This AI notetaker won't sell surveillance to your boss

This podcast mentions Anthropic a few times; my fiancé works there. See my full ethics disclosure here.

Last week on the Platformer podcast, Raycast’s Thomas Paul Mann predicted that within a few years, as much as half the apps on your Mac could be ones you made yourself. For the third episode of our miniseries on work in the AI era, I wanted to learn more about a tool that may already know more about your job than any other: the AI notetakers that increasingly listen to every meeting we take.

And so this week I talked to Chris Pedregal. Pedregal is the co-founder and CEO of Granola, which has been my own AI notetaker since shortly after it launched. Pedregal started the company after a long tenure at Google, which he joined as an associate product manager in 2008. He left the company after a few years to start Socratic, a homework-helper app that Google acquired in 2018. In 2022, he left again to play around with the pre-ChatGPT large language models from OpenAI, and soon he and co-founder Sam Stephenson went on to launch Granola.

In an ocean of chatbot products, Granola is the rare AI tool with real product polish. Granola captures the audio from your own computer, lets you type your own notes alongside the AI, and fills in the gaps with an instant AI summary as soon as the meeting ends. (The app’s surprisingly good summaries were ultimately what prompted me to become a paying customer.) The company was most recently valued at $1.5 billion.

Unlike some rivals, Granola is invisible to the person at the other end of the call — something that has generated some controversy among workers who are leery of being secretly surveilled. Pedregal told me the design is an artifact of how difficult other meeting software make it for third-party bots to join conversations, and walked me through the lengths Granola now goes to in order to disclose itself — displaying an animated watermark on your video feed through a virtual camera, since Zoom and Google Meet offer no better way to ask a room for consent. Pedregal also made a prediction that may chill some privacy-conscious workers: "I think we'll probably end up in a world where in work meetings, the expectation is that you are being transcribed by default, and you ask not to be transcribed."

Pedregal shared some product news on the show as well: Granola, which has never stored audio recordings, is now researching whether to automatically delete verbatim transcripts after some period of time, while compressing the contained knowledge into a working memory. The company is also prototyping "enhanced transcripts" built for AI agents to use as part of a corporate knowledge base. 

But what struck me most was Pedregal’s attitude toward actual surveillance. Granola notes are private by default, and "companies hate us for this," he told me. CEOs regularly ask for access to every employee transcript, and Granola refuses. "If you have a backdoor way for someone in power to use all this context to make judgment calls in that way, that doesn't feel right to me, and that's not a future we want to build towards," Pedregal said.

When I asked my standard question about the risk of AI job loss, Pedregal — who runs a company of around 80 people, and could see it reaching no more than 120 — played down maximalist predictions of labor disruption. "People thought when the computer came out, we'd have lots of leisure — and no, the opposite happened," he said. 

As for the seven other apps that now offer to take notes every time I start a meeting, he is unbothered: "I think we are, like, 5 percent into this new technical revolution," he told me. Transcription, in his view, is merely the word processor of this era: "Spreadsheets haven't come out yet. The GUI hasn't come out yet." The real fight, he says, is over what interface we'll all be working in a decade from now.

In some ways Pedregal’s project reminds me of the original version of Slack — a company whose name is an acronym describing its onetime vision to become a “searchable log of all conversation and knowledge.” Granola is building a log not unlike that one. The question now is whether it can become as essential to the modern workplace as Slack has — or whether, by letting you grant AI agents access to your conversations via MCP connectors, it risks giving away the game to OpenAI or Anthropic.

A long excerpt of our conversation is below, edited for clarity and length. Listen to the entire conversation wherever you get your podcasts — just search for Platformer — or watch it on YouTube at youtube.com/caseynewton. And let us know what you think — we welcome your feedback at casey@platformer.news.

Casey Newton: I find myself using Granola for everything, but maybe searching through my notes less than I expected to. I get a lot of peace of mind from capturing everything, but it turns out I don't really need to check back for what this one person said. How true does that prove across your user base?

Chris Pedregal: I go back less than you'd expect I would, and some companies would look at that and say, okay, how do we drive that number up? I actually think forgetting — not going back and looking at stuff — is a feature as much as a bug. Our brains are actually pretty decent at remembering the most important things. What's powerful is that moment when you're like, ah, crap — what did I say in that meeting? Just knowing that you have it is very valuable.

The thing that has changed: our MCP usage is the fastest-growing chart that we have, and agents have a lot more cycles, so agents do go back and look at stuff. I think that trend will only increase. Transcripts are this goldmine of context. Out of the box, agents are really intelligent, but they know nothing about you. And it turns out we do tons of meetings, and speaking is pretty effortless. There's this context exhaust that you're naturally putting out, and something like Granola just captures it.

Newton: I want to ask you about a design choice you made that I think is very powerful, but that people also have feelings about, which is that Granola is invisible. There had been other AI notetakers before, and often they would involve adding a bot to a meeting. How did you come to a place where you said, we're just going to record the audio off of a laptop?

Pedregal: The reason we made that decision was not so that Granola was hidden. I really don't like bots, but that wasn't the original goal. The original goal was that we thought it was really important for Granola to always work. It didn't matter if you were in a Slack huddle or a Zoom meeting or an in-person meeting: if you hit start, it just starts transcribing. And the only way to do that is by using the system audio on your computer. We also didn't like bots because we thought they're a really inelegant, dehumanizing force in a meeting.

The downside of our approach, though, is that when you use Granola out of the box, other people don't know you're using it. So we're placing the onus on the user to tell others that they're using Granola. And that's actually not our choice. If there were an easy way that didn't involve a bot to make that visible, we would do it — that's where social norms are going anyway, and it's just good for everybody to know whether you're being transcribed or not. The problem is that the Zooms and the Meets of the world make this almost impossible, because they're pushing their own note-taking systems. There's no API call where we could ask everyone in a meeting for consent — we would do that. So we've had to go to some crazy lengths. I can show you what we built: if I turn on Granola right now, do you see this watermark that shows up at the bottom of my screen?

[Pedregal turns on Granola, and a small animated Granola watermark appears on his video feed.]

Pedregal: This is something we built, and users can turn it on. Any company that rolls out Granola tends to want this on, because they want visibility. But the way we had to build it was kind of nuts. This is actually a virtual camera on your computer, and we are injecting this into the video feed before it hits the Zooms or the Google Meets out there, because that's the only way we can control how it shows up and make sure it shows up.

I do think social norms are going to evolve, I think the video conferencing platforms are going to give us more control, and I think we'll end up in a good place. I think we'll probably end up in a world where in work meetings, the expectation is that you are being transcribed by default, and you ask not to be transcribed. I hope that in social environments, the default continues to be that you're not being transcribed, and you ask to transcribe.

Newton: It sounds like you've basically accepted the privacy critique.

Chris Pedregal: It's something that's been on our mind from the beginning. Another design choice we made — unlike, I think, anybody at the time — is that we never stored the audio. We have access to the audio, we could have access to the video, and we never, ever stored the audio, because we wanted Granola to feel like a notepad. At first, we were actually going to delete the transcripts: you'd get the notes, and then we'd delete the transcript. We realized we shouldn't do that — the transcripts are super valuable, and if you want to regenerate your notes, you need the original source material.

Every time a new technology comes up, you have to be thoughtful around the design choices you make, and then society will adapt. I'm very glad we decided not to store the audio — that was the right call. Right now we're looking into whether we should delete the transcript automatically after a certain amount of time for all users. There's a lot of value there that's lost, so we're doing research into compressing it. You probably don't need a verbatim transcript of a conversation you had three years ago, but you might want to still go back and chat with that meeting. There's usefulness and invasiveness, and you need to thread the needle between those two, so you get as much usefulness for the least amount of invasiveness.

Newton: Last question on this point: I recently read a story in the Wall Street Journal that at least one startup founder is using Granola to record her first dates and then have Claude analyze them. Is that a recommended use for Granola?

Pedregal: No. I mean — if they're consenting adults, then go for it, ideally with everyone okay with it. But no, it's not a recommended use. We have built Granola for the work use case. A lot of people are interested in the always-on wearable thing in your social life or personal life. That's not what we're interested in doing. But I wonder if Claude gives good dating advice. I don't know. I've been married for a while, so maybe it could give me good marriage advice.

Newton: Let's talk about agents. I've been using Granola for well over a year now, so it knows a lot of things about me. What are some things you're thinking about being able to do with that, and how do agents get involved?

Chris Pedregal: I think it's a little bit harder today than it should be. You can connect Granola to Claude, or go into the chat in Granola, and ask questions that reference that context. But if you've used Granola for any reasonable amount of time, there's just too much context to search quickly. So Claude will do a relatively dumb job at extracting information — it'll look at the 20 most relevant meetings. But really, what I think we should do is compress that down into apple juice from concentrate. We should have that concentrate — that compressed information about Casey — ready, and that's what should be sent to all the other tools and systems and agents.

So we're in this awkward phase right now where we have all the context, but it's not yet in the form that's most useful. And then — what do you use it for? It's basically black and white. Either you've realized that you can use AI to help you do a lot of things — maybe most things — in which case the critical piece is that the AI needs to have the right context. Or you haven't.

Newton: Before we hopped on today, I asked Granola to give me a critique of my meeting style, and it said: you sometimes let others set the agenda too much — which is a very polite way of saying that I get bored in meetings and stop paying attention. But it's absolutely true. That's a useful piece of information for me, but I don't think the best form factor for that is me going in and asking Granola. It feels like it wants to be something more proactive.

Pedregal: 100 percent. We're in the phase right now where, if you put in a bunch of work, you can create these workflows and make use of the context. But it's not in a state where your average person will just get value out of it. Here's an example: we launched this thing a little while ago called pre-meeting briefs. Granola will write a little brief before you have a meeting with an external person. It's one of the most obvious use cases, right? Anyone who's AI-native has gone into Claude or Cursor or Codex and set up a daily thing that looks at their upcoming meetings and writes them a little "what I should know." It's easy to set something up. It's actually kind of a pain in the ass to make it good. Ours still needs a lot of work, but it's a lot better than what most people had set up for themselves, I guarantee you.

And that's just going to happen with everything. Right now, things are technically possible, but the bar to do them is just too high for most people. That's why it's an exciting time to build product. If you think about the 50 most common use cases that are done with AI, I don't think those should be individuals prompting AI, going into a chat and typing. I think that should be productized, and it should be proactive. But those things are always harder than they sound at first.

I'll tell you some of the things we've figured out so far. One is that people don't really want another inbox of shit they need to do. You already have a lot of to-do lists — your personal to-do lists, your email inbox, Slack, maybe Twitter DMs for you. If all you're doing is adding another place of things I need to check, of work I need to do — I don't know if that's really adding to people's lives. I already feel pretty overwhelmed and stressed out in my work environment. I don't need Granola to make me feel more stressed out about the things I'm not going to get to.

And then another paradoxical thing: oftentimes, we've found that AI is almost most useful with the smaller tasks rather than the bigger ones. The most important things going on — you're probably thinking about them already, and there's not that much for AI to do. But writing this person a thank-you note because they were really kind: maybe you just wouldn't get to it normally. But if AI says, here, I wrote it for you — do you want it? — then you're like, oh, thank you. I feel like I'm a better person; I sent that person a thank-you note. At the end of the day, I think it's all about what the AI can surface proactively, and doing that in a way that's not overwhelming, not distracting, not intimidating, not dumb. That's the name of the game.

Newton: I wanted to ask you one more thing about all this context that you're gathering on people. When Granola is applied in a big company, you're developing something that sort of looks like psychological profiles of the entire organization. I wonder if there are lines there where it felt like: let's not do that. That might feel creepy, or this might not be good for the company.

Pedregal: It's something we think about a lot. In the same way that we decided not to store the audio, the other really non-obvious thing that we did is that all Granola notes are private by default — you're the only person who has access to them. Companies hate us for this — they're like, hey, there's all this content; why am I paying you money? And we're like, no. We're capturing context, but context can be very private, very sensitive, very nuanced. So even though your company's paying for it, those notes are private. Legally, if the company gets sued, we have to hand those notes over — it's the same thing with Gmail. But it's not like your manager, or your manager's manager, can look at those notes or run an agent on them. There are some real trade-offs for us as a company in doing that, and it's something we believe in.

I don't think the line in the future will be drawn so much on whether something should be transcribed or not. The interesting questions will be: should this context be shared, and what part of this context should be shared? Say you have a one-on-one. It can be very valuable, at least for me, to ask: what exactly did I talk about last time? How can I improve? Other people probably shouldn't have access to that. But maybe you have a really open culture and you want to proactively share it — that's cool. There needs to be human judgment and philosophy in how you do that.

And here's the bit that's going to get really interesting. A classic use case for Granola is sales calls. If you're doing a sales call, having great notes and great transcripts is really valuable. Right now, you can either choose to share the notes and the transcript with other people or not. Something we're working on is that you can modify what version of it gets shared. We'd like to strip out all PII — all personal information. We'd like to strip out any conversations about your health or your family. Those are still valuable to you. Say we were friends: I might still want to remember — crap, what's Casey's kid's name? But I don't need my work colleagues to know that. The nuances of human relationships and how we work really matter, and the tools we design need to fit into all these different situations.

Newton: I think you could probably use Granola pretty easily to figure out who is a bad manager at a company. Or, as an employee, being able to ask Granola: hey, I'm being treated in this way, and this doesn't really seem great. Does that feel like an opportunity to you, or is that something that stresses you out?