Chris Pedregal + Sam Stephenson: Making Meetings More Effective with Granola

0:05Hey, everyone, and welcome to Generative Now. I am Michael Magnano. I'm a partner at Lightspeed. This week on the show, I spoke with the co-founders of Granola, Chris Pedregal and Sam Stevenson. Granola is a powerful note-taking app that uses AI to compile and summarize meeting notes. Chris and Sam and I talked about their journey as co-founders, how they have quickly become a must-have tool for many companies in tech, sales, recruiting and beyond, and the areas where Granola might be trying to grow next. We spoke in front of a live audience at the Granola headquarters in London, and it was a lot of fun.

0:37So let's get into it. Hey, guys. Thanks, Mike. Thanks for hosting us in your amazing office. Thanks for coming. Yeah. This office is like brand new for you guys, right? A couple months. Something like that. Yeah. Yeah. Sam just carried all the plants in himself. Is that true? I'm not quite. Almost. A few other people. A lot of trucks. Actually, you went to the plant market at like 4 a.m. We did do that. A truck and like carried it up all the stuff. That's like a great example of building a startup. Yeah. Yeah, that was like the least hard thing he did that. All right. Obviously, everyone, I would hope most people in this room are familiar with Granola.

1:18And we want to get into the app layer and what it means to build at the app layer. But tell us about the product. Tell us about the company. How'd you guys get started? How'd you end up doing what you're doing right now? Give us a little bit of the origin story. So this was three years ago. I quit Google. I knew I wanted to do a startup in London. I didn't know what I was going to do. and like within a week of quitting Google, I started playing with GPT-3, the instruct version of it that had just come out and was blown away as I'm sure everyone here at some point in the last five years has been with Alonso.

1:49And I was like, okay, this is new, this is different. I don't know what it is exactly. So I started messing around with it. I knew I wanted to start a startup. So I also started looking for a potential co-founder and I basically was like convinced there's two things at the time. I'm like, okay, maybe if I need a technical co-founder, it should be someone who knows trained models, which I changed my mind on later. And then the other really hard thing, I was like, oh man, there's a bunch of new UI that's going to have to be designed that's AI native. So I need someone who's really thoughtful at that.

2:18And I started exploring these tools for thought forums, and I stumbled across this online meetup called Tools for Thinking. What are tools for thought? Just to help us all understand what that means. You're going to have to cut me off. Tools for thought are basically, Like humans are toolmakers, right? Like my friend Paul who was here like taught me that. It's like humans are toolmakers. It's one of the things that sets us apart from animals. And basically what we are able to do is really limited by the tools we have available to us, right? So classic Steve Jobs, Bicycle for the Mind, right? But like the original tool thought it's language.

2:51It's like written language, right? Then you can look at like mathematical notation, right? Like the, if you're using like Roman numerals, you can only do so much math in your head. Whereas if you use, what our Arabic, like whatever we use today, like you can do much more complicated stuff with paper and pencil. And basically like every development of like human toolmaking has meant that humans can do more and more and more. And I think AI is like the ultimate turbocharger of tools for thinking. Anyway, so I found this guy online in a tools for thinking group. I didn't even meet him. I just saw his profile and I sent him an email being like, hey, hey, do you want to grab a beer sometime?

3:28And somehow he said yes. Yeah. So I think, yeah, we basically both were very aligned from the beginning on, like AI is going to change the landscape of the tools we use. And either the tools that exist right now are going to have to change everything about what they do or new people are going to come in and take over. And so that felt like an exciting place to be starting a startup from. Like it's an opportunity that's just opened. From both of our prior experiences where Chris talked about how we think building a really good product experience is going to be a lot of what matters in making something successful in this space.

4:03And I think to do that, we were both very keen to have a very specific and painful user problem that we could be designing around. You want to be able to picture a person in your head and picture them struggling with a thing so that you can make the tool that that solves the struggle. And so we spent a while just open-mindedly kind of wandering around looking for that struggle, talking to people about their days, trying to figure out where are the pain points, what sucks about people's jobs that we could possibly make easier. And I think a thing that came up again and again was people whose job revolves around meetings and talking to people.

4:43Every time you have a meeting with somebody, that meeting tends to create a kind of pile of follow-up work, whether it's simple stuff like just writing up notes that you care about or sending a follow-up email to the person you met, or whether it's complicated stuff like updating 20 different fields in a CRM and triggering a workflow and an email campaign to somebody, things like that. A lot of people who have calls have a version of those kind of things after a meeting, and they all universally hated doing them. It's all kind of menial work that isn't what you get energy from in your job. and it felt like the kind of stuff that AI was like primed to be able to help with if not then when we were doing it at least in a few years time and so we started pushing on on that on like how can we how can we kind of make a tool that is in your meeting that eventually will be able to help you do a lot of this kind of menial work that happens around meetings.

5:39Chris said that it's going to be so important with AI to build a great product experience. So I totally agree. But I would say that for a lot of people in the AI community, in tech, in VC, in startups, that was not obvious to many people a year ago. I'm sure you guys remember a year, year and a half ago, there was all this talk about, oh, if you're building at the app layer of AI, you're just a rapper on GPT-4 or GPT-4, you know, whatever, anthropic Claude. Now it feels like we've done a 180. There's so much excitement about the app layer. And again, you know, Granola is an example that's often cited as being, you know, one of the one of the potential winners.

6:20What changed? Why is there? Why do you think I know you guys haven't changed, but why do you think everyone else has sort of changed their mind about this opportunity? I think a few things happen. One is these models just kept getting better and better, faster and faster. And it became very clear that it was just, it made way more sense to just use the best like frontier model out there than try to train your own thing. You're always going to be slower, right? That's one. Two, super hard and expensive to train your own model. So it's going to be a couple big shops that are going to do that. And then three, man, you get all the benefits from those models.

6:56So like if you can apply it to the right use case, it's really powerful. And And I think our view of that from the beginning has always been low frequency use cases that are maybe non-critical are going to be eaten up by the general agents. So I think if it's like a consumer use case that you do like twice a month, it's definitely going to go to ChatGPT or Anthropic. If you are doing something that really matters, like it's like professional tooling where your performance really matters and you want to optimize for that use case, then bespoke tools that are optimized for that are going to be way better.

7:30And I think that's what you're starting to see things like, like Cursors valuation, and like Windsurf just got acquired. I think it's like prototypical, like they're just like wrappers on, on Sonnet 3.7 or whatever. Right. But actually they're like amazing. And there's so, it's hard to build great software and the delta of, if you're using one of those products, it's how much more productive you can be really matters. And I think now the market's kind of, I think this thing has like a pendulum, right? We get really excited about one thing and then it's like, oh, there's a glimmer of the future might be different and people get very excited about that.

8:01We think that's sticking power, right? Because we think the tools you use matter and we think the like professional tooling has always been a thing, right? And if it makes you 10, 20, 30%, 50 % better at your job, like that's going to always have a lot of value and economic value. It does mean we have to be quite selective about the things like the challenges that we choose to bite off and which we choose to like leave alone. When we started, which was like GPT-3 time and real-time transcription had kind of just become a thing that was available by an API. But it wasn't great. You know, like transcription was obviously bad in a bunch of ways.

8:39Like it would kind of miss things. The notes that we wrote weren't amazing just because the models weren't amazing. We had tons and tons of conversations about like, you know, like, should we, you know, what should we be investing our time in? Because like the, some stuff is just going to keep getting better without us doing anything. You know, the quality of the AI output, the speed, the cheapness, all of those things. And some things are not going to get better unless we like push really hard on it and try and figure out what a good solution is. And so I think a lot of the game for us has been like picking our battles and like knowing what to innovate on and what to just like wait for it to get better, you know?

9:15Give us some examples of some things, you know, real-time transcription is one of them, but like what are some other things that you, you know, very intentionally decided not to work on? And maybe what are one or two things that you did? You're like, oh, this is our job to solve. Granola can be the best at this. The obvious one to me was language support. Like when we launched, it immediately was like the most requested thing for Granola was to support multiple languages. I think it still is. And we spent like a week working on it, like trying to figure out a good interface. It was like kind of available in one of the transcription providers.

9:48I guess it wasn't great, the language situation now, which meant that we would have had to, it was looking like it was going to be like a few weeks to a month long project to make a good interface to help you pick the right language for the right time that the meeting that you're in, which is a huge investment, like a month of time, right? And that to me feels like a thing that there are dozens of companies out there really incentivized to figure out multi-language real-time transcription models. And if we just wait, like it's going to happen and the experience will be way better in that world than anything we can kind of think up to like hack around the fact that it's not good right now.

10:22Yeah. Another example is like context window lengths. It was too small and we launched, you can only do like 30 minute meetings. And we could have done a bunch of work to like, okay, try to chunk that. Or we just, we just wait a little bit and the context windows got larger. It just keeps getting bigger. Yeah, exactly. Another one's RAG actually. Can you explain to people? Sorry, RAG. Retrieval augmented generation. Basically the idea is like, so context window, most people here probably know this, but a model can only take so many tokens, so much context into its memory. And if you have more, let's say you have like in our case, a repository of meetings that are larger than a context window, you have to figure out like which of those do you put into the memory.

11:00And like there's all these naive approaches where you basically kind of do a search across those and you choose a subset. Doing that well is hard, right? Doing it not well is super easy. Doing that well is really hard. But context windows keep getting bigger. So you can get away with by just sticking a lot of stuff in there, which is like, and in some ways it's like an unintuitive, I think if you put your engineering hat on, you're like, oh, that's wrong. You know, we should, we should engineer this. Throwing more stuff into the model. Yeah. It's like that, but it, it, it, Unstructured. It, like AI breaks intuitions, man.

11:29It's like, um, sometimes you're like, oh, it's like, it's very imprecise. We're putting all this stuff in there. But like, I think these models are smarter and more intuitive than we expect. And sometimes though, the moments where I'm like, oh shit, are usually where it picks up on something I wouldn't have expected a machine to pick up on. You know, it'll be like, actually in that meeting six months ago, you said this thing. And, uh, and then now you said this and like, we wouldn't put that in the rag if you stick it on the context window, like sometimes magic comes out. Right. Let's talk about the business model of building at the app player.

12:03It feels like so many companies right now are basically just, just charging for inference, right? We see all these products that charge for credits. And really what those credits go to are just hitting the model, right? How do you guys and Granola think about the business model? Like, is that the type of business model these companies should be pursuing? Should they not be pursuing? What's the opportunity to build like a huge business at the app layer on top of the models? I think a lot of the way we think about it is probably not too, like our business model is probably not too different, like pre or post AI.

12:35Like I think ultimately we're trying to make a tool that's like valuable enough that a company will give us money for it. And there are things that we want to push on to kind of make the thing feel more valuable, which are not really to do with AI, to do with like team app building, I guess. Like if we can unlock network effects in Granola where like there's a, that Granola gets better, the more people in your team are using it. And it becomes like this valuable repository in and of itself. I think that's the thing that we have a lot of signal that companies will pay good money for. And it's kind of independent from AI, I guess, although the AI enables you to do full stuff with that.

13:11If you're just monetizing the AI, you're effectively just like a reseller for OpenAI. Whereas if you charge for the repository, you're charging for granola, something only that granola can provide, right? Yeah. I do think we're an interesting moment in history because it's kind of a land grab right now. There's new products that are possible that that couldn't exist before. And we know that the cost of running these products two years from now will be vastly cheaper than they are today. So there's like, and maybe that will always be the case, but it's kind of easier for me to just think about the next few years.

13:42So we're in this world where it's gonna be cheap to run granola or, I don't know, Pica or whatever you want two years from now, but it's quite expensive to run now, but there's a lot of user demand. So like, what do you do in that kind of situation? And I think it's the kind of thing where you have to have, I think you have to build for the future, right? And you have to figure out how to make that work for your company. Because if you build for today, I think you'll make all the wrong optimizations, which does mean it's a capital intensive play. Right. When we forecast our finances into the future, if you don't account for things getting cheaper, then it gets really fucking expensive really quickly.

14:22Well, with exponential. With like 10 % week of a week growth. Yeah, yeah, yeah. Yeah. And so, and so, I mean, part of, part of the company's bet, I guess, is that, is that this stuff is going to get cheaper and, you know, and there's going to be ways, some things will always want to be on the frontier of, I think, like I could see like, um, uh, being able to do chat and document creation on top of the huge body of all of your company's meetings is like the kind of thing where just the more power, the better, but like transcription, I could see hitting a ceiling where where there's a point where it's good enough and then and then like and then again it's just like cool let's now get that transcription for like as little money as we can so that so that you know that's our main running cost like that can that can kind of go away.

15:03Lightspeed you know was one of the first investors a couple of years ago and I will say it was so fun for me and I know everyone else on the team watching you guys build from day one just like from zero lines of code to what it is today. And, you know, I remember the launch moment. May 22nd. May 22nd. So we're coming up on a year. That launch moment, it was amazing. It felt like almost instant product market fit, which is so rare, never happens. I gotta ask, like, tell us about the process of building the first version of Granola so that you could have that day you launched, which again, it's nearly impossible to do.

15:40I don't know. You know, we try and be deliberate about things, but something's kind of by accident too, but I think the things that we were deliberate about were the hard thing about Granola was probably going to be figuring out what's like an interaction or what's that lets a user get the stuff they care about out of a meeting in a way that feels really natural and really effortless. Like figuring out that is like, it's just a huge part of what we have to do to make the product successful. And so I think the first, I don't know, six, nine months of Granola were like just experiment after experiment after experiment, like trying things to figure out what that might be.

16:15You know, we'd build a thing, put it out in the world, be constantly talking to new users and watching how they react to it and how they use it. And, and over time, we, we, you know, threw away a lot of the things we tried and were able to hone in on something that felt like it could work. The first six months was like a gradual growing of complexity in the thing as we like threw more ideas into it, you know, trying this and that and this and that. At some point, I think we felt like we maybe found a thing that could work. Just like, you know, type your notes at the end, Granola fleshes it out on the same piece of paper, that kind of thing.

16:48And then we kind of went through this process of like cutting back and streamlining everything until it was really just that feature. And that's what we launched with. Throughout this, we were kind of like, the goal was to build a daily habit for our users. Like, can we make this a daily use product in the small number of beta testers that we had? And we had this chart called the dot plot, which is like you can see each individual user that uses Granola day by day and how many meetings they did on a given day. And that helped us be really honest with ourselves about like, is someone reliably picking this up and using this in their meetings or are they just kind of dipping in and out or, you know, is it kind of random?

17:26Yeah, we were in closed beta for a year and we had about 150 people that we had onboarded by the time we decided to launch. And we manually onboarded all of them at that point. And I guess looking back on it, it's so funny. We never, like, that's only Connect looking back, but I didn't really think I knew what was ready when we launched it. Like, Mike pushed us to launch. Typical VC. actually been pushing us so long for about nine months before that and we held them off for nine months but i you know it's like oh at that point all we could see were the things that were wrong with it um which is like an interesting lesson right because uh once we put it out in the world it just kind of it actually hit a bunch of quarters uh but we didn't necessarily appreciate the depth of that until we put it out there sam i've heard you talk a little bit about your design process and about how the team really thinks about designing for what people actually need, not what they think they need.

18:26I've heard you use the term lizard brain. Explain. In building software, it's really easy to, I speak as someone who's done this over and over and over again on things I've worked on. It's really easy to get theoretical about what a user might want. This thing would be cool. I've got such a good feeling about this. I think this is how the app should be. And when you interview users, you know, they can tell you all of their great ideas for the product. And it's really easy to just build what they want because they're asking for it. One thing that we were kind of paranoid about from the start was, I guess, especially in our use case, like meetings are a super high stress situation in that when you're in a meeting, especially like a back to back meeting where, you know, where you're, maybe it's two minutes past the hour, you're already late for your next meeting.

19:13You're like, shit, you know, trying to make excuses to get off the call. And then you get off the call and then you've got to rapidly get into the next one as quick as you can. And then you're like, oh my God, who am I talking to? Why are we doing this? You know, all that stuff. You have so little brain space for a piece of software at that moment to try and help you. Like you're just trying to deal with the basics of getting the next person in front of you. We just have like this, I don't know, 1 % of your brain to play with, you know, as people designing a product. And I think keeping that in mind, like keeping the kind of stressed out back to back kind of moment in our heads as we were designing it, like helped keep us honest to what's going to fly, what's going to work in this.

19:53And I think people often talk about how simple granola is and how it feels nice because of that. I think that's just a function of like, we really can't put many buns in front of you when you're in that situation. You don't have the headspace for it. That's really cool. Chris, you built and scaled and sold another company before this, Socratic. I had the pleasure of watching you do that as well, because the company that I was building was on a street right behind you. Yep, in New York. That was a while ago. And, you know, one thing I often think about is, especially with you guys building this, is wondering, like, what's it like to build a company now with AI versus building a company that didn't have AI?

20:34Like, what's the difference in building companies across these two eras? Oof, ask me that in two or three years. I think I'll have a much better answer. Well, I guess one is, So Bas, our CTO, who is not here right now, I look to him because I think he's the best at this on the team, but he really, really pushes us internally to use AI as much as possible. So it's like an active goal to reduce the number of lines our engineers write every day. And I think that you actually do need to push people for that because we all have habits. We've been working on, we've been doing stuff for a while and the world's changing very quickly.

21:07So if the org isn't doing that, then you're missing out. The other thing is, I think people ask a lot about like, okay, what's the makeup of a company going to look like in this post-AI world? How big does it have to be? My view there is that like, I don't know what it'll be like in five years, but for us, the product is core. So we need a bunch of really thoughtful, best in class people working on the product. There are other functions where I, like in the past, we might've built a really big customer success function where I don't expect us to do that. I expect us to use whatever the best and greatest AI tooling is.

21:42And we'll still have a great team there. It's just how that team spends their time. And they might look more like engineers in a way, in terms of building systems, even though they might not be writing code. And the last one is the world is changing and everyone's watching and interested and wanting to try stuff out. That wasn't the case with my last startup. I'm used to startups being a slog of you fighting so hard to get people to care about what you're doing. and I kind of feel like the rug got pulled out from under me with Ranolo because we put it out there and we're like, all right. So I'm like, no one's going to care about this.

22:14You know, I have to like keep working on it, keep working, keep working on it. And all of a sudden, like it just started growing and then stuff just started breaking and trying to mentally prepare for that. So that's like macro environment questions. Those things change quickly, but that's been a defining part of this journey. It's just trying to keep up with the change and keep up with the growth. And I think that inevitably forces you guys and really any team building an AI right now to just move so freaking fast, which inevitably creates a different type of challenge for the company. It's like, how do you maintain quality?

22:46How do you maintain taste? Right. Like taste has been this thing. I feel like it's become this really annoying word, actually. You got to have taste. But I think, again, Granola gets cited as one of these products that just like, oh, it's beautiful, great design, amazing taste. Like, how do you think about maintaining that when you're moving so fast and when you're building a team? It becomes so important, I imagine, with each and every person you bring on. I think we do all right at this, but I think there's much more we can do to make this better. But I think things that I think we do well, a bit of work for us.

23:23we screen engineers as part of the interview process for product thinking, I guess. Can you think from the point of view of a user? And when there's a technical problem put in front of you, get to the why of why is this a problem for the user? And that helps you make the right trade-offs in cutting the scope and really just building the thing that's going to solve the person's problem, not like this beautiful technical masterpiece of an execution. there are there are types of features where you can where like um once we have good systems set up like you know the ui of granola's kind of figured out you can just kind of like ship and iterate and push stuff out very quickly there and and we don't need to kind of be so cautious about about that stuff um you can always roll it back you know a couple days later um and and then that way we can kind of like help like reserve our judgment and the taking the time to kind of pour of the details on the things that really matter or kind of like in the core flow of someone using granola?

24:21New primitives in the app. But basically we're trying to get better at the one-way door versus two-way door. Yeah. So it's like, if it's a two-way door, can we just shift changes quickly, see what people think and go from there? That said, I think what people love about granola is that it's simple, minimal, and gets out of your way. And you add 50 buttons in there with new features, you kind of kill the golden goose, right? And I think we're figuring out how to find that balance because we do have to move quickly, but we also need to keep the soul of the product intact. I want to talk a little bit about building a team here in London.