Products with a Soul: What Mac AI Apps Can Learn from Granola

Most AI apps today don’t feel broken. They feel assembled.

A prompt box here. A “magic” button there. Whatever the latest model can do, bolted on as quickly as possible. The result often works — but it doesn’t feel like anything. There’s no consistent voice, no sense of intention, no personality you can learn to trust.

Following our Granola review and after listening to founder Christopher Pedregal talk about the company on Matt Turck’s Mad Podcast and in his own PMF talk, one idea kept resurfacing: great products don’t feel like feature sets. They feel like people you’ve come to know.

When software can’t decide who it is

Granola’s founders often describe a good product as behaving like a familiar person. You can anticipate how it will react, how it will look, and what it values. When that expectation breaks — when a product feels thoughtful in one moment and oddly aggressive in the next — trust collapses.

On the Mac, this kind of inconsistency is especially visible.

You can often see the org chart in the UI:

• One screen optimized for growth
• Another clearly owned by infrastructure
• A settings pane that feels like legal snuck it in at the last minute

Nothing is obviously wrong, but nothing feels aligned either.

A product with a soul does the opposite. Every surface reflects a small set of deeply held values: how much it respects your time, how it treats your attention, and how much control it gives you.

Granola is a rare AI app that passes this test.

Why Mac users notice “soul” more than anyone else

Mac users don’t evaluate software one feature at a time. They judge it as a whole.

Keyboard shortcuts, menu bar behavior, typography, latency, defaults — these details quietly add up to a feeling of coherence or friction. That’s why so many AI tools feel slightly off on macOS. They function, but they don’t belong.

Granola feels native not because it looks like Apple Notes, but because it behaves like a good Mac app: calm, opinionated, and respectful of the user’s flow. It doesn’t constantly remind you that it’s powered by AI. It assumes you’ll notice.

It’s worth noticing how unusual that restraint actually is. Open the average productivity app and you’re met with banners, tips, streaks, and upsell prompts competing for a slice of your attention — a design logic that makes sense if your business model depends on maximizing time-in-app.

Granola’s interface is closer to the opposite bet: one notepad, one recording indicator, nothing else fighting for your eyes. That’s not a lack of ambition. It’s a decision about what kind of company Granola wants to be — one that makes money by being genuinely useful in a moment, not by holding your attention hostage.

Annotation, not transcription: the real product insight

It’s tempting to describe Granola as “a notepad with AI.” That undersells what’s actually going on.

Most AI tools today fall into one of two failure modes. In the first, AI sits upstream: it reads, watches, or listens to something on your behalf and hands you a finished digest — an AI podcast summary, an AI-condensed article. The problem is that it doesn’t know what you actually care about, so it defaults to the statistical average of what mattered, and averages are rarely what you were looking for.

In the second failure mode, the human sits upstream instead: you write the prompt, and the AI executes — image generation, slide decks, most “agents.” The problem here is the opposite one — writing a genuinely good prompt is its own skill, and most people never quite get there, so the output is serviceable at best.

Either way, human and AI are working in sequence, like a factory line — never actually together.

Granola’s answer is neither. It’s a parallel model: you and the AI work on the same meeting at the same time, but on different jobs. You handle judgment — what’s worth flagging, what made you pause, what you’re unsure about. The AI handles execution — capturing everything that was said, then using your own annotations as the map for what to expand on.

Write “pricing concerns,” and Granola goes back through the transcript and pulls every relevant moment tied to pricing. Write nothing, and you get a generic pass. Write focused notes, and you get a focused, useful note.

This is the literal meaning of “human in the loop,” and it’s a sharper idea than it first sounds: your pauses are the anchor, and AI is only the flesh on top of them. Full manual note-taking loses content. Full AI transcription loses meaning — a three-hour interview turned into twenty thousand words of transcript is, in practice, a document nobody reopens. Signal buried in noise is functionally the same as no signal at all.

That distinction — annotation over transcription, judgment over completeness — is also why several early users and reviewers have pointed to unusually strong retention numbers for a note-taking tool. We’d treat any specific figure as anecdotal rather than verified, but directionally it tracks: a tool that respects what you already decided is important tends to get reopened, in a category where most tools get tried once and forgotten.

Where the philosophy actually came from

Pedregal has said the idea for Granola began the moment he encountered GPT-3 after leaving Google — his prior company, an AI tutoring app, had convinced him AI’s real value was in usefulness, not novelty. His co-founder came out of the tools-for-thought and knowledge-management world, and together they landed on a shared instinct: AI will let people think and work differently, so someone needs to build the tool that actually supports that shift.

Both founders trace that instinct back further, to Douglas Engelbart’s 1950s–60s vision of “augmenting human intellect” — the idea, radical at the time, that computers should exist to extend human judgment rather than replace it. Engelbart is best remembered for inventing the mouse, which both founders consider the least interesting thing he did.

The more important legacy is the framing question it left behind: for any new capability, where does it replace a human, and where does it augment one? Granola is, in effect, a small, specific answer to that question applied to meetings.

Coherence is a design choice, not a byproduct

What stood out in Pedregal’s own account of building Granola wasn’t speed or a clever growth hack — it was how much effort went into subtraction. The team spent roughly a year in private beta before launching, onboarding real users daily and fixing what was broken rather than shipping publicly. By the end of that year, the product had accumulated swipeable panels, a raw transcript view, a “blow-by-blow” log, translated notes — every feature that had ever solved somebody’s specific pain point.

Then, deliberately, they cut about half of it.

Pedregal has called that decision one of the hardest and proudest of the company’s history, precisely because it’s organizationally unnatural. Every individual feature, considered on its own, looks worth keeping — there’s always a real user who asked for it. Coherence only shows up when you zoom out and look at the whole product at once, and that’s a job almost nobody in an organization is naturally incentivized to do. It’s also a decision that gets exponentially harder the more users you already have; cutting half your feature set is a very different cost pre-launch than it is with an audience who’s grown attached to the clutter.

Many Mac AI apps struggle here. In trying to be helpful, they become noisy. In trying to be powerful, they lose their voice. The result is software that can do many things but feels like no one in particular.

A tool, not a bot — restraint as trust

Granola’s decision not to join meetings as a visible bot is usually framed as a privacy choice. It’s really a consistency choice first. Pedregal’s own reasoning: a tool has to work the same way every time, in every context — Zoom, Meet, Teams, WhatsApp, or a coffee shop with no video call at all — because you only get a narrow window to reach for whatever’s already open when you decide something’s worth writing down. In that window, Granola’s real competitor was never Otter or Fireflies; it was Apple Notes, because Apple Notes always works, everywhere, instantly.

The same instinct shaped a smaller but telling decision: Granola doesn’t store meeting audio, only the transcript. Not because it couldn’t, but because a product you’re supposed to trust shouldn’t ask you to just trust it. If the AI gets something wrong — and transcription and models both still make mistakes — you should be able to go back to the source yourself rather than take the summary on faith. That’s a specific, almost old-fashioned kind of humility for an AI product to build in on purpose.

The founder as the product’s nervous system

Pedregal has described aiming for four to six user calls a week, every week, not tied to any particular feature sprint — a standing habit, not a project. His stated reason is blunt: the moment you stop talking to real users regularly, it becomes dangerously easy to convince yourself that people want whatever you’re already planning to build.

It’s the practical version of the old line, often attributed to Jeff Bezos, that customers don’t actually want a quarter-inch drill bit — they want a quarter-inch hole. Granola’s team seems to treat that literally: the product isn’t optimized around what’s technically impressive to build, but around what a person in a meeting genuinely needs in the next ninety seconds.

Early users are treated less like a metrics funnel and more like unpaid product partners — which, not coincidentally, is exactly the kind of relationship that turns a tool into something people recommend by name rather than something they merely use.

Soul as the last real differentiator

As AI models converge, capability differences matter less. Everyone will have access to similar intelligence. What won’t be shared is judgment.

That’s also Granola‘s answer to the obvious question every AI startup eventually gets asked: why wouldn’t OpenAI, or Google, just build this? Pedregal’s own framing is that the giants are going to try to do everything for everyone, and do it well — which leaves a real, if narrower, opening for something built way better for one specific use case and one specific type of user. That’s a bet, not a guarantee.

But it’s a coherent one, and it’s the same bet that’s let a notepad grow into something the company itself now describes as a “second brain” — first for individuals, increasingly for whole teams, with meeting notes acting as a kind of on-ramp for a much larger layer of shared context.

Products with a soul feel predictable without being boring. They have opinions. They respect boundaries. They make users feel smarter rather than managed.

Granola won’t be the last AI app to take this approach — and its own team seems to think the “human in the loop” idea generalizes well beyond meetings, to anything where AI currently either does too much on its own or asks too much of you upfront. But right now, it’s one of the clearest examples of what a Mac‑first, AI‑native product with a coherent soul can look like.

In a world full of feature salads, that kind of clarity is still rare — and deeply felt.

Loved the article, share!