiting, investing, founding, client services.

Co-founder Chris brought the kitchen-utensil analogy. OXO designed for people who struggled to grip, or who had one arm, not because those were the only customers, but because a product that worked for the most extreme user became friendly for everyone else. Accessibility designers know the move. Granola’s version is the person running meeting to meeting, frazzled by context switching, barely able to use the bathroom, calendar a solid block. Those people exist. They are also a cartoon of a state most knowledge workers hit often. Hit that, and you get a simple product for the rest of us on our frantic days.

Labenz used to aim at the center of the distribution and work outward. Stephenson’s flip is the opposite. Software builders, himself included, assume the user arrives calm, giving full attention, able to string together complicated maneuvers and develop mastery. Sometimes that is true, far less often than builders like. Most computer work is reactive and chaotic. People are in System 1: morning inbox, three fires, a meeting in twenty minutes, a standing feeling of being behind. Design has to fit that reality, not the rational self who shows up for a usability session.

Classic user testing immediately ruins the state you care about. Sit someone down, say you will not interfere, and they go into full focus so they will not look stupid. Stephenson is guilty of wanting that validation too. The workaround is to ground every conversation in the facts of the day. When Granola was thinking about folders, they asked people to share a screen, open the home list of meetings, and go one by one: what was this, who was there, who should see these notes, how would you organize it in your head. The person is still in an interview. They cannot lie about what is on the screen.

The trap he fights constantly is abstraction. As soon as someone talks about theoretically or generally what they do, you cannot trust it. You are getting their imagined view of themselves, which is often far from how they behave. Calendars work the same way: pull the real day up and use it as a forcing function. For evaluating what they have already built, the team leans on dogfooding. Stephenson would rather sit quietly on a sales call and watch a teammate use Granola in the wild than hear a filtered account of the experience.

Labenz names the tension. The industry pitch is that AI takes the drudgery so people can be their System 2 selves. Product design, which Stephenson is embracing, says meet people where they are. Will Granola still be for jammed calendars, or will it eventually help people *into* that higher mode, which would imply a different product later? Stephenson wants both, at different speeds. He has tried a pile of “executive assistant for knowledge work” tools. They fail the same way. The work is messy and socially nuanced: what is the relationship, how should this email sound, where does this request sit against everything else on the list.

A screenshot looks helpful. Engage with the actual draft or to-do list and a slight miss makes the whole thing useless. You are not quite comfortable sending it. Models will get better, and connecting tools will get easier, but it may take a surprisingly long time. The principle that keeps Granola effective meanwhile is to be surprisingly unambitious about the jobs they promise. It will be a while before Granola can touch the thorny, top-of-mind fires. It can already take the menial stuff people postpone because it is unimportant, plentiful, and boring. Be deliberate about that, and you help sooner.

System 2 is still the lens, just on different surfaces. The full-screen chat, which can see all of your meetings and the company’s, is meant to be sat with: Labenz uses the multi-meeting view heavily; Stephenson uses it for posts, job descriptions, and analyzing a part of the company. That is a different mode from the notepad in a back-to-back slot, so the design can loosen. The notepad, he later says, can expect about two percent of attention, with ninety-eight percent on the other person. Chat can take more like eighty percent, and they throw more at you there.

Labenz has spent months dumping five years of email, Slack, DMs, call transcripts, and diarized podcasts into a local database his agent can query. Intros that used to fail for lack of history now come back close enough to send. He is an admin on every system. What, Stephenson thinks, will keep ordinary people from that for a quarter or two — trust, export hostility, something else? Stephenson agrees the machine is only as good as the context. On a personal machine, gathering your stuff is more possible than it used to be. Inside a company it is still fragmented: permissions, security sign-off, tools that do not connect.

Some Granola customers are surprisingly willing to let employees explore, because the alternative looks like getting left behind. Context is still the bottleneck, and so is the amount of personalization and memory an agent needs — you, the people you work with, the projects, the priority stack. A proactive power user can teach OpenAI’s or Anthropic’s general agents those things today. The average computer user will not. Granola’s job is to make as much of that happen in the background as possible, so useful work is waiting when you open the app, and the user never has to become a configuration engineer.

Sharing is the other wall. Granola defaults every note to private. Conversations are personal; one dodgy line can pollute a transcript for the whole org. The value of transcripts, though, is a shared space the company can use. They have not nailed the design. A basic version already runs: finish a note and an LLM suggests a folder from the call plus what usually lands in each folder. It is not accurate enough to turn on automatically. Labenz’s instinct is to run transcripts through a model that flags what a guest might regret; Stephenson is tinkering. Even 98 percent automatic routing may not be enough if a tiny risk of a sensitive line going public is unacceptable.

Labenz has never hit a credit meter, a limit warning, or a per-action cost in Granola. Is that the Uber phase — who cares what it costs — or does the usage actually work at a fixed seat price? Stephenson says it is a function of stage. Pre-launch and around launch they told themselves there was no budget. Building a good product was already hard enough; with small user numbers, optimizing cost was a waste of time. Transcription APIs made the early product genuinely expensive. Scale and cheaper, more commoditized speech models pulled that down. Notes are now a predictable cost: meetings have physical limits and average out over a month.

Heavy users who chat across many meetings rack up bills the company currently swallows. As Granola moves from companion to something that more directly does work, usage-based pricing in the Cursor style will probably have to arrive. While people still think of it as a notes app you bring to meetings, they like a simple seat price. A year after launch, if you plotted user growth against then-current unit cost, the line looked terrifying. At one point half of company burn was transcription. It is under control now. LLM-heavy assistant features may swing the pendulum back and force the pricing conversation. For now they are comfortable.

Transcription itself is simple and deliberate. They record microphone and system audio at the OS layer and pipe both to third-party real-time APIs — Deepgram and AssemblyAI, with traffic split in the desktop app for redundancy, because if Granola goes down users notice immediately and it hurts. Real-time is comforting, lets notes generate the moment the meeting ends, and lets them discard the audio. Quality and speaker separation are worse than a full-file batch transcript. They still think it is the right trade. Speaker diarization, if they ever nail it, would make everything else ten times better. The stack is built to swap models quickly. On-device models may win on speed and cost later. They stay nimble.

Working at the OS audio layer means Granola can run without anyone else on the call knowing. They ship calendar-invite notices and more disclosure tools because a glowing join-bot would have been a growth channel they refused. Stephenson’s stance is that Granola is a tool, like Voice Memos or a notepad. Following local law and telling the other people is on the user. Hiding it is not in Granola’s interest; they want it in the open. They used to get far more privacy questions two years ago. They get fewer now. Work transcription, he thinks, will become default because the upside is hard to pass up, and a meeting is already a social contract to sit down and work — unlike an always-on wearable in the rest of life.

They toyed with hiding the transcript entirely, even letting it self-destruct, because notes were supposed to be the value. They learned quickly that notes are good for humans and transcripts are amazing for LLMs, so the transcript stays. They are not interested in a word-for-word record that would hold up in court. Granola needs enough detail for the model to work. Audio does not. In the first week of putting the product in people’s hands they stored recordings in S3, telling themselves they might train on them later. They lasted a week. It felt too creepy, heavy, serious — a thing you use against people. They turned it off.

Labenz is splitting a co-pilot agent on his main machine, with deep access, from a more autonomous agent on a Mac mini that he may let send email. He is experimenting with monthly and annual summaries, project and relationship views, so the autonomous one gets a blurrier picture. Stephenson reaches for their executive assistant: only as good as the context you give, and you still would not want them in the room for every conversation with a co-founder. Forgetfulness is underrated. Human memory fidelity drops fast; for agents that can be a feature, an elegant way to diffuse dangerous detail from transcripts and mail. Labenz likes the idea of agents that literally shrink over time. Stephenson has no fully worked protocol. He thinks we will need those kinds of agent-to-agent manners, and nobody has them yet.

Among AI products Labenz has used, Granola is the most disciplined about feature bloat, in a year when you can vibe-code ten features before lunch. Stephenson says the calm is purposeful, especially on the core flow: popup, notepad during the meeting, generated notes after. He is almost paranoid about adding to that view. The bar is high because extra chrome makes the notepad less of a place you want to spend time. The cost is burying useful things. Templates can structure notes however you want; many users never find them, which he thinks is correct — most people should not have to think about it. Folders and multi-meeting chat get more experiments because the user is in a different frame of mind. He would like them to be *more* experimental there. Calm sometimes delays shipping good ideas.

Recipes are saved prompts that run at a click. Power users had already discovered they could chat across a body of meetings and pull out something profound; recipes make that repeatable. Internally, the CX team meets on what customers are confused by, then hits a recipe that reads the transcript, looks at the docs, finds holes, and suggests edits. Hiring is similar: get the relevant people in a room for half an hour, point the model at existing job posts plus the transcript, and the first draft is strong. Granola is fully in person. Someone puts a phone on the table; people argue unstructured; the robots sort structure and format. Stephenson has been using the same loop to write design principles in the team’s own language.

The surprise hit is personal coaching. In recipe interviews they had people run Matt Mochary’s Coach Me template across their history. Nobody quite cried. Almost everyone had a strong emotional reaction. Stephenson has not shipped anything else that landed like that. Granola has only been in the meetings, and it already knows a scary amount. Labenz adds Dan Shipper’s implicit-culture recipe: scan the company’s calls and report the culture that is actually running, against the website. For a leader that is a harsh mirror. Recipes travel on social media. The growth engine remains people sharing notes, plus teams that now benefit from everyone being on the same context and procurement buying seats. Labenz’s growth lesson: you only need one hook if it really works.

Granola is about sixty people: twenty-five to thirty engineers, three product, three designers plus Stephenson, a couple of design engineers, and everyone else. For designers, actively playing with AI coding tools is a hiring prerequisite. The notepad cannot be judged in a Figma mock. What matters is real content in it, and the ergonomics of glancing from the side of your eye while talking. There is no substitute for a prototype in the real app, used in a real meeting. All the designers now hack into the live product. Paywall and onboarding flows are still easier on a canvas, where the problem is copy and sequence. Lines are blurry: designers submit code; engineers do what used to be called design. Stephenson is having more fun as a designer-builder than he ever has.

The speed that actually changed is idea-to-evaluation. Chat used to live in a sidebar in parts of the app. They argued for a long time about making it bigger and global. Stephenson had mocked versions. The trigger was a live-code prototype of a floating chat they could all live with internally. Within a day it was obvious. They shipped it in September. Figma is not dead; it is specialist. He used to start, middle, and end in Figma. Now he often builds the first thing in his head in the product, jumps to Figma to try ten variants of the bit that feels wrong, then back to code. Full-app schematics of every screen are gone. If Figma can remain the best place to ideate, fine. Claude Code and others will run at that job from the other direction. He would be nervous in their seat.

Culture has to absorb a high idea-to-ship miss rate. They hire for humility and open-mindedness: be curious, assume nine of your ten ideas will die when they hit real people — not because a founder vetoed them. Every engineer talks to users. Friday demo day is the highlight of Stephenson’s week: roadmap work, and also left-field pitches like deleting the sidebar. Trying, showing, learning, putting it to rest is the point. His own ideas are usually terrible too, and that is okay. About 10,000 people sit in beta. Core product decisions are almost purely qualitative gut and observation; growth gets some measurement and A/B tests. Pre-launch they iterated with a handful of people rather than chasing volume. High-fidelity closeness still outweighs talking to many people quickly. By the time something hits beta it is usually baked; the program is bug catching, not a vote.

What scares him most? Labenz offers his friend Andrew Critch’s phrase, the big tech singularity: a few labs so far ahead they can clone any niche, even subsidize it. Stephenson has worried about that since day one. The morning ChatGPT launched you could draw a hand-wavy line to a future where one tool does everything and no startup needs to exist. It is a depressing version that could pan out. He does not have a great rebuttal. He would rather build Granola until that day than sit it out. Tech loves to assume the next big thing swallows all of it. He felt that about Alexa, and about Facebook Messenger chatbots around 2013. Life is messy. Specialized tools keep earning their keep.

Granola’s survival plan is blunt. Architect the product so that every model upgrade makes it stronger. If the next Claude or GPT ships and Granola does not get better, over time they are cooked. Stand on the giants, be a little better in one painful slice — back-to-back meetings — and trust that people will pay for that. Labenz adds antitrust. The positive vision is more concrete than the fear. Study a knowledge worker’s day and much of it is slavery to the laptop: a machine that moves a mouse and shuttles information. Help with the mindless part, and work gets more interesting. The number-one piece of praise they hear is that Granola makes people more present, brain firing, because they are not frantically taking notes. He is proud of that. Played out across the rest of work life, that is the future he wants.