The note-taking app that’s trying not to become a snitch
Granola has one of those product pitches that sounds almost too neat for office life in 2026: let the software listen, take the notes, and keep itself out of the way. The app is now valued at roughly $1.5 billion, which says plenty about how eagerly investors are betting on AI note-taking, but the more interesting part is the social contract it’s trying to write. It wants to sit in meetings, catch the details, and still avoid becoming the kind of tool employees squint at whenever a manager says, “Don’t worry, this is just for notes.”
Chris Pedregal built the company after stints at Google, where he also helped build Socratic before Google bought that education app in 2018. That history matters because Granola doesn’t feel like a quick slapdash startup trying to ride a buzzword wave. It looks considered, almost overdesigned in the best possible way. The interface’s polished, and the pitch’s tidy. And the underlying bet is messier: if software can remember every meeting, who gets to decide what that memory’s for?
The appeal of AI note-taking ends the moment it starts sounding like a listening post.
That tension has only sharpened as more teams adopt AI tools that can summarize calls, draft follow-ups, and pull action items out of a half-hour of rambling about launch timing and budget drift. On paper, this sounds delightfully useful. In practice, it nudges the workplace into awkward territory. A meeting note isn’t just a note anymore. It can become a record, a search object, a performance review aid, or a quiet weapon in the wrong hands. The mood in the room changes, once that possibility exists.
Workers already know what monitored software feels like. Time trackers, keystroke logs, screenshot tools, productivity dashboards that turn every blink into a metric. AI note-takers arrive carrying a friendlier face, but they still sit close to the same power and politics problem. If a system records what was said, who controls the transcript? If a boss can request access, does “team productivity” start to mean “management gets to replay every sentence”? And if nobody is sure who owns the record, does anyone truly consent?
That last question is where ai policy stops sounding abstract. Consent in a meeting is messy even before AI gets involved. Some people join late. Some are on the phone. Some are in a conference room with one camera pointed at a whiteboard and another at nobody’s best side. In digital culture terms, the old “mute yourself if you’re not talking” etiquette was simple compared with “please agree to be transcribed by software that may store, summarize, and redistribute what you said.” The rules are still being written, and the paperwork is always behind the product.
This means the timing also explains why Granola’s getting attention beyond the usual tech news crowd. AI notetakers are moving from novelty to routine. Sales calls, hiring interviews, internal planning sessions, customer meetings, therapy-adjacent wellness check-ins, and the random in-person catch-up that somehow becomes a project plan by minute twelve. Once transcription becomes normal, the record of a meeting stops feeling temporary. It starts looking like a workplace asset.
Or, depending on your seat at the table, a workplace hazard. Pedregal’s company’s now trying to sell a useful version of that future without sounding like it wants a surveillance badge on the lanyard. That’s a delicate line. If it works, Granola could become one of those everyday tools people barely notice until it disappears. It may end up as a warning label for the entire category, if it doesn’t. Either way, the question is already in the room: can an AI notetaker help people remember what was said without turning every meeting into evidence?
How Granola works when it refuses to behave like a bot
Granola’s whole pitch depends on a small piece of product weirdness: it listens without arriving as a little robot in the meeting roster. Instead of joining a call as a visible bot, the app captures audio straight from the user’s computer. That sounds like a technical detail. It’s really the entire trick. By pulling sound locally, Granola can follow the action in Slack huddles, Zoom calls, Google Meet meetings, and even the old-fashioned kind where people sit around a table with coffee and regret. No extra attendee. No “Granola Bot is joining the call” moment. Just the note taker doing its job in the background.
That design makes the app unusually flexible, which is part of why it has caught on with people who move between tools all day. A meeting can start in one place, continue in another, and end with somebody standing in a hallway saying, “Can we just grab five minutes?” Granola is built for that mess. It does not care much whether the conversation is happening through a conference link or across a laptop mic at a desk. It just needs access to the sound coming out of the computer.
The privacy trade-off is sitting right there in the same machinery. Granola says it’s never stored audio recordings, which is a nice line to have if your product spends all day hearing other people’s voices. The company also started with a stricter idea around transcripts. At one point, the plan was to delete them even more aggressively. That changed when the team ran into a boring but unavoidable problem: if you want to generate notes again later, you need the raw text that produced them. AI can do a lot of theatrical gymnastics, but it still needs a script.
The cleanest way to avoid becoming surveillance software is to be honest about when the machine is listening.
That honesty is harder than it sounds, because meeting platforms do not always make consent easy to manage through normal APIs. A bot can announce itself. A hidden recorder cannot. Granola’s workaround is a visible animated watermark shown through a virtual camera, so users can signal that transcription is on. It is not subtle. That is the point. The watermark is there to keep the recording from becoming the sort of thing people notice only after the fact, once someone asks, “Wait, was this transcribed?” and the room suddenly goes quiet in the worst possible way.
This is where Granola runs into the broader anxiety around workplace surveillance. People don’t usually object to note taking itself. They object to the possibility that the note taker is also a quiet witness, storing material for future inspection by managers, HR, or whoever else gets curious after a rough quarter. The fear isn’t imaginary. Meeting transcription’s become easy enough that the main question is no longer whether it can be done. It’s who gets told, who gets access and how visible the process’s while it’s happening.
Platform rules are shifting around that same issue. Google Meet now requires explicit consent for its Take Notes with Gemini recordings and transcripts, which shows how messy the whole business has become once transcription is treated as a default feature rather than a special add-on. Other note-taking products, including Fellow’s feature set, have also had to be plain about how they handle meetings, and Fellow’s security and privacy notes spell out the sort of disclosures buyers now expect to see before they let software sit in on a call. Nobody wants to discover the fine print after the transcript has already been mailed to half the company.
Pedregal’s own view’s fairly blunt. He seems to think the etiquette around meetings may flip. Today, the default assumption’s often that a meeting’s private unless everyone explicitly agrees to record or transcribe it. He thinks that may change. In his version of the future, transcription becomes the default state and people opt out if they don’t want the machine present. That’s a pretty large social shift for a feature that started as a convenience tool. It also says a lot about how fast AI notetakers are moving from novelty to routine office infrastructure.
For now, Granola’s trying to sit in the awkward middle. It wants to be useful in the exact moments when people need notes most, which is usually when they’re too busy talking to keep writing them down. At the same time, it can’t afford to look like a stealth recorder with a shiny interface and a startup valuation. The product has to feel easy, but not sneaky. Helpful, but not creepy. In this corner of lifestyle tech, that balance may matter more than any summary the app produces afterward.
From transcripts to ‘apple juice from concentrate’
the real question starts: what should happen to all that text after the call ends?, once Granola’s captured a meeting. For the company, it’s now testing a setup that’d automatically delete verbatim transcripts after a period of time, while keeping a compressed working memory around for later use. In plain English, that means the full transcript wouldn’t sit there forever like a filing cabinet stuffed with every cough, tangent and half-formed thought. What survives is the useful residue: who promised what, which decision got made, what changed since last week and what still needs follow-up.
The sell is no longer just “we can record the meeting.” It’s “we can keep the parts a machine will actually need.”
That shift matters because Granola’s increasingly being built for two audiences at once: people who want a cleaner record of their work, and software agents that need context before they can do anything sensible. The company’s been experimenting with what it calls improved transcripts, a version of meeting text designed less for human rereading and more for AI systems that need structure. A verbatim transcript is fine if someone wants to search for a phrase they half-remember. It’s much less elegant when an assistant has to figure out whether a discussion ended with an action item, a decision, or a polite shrug.
The idea gets more interesting when you add MCP connectors into the mix. Those connectors let outside tools read Granola context without forcing every workflow through the same app. That can mean a knowledge system pulling in meeting memory, a task tool seeing the outcome of a client call, or an agent checking prior conversations before drafting a reply. In a world where companies keep rewriting their AI policy every few months, that kind of controlled access matters. It gives IT and legal teams something clearer to think about than “the bot heard everything, good luck.”
Granola’s team says agent usage is the fastest-growing part of the product, which is a tidy bit of evidence that the market’s moved past novelty. People are no longer just asking for notes because notes are convenient. They’re asking software to do the sort of work a capable assistant would do with a full brief in front of them. That only works if the transcript carries enough context to be useful. Names matter. The tone of a disagreement matters.
So does the tiny sentence that tells you a deal’s basically done, except for one annoying clause that nobody wants to touch yet. That’s why the company sees raw transcripts as a goldmine of context. The more detail the system keeps, the less guesswork an assistant has to do later. It can surface the right thread before a follow-up call. The reality: it can remember that a client hates quarterly check-ins but likes written recaps. It: it can remind someone that a promised intro never actually happened. The raw transcript’s messy, sure, but it contains the material that makes an assistant less generic and more useful. Strip too much out too early, and you’re left with notes that sound polished but don’t carry enough weight to help another tool act on them.
There’s already a visible pattern in how this plays out across the rest of the meeting software market. Zoom’s own AI Companion security and compliance documentation shows how the enterprise conversation has moved toward guardrails, retention, and who can see what. Microsoft Teams makes live transcript capture a routine part of meetings now, with its start, stop, and download transcript controls sitting right in the product. Even botless tools like Fellow’s recording workflow show how normal it has become for meeting text to flow into software without a little robot avatar joining the call. Granola is pushing that same direction, just with a sharper focus on what happens after the call is over.
That’s where pre-meeting briefs enter the picture. A brief can pull together what matters before an external call: the last conversation, the open issues, the names involved, the tone of the relationship, and maybe a warning if the other side just changed direction. For anyone who has opened a meeting invite five minutes before the start time and stared at the calendar like it’s a bad joke, that kind of prep’s easy to appreciate. Right now, some users are cobbling this together themselves by copying notes into other tools, asking an assistant to summarize old transcripts, or stitching together prompts that work well enough until they don’t. Granola’s view’s that this should be productized, not left to power users who enjoy assembling their own tiny workflow contraptions before lunch.
The deeper bet is simple enough to say and messy enough to execute: a meeting record shouldn’t have to stay verbatim to stay valuable. If the system can keep a compressed memory layer, feed it to agents, and expose it through MCP connectors in a controlled way, then the transcript becomes something more like working infrastructure than dead storage. The catch, of course, is that once software can read the memory, someone will want to know who else can. And that question, as it turns out, doesn’t stay buried for long.
Who owns the meeting log?
Granola’s founder, Chris Pedregal, sounds unusually blunt about the line he won’t cross. Notes are private by default, he says and the company won’t hand CEOs a secret route into every employee transcript just because a boss asks nicely, or not so nicely. That stance gives the product its personality. It also puts a very real edge on the question sitting under all this cheerful productivity software: if an AI notetaker remembers the room better than any person can, who gets to own that memory?
The fight is not about typing less. It’s about who keeps the record after the meeting ends.
That question lands because meeting logs have always had a slightly suspicious twin life. On one side, they’re convenient. A messy half-hour call turns into a searchable record, a list of action items, a reminder of who promised what and when. On the other. They can become a disciplinary artifact, a little archive of every cautious phrase and half-finished thought. Slack ran into a similar tension when it turned workplace chat into something closer to a filing cabinet than a pile of stray messages. Useful? Absolutely. Innocent? Not quite. Once everything is searchable, everything feels one admin login away from being repurposed.
Granola sits right in that gap. Its pitch’s that a meeting log belongs to the person who sat through the meeting, not to whoever controls payroll. That sounds simple until you imagine the office politics. A manager wants the transcript to settle a disagreement. HR wants a cleaner record. A worker wants the notes because they missed a sentence while muting their dog, chasing a toddler, or both. The same document can be a memory aid, a corporate knowledge base and a liability.
Pedregal seems comfortable with that tension, partly because he’s not buying the loudest doom-and-gloom predictions about AI wiping out jobs overnight. He’s been around enough product cycles to know the industry loves a grand exit line. His view’s much calmer, and frankly more believable: AI is still early, transcription is still early and what we’re using now may look primitive later. He compares this stage to the word processor era, which is a nice way of saying we may be arguing about the wrong interface while the real one hasn’t arrived yet.
That doesn’t mean the stakes are small. It means the stakes are messy. Today, an AI notetaker might be a convenience feature. Tomorrow, it might sit between workers, managers and the model vendor that processes all the text. If the transcript feeds an assistant, then the assistant wants access. In a way, if the assistant gets access, someone wants controls. If someone wants controls, somebody else wants exceptions. That’s how a tidy product feature turns into a governance problem with snacks.
Pedregal’s bet is that people will accept AI in the room as long as it behaves less like a spy and more like a notebook they can trust. The harder question’s whether that trust can survive once the notes become part of the infrastructure. Can Granola become standard workplace plumbing without handing the whole stack to OpenAI or Anthropic, and without turning into the tool managers use when they want receipts from every meeting? That’s the real contest now, and it’s a lot less cute than the software’s polished interface would suggest.
