Aug 21, 2025 · 1h 8m · mad
How to Build a Beloved AI Product - Granola CEO Chris Pedregal
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
In this episode of The MAD Podcast, host Matt Turck speaks with Granola Co-founder and CEO Chris Pedregal about building a breakout, opinionated AI workspace for knowledge workers. Chris discusses human cognitive augmentation, product simplicity, AI stack architecture, launch strategies, and the vision of transforming meeting notes into deeply contextual team intelligence.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. Matt holds 28.6% of the talking time here. How this is scored →
Matt as informed peer 3.7 Guest teaching 4.7 Guest disagreement 0.7 Matt pushing back 1.3
speaking balance: gold is Matt, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 1:00:46 Rejection of OpenAI commoditization threat
The guest directly pushes back against the host's premise that OpenAI will subsume personal memory tools, arguing that horizontal platforms try to be everything to everyone while specialized power tools win on tailored user experience.
Hardest push from Matt ▶ 1:00:46 Host challenges Granola's moat against OpenAIThe host refuses to accept the guest's easy dismissal of competition, explicitly pressing that building long-term user memory is the ultimate strategic prize for foundation model giants like Google and OpenAI.
Biggest teaching moment ▶ 43:10 Why RAG fails for qualitative meeting queriesThe guest educates the host on technical limitations of vector search and cosine similarity RAG, demonstrating that qualitative queries require cramming full context into model context windows despite higher inference costs.
Matt holds his own ▶ 45:03 Probing AI unit economics and gross marginsThe host demonstrates sharp financial and technical understanding by citing Cursor's well-reported negative gross margins, pressing the guest directly on Granola's underlying inference and transcription unit economics.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | Matt as informed peer | Guest teaching | Guest disagreement | Matt pushing back | Why |
|---|---|---|---|---|---|---|
| Granola's Early Viral Adoption in Silicon Valley | 2 | 3 | 0 | 0 | The host opens with high praise for Granola, sharing his firm's rapid adoption and citing recent media coverage. The guest responds appreciatively, sharing how an SF user event demonstrated organic community enthusiasm beyond their initial product design. | |
| The 'Second Brain' Concept and Human Cognitive Augmentation | 3 | 6 | 1 | 1 | The host brings up the philosophical implications of outsourcing memory and navigation cognitive atrophy. The guest expands on this with references to Douglas Engelbart, Google Maps, and sci-fi tropes like WALL-E vs. Iron Man's Jarvis. | |
| Entering a Saturated Category and Finding Granola's Entry Point | 4 | 5 | 1 | 2 | The host asks how Granola broke into an already crowded market of AI note-takers. The guest reframes Granola not as a meeting recorder, but as a personal tool for thought using meetings as a high-context entry point. | |
| Product-Focused Founder DNA and LLM Engineering Needs | 3 | 5 | 1 | 1 | The host asks about founder DNA and when technical LLM specialization is required. The guest explains why product and design taste mattered first and why hiring specialized LLM researchers was deferred until after product-market fit. | |
| Building out of London with Silicon Valley DNA | 3 | 4 | 1 | 2 | The host notes the narrative that great AI startups must be built in San Francisco. The guest outlines the benefits of building in London, including access to deep talent and insulation from Silicon Valley hype thrash. | |
| Stealth Mode, Closed Feedback Loops, and Launch Strategy | 3 | 5 | 1 | 1 | The host asks about the decision to spend a year in stealth versus public building. The guest articulates the strategic rationale that closed feedback loops learn faster when fixing obvious flaws before taking on public maintenance burdens. | |
| Initial User Acquisition Strategy: From VCs to Founders | 4 | 5 | 1 | 2 | The host asks about user acquisition wedge strategies and Granola's opinionated choice to forgo a visible meeting bot. The guest explains how bots create discomfort and why Granola chose a personal, non-intrusive architecture without saving audio. | |
| Achieving Simplicity by Radically Cutting Features | 3 | 5 | 0 | 0 | The host asks how Granola achieved its simple product feel. The guest reveals that they radically cut 50% of the app's feature surface right before launching, a painful decision made easier by being pre-launch. | |
| Balancing Product Intuition with Quantitative Data and User Calls | 3 | 4 | 0 | 0 | The host asks how product intuition is balanced against qualitative and quantitative feedback. The guest explains that intuition leads product direction, while daily user calls maintain grounding and context. | |
| Navigating Enterprise Demands vs. Building for Tomorrow | 4 | 5 | 1 | 2 | The host asks how Granola balances enterprise requests against product simplicity. The guest emphasizes that the biggest failure mode is optimizing for current user feature requests rather than building for tomorrow's capabilities. | |
| AI Model Orchestration, Evaluation, and Prompt Engineering | 4 | 4 | 0 | 1 | The host asks technical questions regarding multi-model routing and maintaining consistent UX across non-deterministic models. The guest explains prompt engineering adjustments per base model release and defaulting model picks by task. | |
| Context Curation, Personalized Prompts, and Beyond Standard RAG | 5 | 6 | 1 | 2 | The host brings up context window limitations in long meetings like board meetings. The guest educates on why standard RAG and vector cosine similarity fail for nuanced qualitative queries, forcing them to feed full context into context windows. | |
| Unit Economics and Gross Margins of AI Tools | 6 | 5 | 1 | 3 | The host directly probes financial health and unit economics, citing coding assistants like Cursor operating at negative gross margins. The guest clarifies that real-time transcription is their primary cost center rather than LLM inference. | |
| Product-Led Growth, Bot Trade-Offs, and Meeting Privacy | 3 | 4 | 0 | 1 | The host inquires about growth loops without meeting bots. The guest shares organic referral dynamics and tells an anecdote about a confidential meeting almost leaked by an automated Google Meet bot. | |
| Enterprise Adoption, Compliance, and Legal Liability | 4 | 5 | 1 | 2 | The host asks about legal liability risks of creating permanent conversational records in enterprise settings. The guest outlines the structural tension between corporate retention policies and AI context requirements. | |
| User Retention Metrics and Habit-Building Triggers | 5 | 5 | 2 | 4 | The host pushes on why foundation model providers like OpenAI won't commoditize Granola by building personal memory directly. The guest resists the threat framing, contending that general models won't beat specialized native power tools. | |
| Product Roadmap and Deep Context Search | 3 | 5 | 0 | 0 | The host prompts for roadmap insights. The guest details upcoming features including deep research capabilities that search across thousands of meetings in seconds to produce dynamic memos. | |
| Vision for AI Executive Coaching | 4 | 3 | 0 | 0 | The host suggests Granola evolving into an AI executive coach based on meeting history and praises Granola's multi-language translation feature. The guest expresses excitement and concludes the episode. |
Statements from this episode (40)
Granola's SF user meetup packed a two-floor bar despite expecting five
“I we did this was last maybe November so we're based in London, right, and we went to SF for a board meeting, and someone on the team said, hey, should we rent out a bar and just email users and say if you'd want to come? And we'd like, sure, you know, and we …”
Pedregal: The computer mouse is Douglas Engelbart's least important invention
“Like he's known for being the inventor of the mouse. And I think the mouse is like the least important thing he's come up with.”
Pedregal: AI's long-term effect on human agency will be decided within 10 years
“And I think it's a little bit What tools we build, what bets we make as a society, what rules we make, I think that's going to be decided over the next, like, 10 years.”
Pedregal: Granola's biggest competitor has always been Apple Notes
“The biggest competitor that Granola had from day one and even today is Apple notes.”
Pedregal: Applied AI startups don't need deep technical chops for MVPs
“It used to be that you need really strong technical chops just to build an MVP to understand if you, this is something people wanted or not. And I think the reality now is that that isn't the case. You can usually figure out MVP or like, is there may be even e…”
Granola paused hiring deep LLM experts until reaching product-market fit
“And in, in those early phases, like when I was looking for a co-founder, I met Sam, but I was also, I met all the LLM experts from Imperial and Oxford and Cambridge. Because I thought that was DNA we would need on day one. And as Sam and I started prototyping,…”
Pedregal: London is seeing a huge influx of Russian tech talent
“And there's a huge influx of Russian tech talent that's coming to London.”
Granola enforces American spelling to present itself as a US company
“If I ever see any copy that has English spelling instead of American spelling that goes out, I throw a hissy fit because I want everyone to think we're an American company.”
Pedregal: Granola spent a year privately onboarding users before public launch
“So we basically spent a year onboarding people, learning what was wrong about it, making fixes to that onboarding a new set of people.”
Pedregal: Startups need polished products to stand out in crowded markets
“I think today there are so many products and companies coming out and vying for your attention that Launching something more polished so that when people use it, they're wowed by it is, is a way to stand out.”
Pedregal: Granola will eventually be a horizontal product used by many user types
“We know granola will be a general product, a horizontal product, lots of different types of people are going to use granola”
Pedregal: Building a great product for founders creates a decent product for other roles
“A great product for founders would be by default, a decent product for folks in these other roles.”
Pedregal: Everyone will use AI meeting note tools within 2-3 years
“I'm sure that two years from now, three years from now, everyone's going to be using something like granola. Like, I'm hoping it's granola, but if it's not granola, something like granola, just because it is so useful and will get so much more useful over time…”
Granola does not record or store raw audio from meetings
“Even though we could store the audio and that would be useful, we do not store the audio. So we don't record the audio”
Pedregal: AI tools must let users verify LLM outputs against source data
“In the world of AI, where AI makes mistakes, transcription makes mistakes, it's really important that I don't have to trust, I don't have to trust the LLM output. I can kind of go back to the source.”
Granola cut 50 percent of its product features prior to launch
“We looked at it all and we cut out 50% of it. We basically redesigned and cut out 50%.”
Pedregal: Cutting Granola features post-launch would have been nearly impossible
“I think that would have been impossible or extremely hard to do if we had been publicly launched.”
Pedregal: User requests pose the greatest danger to product simplicity
“The real danger there is user requests. Like, you know, people always ask for the things they don't have. People rarely say, oh, actually, can you cut out half of the functionality of the app?”
Pedregal: Granola makes most product decisions based on intuition
“So our general our philosophy that's gotten us here, and it may not get us there as we scale, is we make most product and design intuition, sorry, decisions based on Intuition.”
Pedregal: Granola co-founders aim to conduct four to six user calls weekly
“And we aim to do, I think Sam and I aim to do four to six calls a week with users, but constantly not like, oh, we're doing a sprint on this feature. It's actually, we try to book them Every day, always.”
Pedregal: AI startups fail when optimizing for current user requirements
“I think the failure mode is that we optimize for today's world. So we optimize for today's product and today's world and today's needs. And it's easy when you just talk to users and you get these requirements or you talk to enterprises. You just kind of, it's …”
Pedregal: Granola is a Trojan horse to aggregate user context
“And in a way the product we have today is, is a Trojan horse to collect a lot of your context so that you can then use all the information in that context to do future work.”
Pedregal predicts cross-meeting AI research will drive massive value for Granola
“We're doing a kind of a, this really incredible, like deep research mode across that can look at thousands of meetings in a matter of seconds and pull out these insights and, You know, that's not something that our enterprise customers are asking for right now…”
Pedregal: Granola will only fine-tune or train models after hitting base-model limits
“Our strategy has been to use the latest and greatest. And when we feel like we hit a wall and the only way to make the experience better is then to fine tune or train models. And we will do that.”
Pedregal: Granola abstracts model choice for note generation, allowing choice only for chat
“And we only let users choose their model on chat. On the note generation side, we completely abstract it away.”
Pedregal: Every new base model requires Granola to overhaul its prompts
“Every time a new model comes out, we have to completely change or tweak the prompts that we use for note generation to provide consistency of experience and an improvement of experience.”
Pedregal: Complex meeting queries fail with standard RAG without full context
“What we found is that a lot of the most interesting queries that people have like would completely fail with that type of method. So for example, query might be like, what are all the things I didn't do a good job explaining. Or tell me what are all the bugs t…”
Pedregal: Granola builds for AI capabilities expected a year out
“And our philosophy since the beginning of Granola is always to build for the world a year from now, because by the time we build it and it gets distribution the costs of those models or those capabilities will come down to a reasonable place.”
Pedregal: Audio transcription is Granola's largest cost, exceeding LLM inference
“So the most expensive thing about our business is actually transcription. And historically it's been actually transcription and high quality transcription versus LLM inference.”
Pedregal: Granola does not operate with negative gross margins
“We're not at a negative gross margins right now”
Pedregal: Granola's inference costs will stay flat or rise as queries expand
“What I do expect is I expect the cost of inference to stay the same or go up as we allow users to do much more complicated queries over much larger data sets.”
Pedregal: Real-time speaker diarization quality is still in its infancy
“Real-time diarization is still in its infancy in terms of quality.”
Pedregal: Incorrect diarization confuses AI models more than letting them infer
“If you give like incorrect diarization, To a model, it'll oftentimes confuse it more than if it just has to try to infer who's speaking.”
Granola's next goal is becoming a second brain for teams
“If granola acts as a second brain for you, like our goal, the next step there is to be a kind of a second brain for your team or for your company.”
Pedregal: Granola's retention relies on utility and timely calendar triggers
“It's the, I think the combination that granola is useful to folks and we can send notifications at the right moment to start it.”
Pedregal: OpenAI will try to do everything for everyone
“OpenAI is going to try to do everything to everyone.”
Pedregal: The future of software is generating artifacts on the fly
“I think that the world we're moving towards is You have a bucket of context, and then you generate documents or artifacts on, on a per need basis on the fly.”
Granola built an internal prototype searching 2,500 meetings in 20 seconds
“I have a version of granola that will go through my 2500 meetings and spit out a remarkably intelligent answer to that in 20 seconds.”