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

We open a lot of AI apps at TheSweetBits. Most of them work. Fewer of them feel like they were designed by someone with a strong opinion about how software should behave.

That’s the difference we’ve been trying to put our finger on with Granola.

Following our Granola review and listening to founder Christopher Pedregal talk about the product, we kept coming back to the same idea: great software isn’t defined only by what it can do. It’s defined by the decisions its makers make about what it should do—and what it should leave alone.

That’s what we mean by a product with a soul.

Soul isn’t personality. It’s judgment.

We don’t mean that good software needs a cute mascot, a distinctive color palette, or a founder’s personality baked into every screen.

Granola’s “soul” is much less superficial than that. It’s a consistent set of decisions about what the product is willing to be.

During a meeting, Granola doesn’t compete with you for attention. You write what you’re thinking; it captures what everyone else is saying. Afterward, it uses the conversation to fill in the gaps in your notes.

That sounds like a small interaction choice. It isn’t.

It reflects a larger belief: AI should handle completeness while the human retains judgment.

You decide what made you pause. What seemed important. What needs following up. Granola supplies the memory around those decisions.

That’s a very different philosophy from simply handing an AI a three-hour recording and asking it to tell you what mattered.

The difference between transcription and understanding

We’ve all seen the pitch for AI meeting software: record everything, transcribe everything, summarize everything.

The problem is that completeness isn’t the same thing as usefulness.

A transcript can contain every word from a three-hour conversation and still tell you very little about what you cared about. The signal is there, but it’s buried in the archive.

Granola takes a more interesting approach. Your own notes act as an anchor.

Write three words about pricing and the system can use the surrounding conversation to reconstruct the context. Write down a question, and the resulting note can preserve the discussion that prompted it. Your incomplete notes aren’t a failure to capture the meeting. They’re the map for the AI.

That’s why we think annotation is more interesting than transcription.

The human isn’t being removed from the process. The human is providing the thing the AI doesn’t naturally know: what matters to them.

What a product refuses to do matters too

This is where Granola becomes particularly interesting.

Pedregal has talked about the product’s development in terms of subtraction. After a long private beta, the team had accumulated features such as raw transcripts, additional panels, translated notes and other ways of exposing everything the system knew. Then they cut roughly half of it.

That’s not just minimalism.

It’s a willingness to say no.

And that willingness shows up throughout the product.

Granola doesn’t store meeting audio, even though it could. Pedregal has described that as a deliberate decision because the product should feel like a notepad rather than an archive of recordings. The company also keeps notes private by default and says it refuses requests from companies that want a backdoor into employees’ private notes.

These decisions don’t make Granola more capable.

They make its boundaries clearer.

What a product refuses to do can tell you as much about the product as what it does.

Even the invisible bot is a design philosophy

Granola’s decision not to join meetings as a visible bot is often described as a privacy feature. But the origin of the decision is more interesting.

Pedregal says the original goal was reliability: if you hit start, Granola should work whether you’re in Zoom, a Slack huddle, Teams, or an in-person conversation. Using the computer’s system audio made that possible without depending on a third-party meeting bot. He also describes bots as an inelegant, even dehumanizing presence in a meeting.

That distinction matters.

The product wasn’t designed around a checklist that said “privacy: yes.” It was designed around a broader question: what should a meeting tool feel like when you’re actually using it?

The answer was: like a tool you can simply reach for.

There is a trade-off, of course. An invisible recorder creates a transparency problem, which Granola has tried to address with meeting-chat notices and a visible watermark option. But even that response follows the same philosophy: keep the interaction simple without pretending the social question doesn’t exist.

Good AI doesn’t necessarily mean more automation

This is probably the biggest lesson for AI apps.

There is an understandable temptation to automate everything. If AI can summarize the meeting, why make people take notes? If it can write the follow-up, why ask what they thought? If it can decide what matters, why involve them at all?

But Pedregal’s thinking points in the opposite direction.

Granola’s pre-meeting briefs are proactive. Its Chat can work across your meetings. Its MCP connections can make meeting context available to other AI tools. The product is clearly moving toward doing more with the information it captures.

Yet the team also talks about a limit that many AI products seem to miss: people don’t want another inbox full of things AI thinks they should do. The useful AI, in Pedregal’s description, is proactive without becoming overwhelming, distracting or demanding.

That’s a subtle distinction.

The goal isn’t maximum automation.

It’s the right amount of intervention.

Context is valuable. Ownership matters more.

There’s another reason Granola’s design choices matter.

Meeting software doesn’t just capture words. Over time, it captures a picture of how you work: what you care about, who you talk to, what decisions you’ve made, what problems keep coming back.

Pedregal calls this accumulated conversation “context exhaust.” His argument is that today’s AI systems are intelligent but know very little about the person using them. Granola can provide that missing context.

That creates an enormous product opportunity—and an equally enormous trust problem.

Who owns that context?

Granola’s answer is unusually clear: the notes belong to the individual and are private by default. Even when a company pays for the product, managers don’t automatically get access to employees’ notes. Pedregal explicitly frames the distinction as the difference between AI being your tool and AI becoming the company’s tool for monitoring you.

That is not a minor privacy setting.

It’s product philosophy.

The Mac lesson: coherence beats feature count

This is also why Mac users tend to notice these things.

Mac software has historically been judged not just by its feature list, but by the accumulation of tiny decisions: how fast something responds, whether a shortcut feels natural, whether a preference is where you expect it to be, whether the interface stays out of your way.

AI makes those decisions more important, not less.

When models become commodities, the underlying intelligence starts to converge. The differentiator moves upward into the product: the workflow, the interface, the defaults, the boundaries, the context and the moments when the AI decides to stay quiet.

Granola is interesting because these decisions point in the same direction.

It doesn’t try to make the AI the star of the meeting. It doesn’t insist on storing everything forever. It doesn’t automatically turn private context into company knowledge. It doesn’t try to become another task inbox.

None of those decisions is a moat by itself.

Together, they create something much harder to copy: a coherent point of view about what the product is for.

“My tool” is the real test

Pedregal put the idea particularly well in his conversation with Platformer: humans shape their tools, and eventually their tools shape us. The important question for AI, he argues, is whether it feels like my tool—something I control that makes me better—or the company’s tool, the system’s tool, or somebody else’s tool.

That may be the more useful way to think about “AI product design.”

We spend a lot of time asking whether an AI application is smarter than the competition, whether it has the newest model, or whether it can automate one more task.

We should probably ask a different question too:

After using it for six months, does this still feel like a tool I chose—or does it feel like a system I have to work around?

That’s where software develops something resembling character.

Not through personality.

Through consistency.

Through boundaries.

Through thousands of small decisions that all seem to come from the same place.

Why Granola is worth studying

Granola isn’t proof that every AI app should look like Granola. And it certainly isn’t proof that restraint alone creates a durable business.

But it is a useful case study at a moment when building software is becoming dramatically easier.

When almost anyone can bolt a model onto an interface, the hard part becomes deciding what the product should actually be.

That means knowing what to automate—and what to leave to the person.

Knowing what context to keep—and what not to keep.

Knowing what to expose—and what to hide.

Knowing when to add a feature—and when to cut one.

Those decisions don’t show up in a benchmark.

But users feel them.

That’s what we think a product with a soul really is: software with a point of view.

And in the AI era, that may become one of the few things that gets harder—not easier—to commoditize.

Loved the article, share!