Granola’s Chris Pedregal: people are starting to question what’s actually useful

Granola’s AI note-taking app has become a viral hit in Silicon Valley since it launched in 2024. It automatically transcribes meetings and then turns them into useful summaries.

Customers include several other hot AI companies, including Cursor, Lovable and Replit, as well as venture capital firms, such as Index Ventures, which led a $125mn financing in March that valued Granola at $1.5bn.

The London-based company was co-founded by Chris Pedregal and Sam Stephenson. Pedregal sold his previous start-up, the AI-powered education app Socratic, to Google in 2018.

Here, he talks to the FT’s global technology correspondent Tim Bradshaw about how AI is percolating through the workplace, what it’s like building on top of rapidly evolving models and the privacy implications of always-on AI listeners.

Tim Bradshaw: Why do we need an AI note taker?

Chris Pedregal: This is going to sound ridiculous but if you look at the history of humanity . . . essentially what humans have been capable of has been defined by the tools that we use. I’m talking about written language, mathematical notation and the printing press.

And I know that’s a bit of a ridiculous start to this question, but I actually think a notepad is one of the best tools that we’ve come up with. It’s simple, it’s reliable, it’s open-ended. It’s a way for us to extend our memory and to organise our thoughts.

The promise of an AI notepad is that it can write all the boring stuff for you, capture all the rote stuff or specific details that might be important when you go back and look at it, but still give you the space to lead the thinking, decide what’s important, have better ideas and better thoughts than you would without the tool.

TB: When you started using large language models for the first time, why was Granola what you decided to build?

CP: I started playing with LLMs eight or nine months before ChatGPT launched [in 2022]. It was a beautiful time because it didn’t feel like everyone was sprinting, so I could actually take some time and play.

In that time, before everyone was talking about them, it became really clear to Sam and me that all the tools we use for knowledge work were going to be reinvented on the back of these LLMs. As I started building these prototypes, it became very clear that tools having an understanding of who we are, our context and what we were trying to achieve, should just lead to way more helpful tools.

Even today, [something like] Microsoft Word, it’s a completely generic experience. Two or three years from now, I think the idea that you and I are going to have the same experience [with] a piece of software — that software is not going to be personalised or the experience is not going to be bespoke to me — is going to seem weird.

Our first pitch deck was for “the contextually aware workspace”. Then we were like, OK, where do we actually start? It turns out meetings are extremely rich in context, and people do a very bad job of capturing and making use of that context. We had different ideas that we put in front of people, but their eyes only lit up when they saw an early version of Granola.

The vision was always much larger, like: how can we actually use your context to help you do work? Granola is only just starting to do that at the margins

The vision was always much larger, like: how can we actually use your context to help you do work? Granola is only just starting to do that at the margins

The vision was always much larger. It was like: how can we actually use your context to help you do work? Granola is only just starting to do that right now at the margins. It’s really mostly used as a context-capture and retrieval tool, less of an action tool. But that’s still the vision and still what we’re working towards. It just turns out capturing meeting notes is hard and really important to do well.

TB: How has the market for AI apps changed since you founded Granola in 2023?

CP: In the early days, we had to handhold so much more to get decent quality out of the early models. The models do much better out of the box now. They can also handle much more context. Generally, the quality of the output you’re going to get is going to go up the more relevant context you give the model to do the task.

I’d say over the past two years, a lot of companies, at least in tech, have been [thinking], “Oh, the age of AI is upon us, I have to lean in, I don’t want to be left behind, I want to use all the tools.” And now, people are still very excited about it, but it’s a little bit like ‘OK, which tools are actually being useful, how much are we spending on tokens [a measure of AI computing power]? Do we actually need to use Claude Fable [one of Anthropic’s most advanced and expensive models] for that?’ So I think we’re in a moment right now where there’s a bit of a questioning of what’s actually useful and what are the actual unit economics.

TB: How does that affect your business? Can you pass on those token costs to customers?

CP: For us, we’re right now in this transition between capture and retrieval [of meeting notes] versus helping you actually take actions [based on those meetings]. That’s a moment where we think we’ll have to change our pricing model. Right now it’s a fixed pricing model: you pay a certain amount per seat per month and it doesn’t matter how many meetings you do or how much you chat with [Granola’s chatbot]. So some users are much, much more expensive than others for us. But I think if users can use us in a more open-ended, agentic way to do work, then we have to move to variable pricing in some shape or form — just so that the 1 per cent of users that would bankrupt us are going to pay more money than the average user who’s perhaps just doing the regular things that you’d expect.

TB: What are the new ways in which Granola is trying to help people ease their workload?

CP: Our bread and butter is meetings, people do lots of meetings a day with us. And there’s just a lot of work that’s very tied to meetings that I think we are very, very well positioned to help with. It could be as simple as if in a meeting you say, “Oh we should have another meeting about this” it actually creates that calendar event for you and shows it to you and asks, “Hey, do you want me to schedule this?” And again, that sounds silly and small, but for all the talk of agents out there in the world, that’s not an experience that most people have that actually works. So our plan is basically to take that step-by-step-by-step, so we’re actually useful in increasingly large ways to our users over time.

The trillion-dollar question that everyone in Silicon Valley seems to be chasing is: how is AI going to help me actually do work? I have very high confidence that a lot of knowledge work will be done with AI. So it’s going to change a lot how we work. I think what’s not clear to anybody is what does that actually look like. Are we all typing into a chat box all day, every day? I don’t think that’s what it’s going to look like. Is it one tool — are we using OpenAI or Anthropic for everything or do we have lots of tools? Or do we have lots of tools that work in the back end and there’s one user interface? I think those are the interesting questions.

It’s the kind of thing where it’s very hard to sit and predict the future from what you know. It’s a little bit like the internet has just appeared and we’re trying to guess what types of websites are going to be popular or not.

TB: For all the talk in Silicon Valley about “taste” being so important in the AI era, there are still a lot of AI apps that are just being “vibe coded” and thrown out there. But actually, what things like the iPhone App Store have taught us is that polishing and paying attention to the details of the user experience can really pay off.

CP: Just because we can build [software] faster than we ever could in history, it doesn’t mean that people are adopting way more products.

It’s very easy to build things with AI now — but it’s very easy to build a first version of something, not something that’s actually really great. And therefore I think in AI you end up with a lot of products that sell this massive vision that maybe the world’s not quite ready for. And then maybe the product quality is not at all good enough. I think people are a little bit fatigued with that.

When you look at companies, oftentimes their cultural DNA is forged by where they started. And I think the reasons why we did well were that we were very user-centric in the design, and we took that to an extreme. And that’s why the form factor of Granola is completely different from an AI bot that joins meetings and creates generic notes and shares them with everyone automatically. We also spend a lot of effort trying to make the quality of the notes actually useful and we try to keep the app very minimal.

We have kept that philosophy but there’s a cost to that, because it means you try to ship things that are nice and it takes time. In a moment where everyone can build so many things so quickly, it’s a constant tension internally.

TB: How has the cost of using the AI models that sit underneath Granola changed over time? How do you choose which model to use and how easy is it for a company like Granola to switch between them?

We always use the most expensive, real-time, top-of-the-line transcription model, which is quite expensive. And in the last two years the cost of that has dropped precipitously

We always use the most expensive, real-time, top-of-the-line transcription model, which is quite expensive. And in the last two years the cost of that has dropped precipitously

CP: Historically, our biggest cost has always been transcription. We always use the most expensive, real-time, top-of-the-line transcription model, which is quite expensive. And in the last two years the cost of that has dropped precipitously but the quality’s much better too. And then the other inference cost is just running the models. The models have gotten more intelligent. The cost of running note generation has gone down.

The problem is — and I don’t know if it is a problem — people just do so many more other AI actions in Granola. So the number of users who are chatting with their notes or chatting with Granola has gone up . . . significantly.

We want to provide a really good experience there. At first that was a simple one-pass process — you ask a question, the chat would come back. Now we give it all these tools. It’s an agent so it can look up things from the web, it can look up different notes, it can deep-dive into different parts of transcripts. So what used to be one chat call is now kind of the equivalent of 20 chat calls.

And that’s just chat. We also have this feature where if you’re going to have a meeting with someone outside of your company, it goes off into some research and writes you a little brief. Pre-meeting briefs cost us an arm and a leg. They’re horrendously expensive. They cost, I think, as much as all the notes that are generated on Granola.

We know the notes experience, we have product-market fit, users love it. We’re always trying to make it a little bit better, but that’s in one box and that’s scaling. All the other ways where Granola tries to be helpful to users, that’s still in the ‘what actually works here?’ bucket. If we decide briefs aren’t useful enough to users, we’ll turn them off and we’ll try something else.

In this world, the speed of figuring out what are the features or the form factors that actually resonate with users and are worth investing in really does matter. Therefore, we’re very much [taking the approach of]: build something, even if it’s very expensive; launch it; decide if it is something we want to double down on or not; and then figure out the cost. And pre-meeting briefs, I guarantee you, at the same level of intelligence, that cost is going to drop dramatically over the next 18 months. So I have this conviction that you should try to build for the future.

TB: Which models do you use the most and how easy is it to swap them in and out? How sticky is any individual model provider?

CP: We use Anthropic, OpenAI and Google, and maybe one or two others.

TB: Any open-weight models [which are cheaper and more adaptable than the big US models]?

CP: Internally, not for any big production feature yet. Not against it, it’s just that it just hasn’t happened yet. Technically, it’s trivial to switch from one model to another. But if we talk about note generation, it’s interesting because actually every model has its own personality and its own quirks.

We are always trying to provide the best experience to our users, so if tomorrow there’s a model that drops that’s significantly better than what we have, then we’ll switch to it. But it’s a non-trivial amount of human work to make sure that the note quality is really good for that model.

TB: What does that mean in practice?

CP: We run these [evaluations] and we have these dashboards that show that, say, this model does way better on these meetings that are of this length, on this topic, with this many participants but [it] take[s] a short meeting and write[s] really long notes. So then we have to go through and try to counter that inherent behaviour of the model through prompting and other means. So there’s always this, like, fixing of the model. You actually have to have pretty sophisticated telemetry to test that.

We build a lot of internal tooling to measure the model but at the end of the day, the feel of it really matters. So you have a human looking at lots of examples and tweaking the knobs.

TB: What else have you learnt about how these AI models work by being down in the weeds of building with them, that your average armchair expert on X doesn’t understand?

CP: I think there are a bunch of tasks right now where the models don’t really need to get more intelligent. And then there are tasks where it makes a huge difference.

Long-running agentic tasks right now are a place where you really feel the difference between the latest model and the model that came out a few months ago. But for note generation, for example, newer models don’t actually make that big of a difference. I suspect there are lots of use cases where the intelligence is probably good enough at this point.

TB: When you look at the billions of dollars that people are ploughing into infrastructure to train the next models and run them, does the fact that you feel like we’ve topped out in some of these areas make you nervous about the state of the industry?

CP: I think that we have really intelligent models that are widely available — and open source isn’t that far behind — so we’re all going to have access to really cheap, good intelligence regardless of what happens in the macro environment. So I think that’s great for anyone who’s building and for folks like us.

I do think there’s a crazy high-stakes game of chicken that’s happening at the frontier. If you’re saying, “artificial general intelligence — an AI that can do any human task — is around the corner and AGI is priceless”, then you can justify any kind of investment. I don’t know how that’s going to play out.

TB: Can the best outcome of that ever be as amazing as some people in Silicon Valley think? Is there a real chance that we do get to something really transformative?

CP: Oh, I think so, absolutely. I think we are living through a transformative moment in history. What I can’t judge is just how many zeroes you need to justify that investment.

TB: A lot of AI assistants that pop up in online meetings have been criticised by people who don’t want to be recorded or feel they are intrusive. Is this a product problem or an etiquette question?

CP: Ultimately, we are a tool used by humans. It’s on those people to know when to use this, what’s the right way to use us, and they’re legally responsible for the decisions they make. I think that’s true with any tool and we don’t infantilise our users.

But I think from our perspective, the design decisions we [at Granola] make matter. I would like Granola to be in this world where whenever [someone is] using it, the people around you know about it. We have a gap right now there and it’s due to [restrictions on access to meeting apps imposed by] the platforms.

We wanted Granola to feel like a notepad — you pick it up, it always works, it doesn’t matter what room you’re in or if you’re on a video call or not. And the only way to build that, technically, that we could find was basically using the system audio on your computer as opposed to a bot [that joins and listens to an online meeting as if it was a human participant]. There’s no good way, in our opinion, for Google Meet or Zoom or those platforms to enable that. I think those platforms are going to evolve and this will just be very different in a year or two.

TB: As we have also seen with some of the reaction to AI smart glasses, there seems to be a backlash brewing against technology that might be construed as surveilling people without their consent. To me, it all seems to be caught up in the broader opposition to AI from some people.

CP: To any new technology, there’s going to be resistance and backlash, and I think AI’s going to be very disruptive in ways that we don’t fully understand. So I fully expect there to be lots of waves over the next few years of resistance or backlash to AI, some of which I feel, even though I’m working in this space.

But I think there’s a very big difference between AI glasses and meeting notes. If you’re in a work context, the meeting notes feel genuinely useful. If you show up at a party wearing AI video glasses, I get nothing from that but it makes everything weird. Unless I’m your friend and I’m going to get these great photos from you afterwards, which maybe sounds great, it’s mostly a net negative for me. Whereas in a work environment, I don’t think people would be using Granola if it wasn’t actively useful to them and their colleagues. That’s very specific to work. I think if you start transcribing parties, that’s a completely different thing.

I think there’s a very big difference between AI glasses and meeting notes. If you show up at a party wearing AI video glasses, I get nothing from that but it makes everything weird

I think there’s a very big difference between AI glasses and meeting notes. If you show up at a party wearing AI video glasses, I get nothing from that but it makes everything weird

When it comes to transcription [at work], the important question is who has access to it. If you have notes or a transcript of a meeting that help you [as an individual user] be better and you choose how to use that, that is very different than if every meeting is transcribed and is visible to your boss and their boss and their boss. Or if, say, an [AI] agent can use that data to assess your performance or something. There’s a world where AI feels very empowering to the individual and there’s a world where AI feels empowering to the authority, whether that’s a government or a company.

Granola’s position is that the end user is in control. If you use Granola in a meeting, it is private by default. It is up to you to share it. Ultimately, the [individual user] should have control over where that information goes and how it gets used, not the company. Companies pay, so companies are our customers. But we think, ultimately, to serve the company’s interest, you also have to build a tool that’s good for the employees and they use their judgment on how that context should be used.

TB: I suppose this privacy debate has blown up around AI meeting bots but actually applies to all sorts of ways that AI agents could be used in the workplace.

CP: If you want to use AI for work, it’s going to have to have access to all kinds of context. Once you’ve captured that context — who that gets shared with, who has access to it, is it the full context, is it redacted, is it a different artefact — that is one of the hard questions that we’re going to have to resolve for AI to be really useful in the workplace over the next few years. And it’s incredibly messy and hard and we don’t have good solutions to it. This is one of the areas where the technology has leapt ahead . . . and I don’t see that many people thinking about it yet. We’re going to have to figure out new norms, new systems by which information gets shared.

This transcript has been edited for brevity and clarity