Summary

Granola is using meeting notes as the wedge into a context-rich AI workspace where communication-heavy knowledge workers could eventually do most of their work. Today it combines a user’s sparse judgments with a live transcript; Pedregal’s broader vision is real-time dossiers for people entering meetings, while Granola should eventually take follow-up emails, investment memos, and scheduling “eighty, ninety, ninety-five percent of the way.” He considers the product only “five percent down the path to our vision.”
The defensible app-layer opportunity sits where a task is both high-frequency and must be done exceptionally well; low-frequency, “pretty good” work will collapse into general assistants. Granola freely swaps among competing foundation models, so its stated differentiation is not owning a model but offering a workflow-specific interface. Pedregal’s answer to “Won’t Anthropic build this?” is that a dedicated power tool can remain better because “the UI [is] optimized for this use case.”
Granola’s current advantage is accumulated personal context plus relentless product velocity, not a permanent lock on foundation technology. More meeting history raises switching costs, but Pedregal calls those costs and other moats small: “You get complacent for three months, you’re in trouble.” The competitor he worries about is the startup that has not launched yet and can begin from everything incumbents have already learned.
Its clearest product lesson came from killing the most magical demo: real-time AI note completion made users less present, directly violating the product’s purpose. Granola spent six months trying to perfect that interaction before moving the AI work to the end of each meeting. The resulting “magic moment” arrives later than the ideal first 20 seconds, but users retain control through an ordinary editable notepad.
Pedregal expects AI memory assistance to become so valuable that working without it might feel like an impediment within “maybe 18 months,” but adoption depends on a visible social contract. Granola transcribes without storing audio, sacrificing replay and tone to be less invasive than meeting bots that retain audio and video. He favors placing a phone openly on the table and rejects the vision of a hidden pendant listening to everything.
Foundation models can radically compress the labor required to build a large software business, though they do not eliminate product taste or messy edge cases. Granola’s CTO actively minimizes engineer-written code, while Pedregal expects customer-experience departments created after 2025 to be much smaller and structurally different. He agrees a $10 billion company with roughly 20 employees is conceivable, even if Granola’s ambition will still require hiring.
The long-run design boundary is to outsource rote work without outsourcing human judgment, because “writing is thinking” and generated ideas can silently narrow what people consider. Pedregal wants AI to surface the best context while humans interpret, decide, and create; surprisingly, today’s models still give different people nearly identical answers. For investors, that makes the application layer an “explore mode” market where specific product insight and a shared worldview matter more than confident execution forecasts.

Deep dive

1. AI’s next leap is dynamically generated context

Pedregal begins with “humans are toolmakers”: writing and notepads externalize memory, effectively “extending the RAM” available beyond the brain’s physical limits. Mathematical notation similarly changes what humans can calculate; Roman numerals constrain mental arithmetic, while modern notation makes long division with enormous numbers manageable.
His favorite example is a figure called Playfair—he thinks the first name was William—who roughly 200 years ago mapped numerical data onto a visual plane. Because humans rapidly interpret images, a graph lets someone immediately feel that a series is rising, falling, or accelerating—an intellectual capability that seems obvious only after the tool exists.
LLMs add something qualitatively different: they can “bring extremely relevant context to the person in the moment they need it,” rewritten dynamically for the exact situation. Pedregal can imagine only the next few steps; in 10 or 20 years, he guarantees these tools will look “nothing like” today’s products.

2. Meeting notes are the wedge into a workspace for knowledge work

Granola currently behaves like a digital notepad that also listens and transcribes. Users write only the insights they consider important; after the meeting, AI expands those fragments using the transcript, outsourcing rote capture work while leaving the human’s attention and judgment in the conversation.
Heavy users consequently write just a few subjective observations: “This person was a bit aggressive,” “They seem kinda down,” or concern that someone evaded a question. When they return, they increasingly ask a specific question through chat instead of rereading pages of notes.
Pedregal’s blog-writing workflow illustrates the broader behavior change. He recorded advice from several people, walked around speaking a separate brainstorm aloud, placed everything into a Granola folder, and asked AI for themes and possible structures. He still writes the post, but fewer useful ideas disappear during synthesis.
The larger vision is a real-time diplomatic “dossier” assembled before any consequential meeting. An internal, unreleased feature already finds themes across meetings with one person or on one topic; eventually, Granola should use that context to draft emails, memos, and other post-meeting work “eighty, ninety, ninety-five percent” of the way.

3. AI memory becomes normal through an explicit social contract

Pedregal believes that in “maybe 18 months,” working without a Granola-like tool might feel needlessly limiting. He is less certain where society will place the boundary between usefulness and surveillance: the goal is “maximum usefulness” with the “minimum amount of invasiveness,” but he offers no confident endpoint.
Granola deliberately transcribes in real time without recording or storing audio. That eliminates exact replay and vocal tone, creating a real loss of user value, but Pedregal considers it meaningfully less intrusive than bots that enter calls and retain both audio and video for an uncertain period.
An iOS app is “launching soon”; users with a third of their meetings in person report feeling “flying blind” or “naked.” Pedregal favors an openly placed phone that makes note-taking visible to everyone; he hates hidden pendants and expects hidden-pendant recording to provoke a Google Glass-style backlash even as workplace capture normalizes.

4. AI-native companies can need fewer people—and a new operating intuition

Granola inherits enormous capability from foundation models, allowing a small team to concentrate on end-to-end experience. Yet quality still depends on obscure technical cases: removing AirPods during a multichannel Zoom call can require precise behavior that nobody anticipates until the broken version feels “crappy.”
Pedregal expects customer-experience departments created after 2025 to look fundamentally different. Granola’s first customer-experience hire is part of a broader shift toward using AI across the function, with fewer people doing different work; retrofitting an established department will be harder than creating one around the new paradigm. A $10 billion company with 20 employees is plausible, though not a forecast for Granola.
CTO Vas’s explicit goal is to minimize the code each engineer writes. At an off-site, he also corrected Pedregal’s weak barbecue prompt: photos of the grill and Spanish-labeled shrimp supplied the missing context, revealing that the shrimp were already cooked. The lesson was “give it the context”—AI-native workers will instinctively supply more evidence rather than assume they already understand the situation.
Collecting emails, notes, documents, and tweets will become easy; selecting what matters now is harder. Pedregal compares chat interfaces to early cars controlled by a crude stick: adequate at low speed, dangerous when moving fast. AI still needs its “steering wheel”—a shared canvas with granular control and fluid collaboration instead of alternating commands and responses.

5. Model competition shifts differentiation toward focused workflows

From an application builder’s perspective, competition among model providers is “the best thing ever.” Granola uses many models in different places and will switch to “whatever the best model is on any given day.” The hot swap is conceptually simple, although building trustworthy evaluations is not.
Pedregal maps applications on two axes: frequency of use and required output quality. General assistants should absorb infrequent tasks where “pretty good” suffices, especially consumer use cases too rare to create a standalone habit. High-frequency, high-quality work remains the “power tool quadrant,” protected primarily by purpose-built product design.
Education reinforces his forecast. GPT-4 voice mode playing hide-and-seek with his five- and seven-year-olds revealed a striking new child-computer interaction, yet he expects generic assistants to capture much of tutoring. Because one-to-one tutoring can move a median student toward top-5% or top-10% performance, he thinks it “should be free” and built on an open-source model rather than distorted by business incentives.
Proprietary data is less categorical as a moat than under older machine learning. A specialized system can sometimes get away with 50,000 examples rather than “millions and millions and millions.” As app creation becomes accessible like phone photography, Pedregal suspects taste may command a greater premium precisely because everyone can produce something.

6. Great product design begins with emotion, then survives painful reversals

Pedregal’s core design test is simply, “How does this make me feel?” Within roughly 500 milliseconds, a user may register clutter, uncertainty, complexity, or insecurity. An imaginary “emotional recorder” played in slow motion would expose much of what must change. Granola’s corresponding principle is user control, including editable notes rather than a static PDF-like output.
Choosing a Mac app initially restricted Granola to macOS 13.4, then about 15% of Mac users, and created substantial technical pain. The choice nevertheless made the product behave like “a notebook and a pencil”—consistent across Zoom, Slack huddles, and in-person contexts, immediately accessible rather than buried among 50 browser tabs.
Granola’s original demo let users type a keyword, press Tab, and watch AI complete the note in real time. After six months, the team accepted that even excellent generated text compelled people to read and edit instead of listening. Moving the magic to the meeting’s end weakened instant gratification but made the underlying product substantially better.
Usage escaped the original work-meeting brief: one user found Granola invaluable during meetings with doctors about a partner’s cancer, individuals talk aloud to brainstorm or prioritize a day, and learners take notes while watching YouTube videos. Its AI-forward user base—founders, investors, and employees across AI startups—was unusually vocal after the May launch, with famous CEOs turning Pedregal’s Twitter DMs into customer support.

7. Product velocity depends on distinguishing exploration from execution

Granola explicitly distinguishes “explore mode” from “exploit mode.” When the answer is known, teams should build the minimum, impose a deadline, ship to real people, and shorten iteration cycles. Applying that playbook to an unsolved problem merely rewards shipping something bad without actually discovering a good solution.
The company worked for a year before launch despite being, in Pedregal’s phrase, already “seven years late” to AI note-taking. That private period preserved its ability to replace the core interaction. A public release would have trained retained users around real-time completion and made the necessary pivot nearly impossible.
His competitive answer is blunt: “You need to build something better than other people faster.” Accumulated context creates some switching cost, but it is not permission to slow down. Granola must deliver the next useful meeting features while taking the much larger swing from note-taking into the workspace where users write documents and do most of their work.
The competitor Granola chooses to fear is “the one that hasn’t launched yet”—a startup able to build from the lessons others have learned and execute from that head start. Pedregal was impressed by big tech’s rapid AI pivots, but deciding to respond is not the same as executing; startups often become big tech’s R&D wing, while generational companies can leverage an earlier discovery into something massive.

8. Human judgment is the design boundary—and exploration is the investment test

Pedregal wants “tools that make us more human and better humans.” AI should remove boring and mindless work while preserving judgment, creativity, and the thinking embedded in writing. His personal motive echoes a phrase from an early boss: “the active realization of human potential,” both his own and humanity’s.
His dream tool would dissolve information silos, dynamically surfacing the best material from a person’s life and the wider world. A friend’s prototype rendered slightly divergent Midjourney-like imagery at roughly five or eight frames per second as people spoke, hinting at richer thinking aids—but also at the difficulty of making live augmentation helpful rather than mesmerizingly distracting.
Today’s persistent surprise is the lack of personalization: Patrick and Chris can ask the same model a question and receive nearly identical answers. Granola deliberately produces different notes for different participants because what matters to each person differs; Pedregal sees that individualized relevance as a basic, still-underdeveloped capability.
His advice to AI investors follows the same operating distinction: foundation models may be entering exploit mode, but applications remain “total explore.” Cold outreach stands out when it contains a specific observation about product behavior, including something Granola may have wrong. Pedregal wants long-term partners who share a worldview and can reason deeply about products because every execution detail will change.

The rise of AI agents with João Moura of CrewAI 102: DeepSeek Launches Open Source Week—What Exactly Is Being Open-Sourced in Large Models?