Granola’s AI Meeting Notes Raise a Bigger Question: Who Controls Workplace Context?

AI meeting notes are becoming a store of workplace context, and the key question is who gets to use that record. Granola CEO Chris Pedregal says the company keeps notes private by default, even when an employer pays for the service. Its product challenge now is to make meeting information useful to AI agents without turning a personal notepad into a tool for monitoring workers.

Granola captures audio from the user’s computer rather than sending a bot into a call. It combines transcription with notes typed by the user, then produces an AI summary when the meeting ends. The design lets it work across video calls, audio calls and in-person conversations, but it also means other attendees may not see an obvious notice that transcription is underway. Granola has built an optional animated watermark that appears in the user’s video feed to make that activity visible.

Pedregal’s larger ambition reaches beyond summaries. He sees meeting transcripts as a useful source of context for AI systems that otherwise know little about the people using them. The company is exploring ways to turn that raw record into a compressed, structured memory that can help with meeting preparation, follow-up and other work. That direction could make Granola an important workplace tool—or leave it supplying context to agents made by other companies.

A notepad that captures meetings

Pedregal started Granola with co-founder Sam Stephenson after leaving Google. He joined Google as an associate product manager in 2008, later founded the homework-help app Socratic, and returned to startup work after Google acquired Socratic in 2018. In 2022, he began experimenting with large language models from OpenAI, before the public arrival of ChatGPT. Granola grew out of that period of exploration.

The product is designed to stay close to the user. It records system audio on the computer, so a user can start transcription without adding a separate meeting participant. People can take their own notes while Granola listens, and the app fills in details with a generated summary once the conversation is over. The combination matters: an automated transcript provides a record, while the user’s notes can capture emphasis and interpretation the transcript alone may miss.

For the interviewer, the product’s appeal was partly practical and partly emotional: having a record offered reassurance, even when there was little reason to search it later. Pedregal says that pattern is common. People may return to their notes less than they expect, but the ability to recover a forgotten detail can still be useful. In his view, not revisiting every conversation is not necessarily a product failure. People already retain what matters to them; the archive is there for the moments when memory falls short.

That idea becomes more complicated when software agents can search the archive on a user’s behalf. Pedregal says Granola’s connections to other AI tools are among the product’s fastest-growing uses. A person may not need to reread old meeting notes, but an agent can draw on them to answer a question or prepare for a task. Spoken conversations create a steady stream of information about projects, decisions and relationships. Granola’s bet is that this accumulated context can make AI systems more useful.

That context is also sensitive. Meeting records may include unfinished ideas, private concerns, personal details and candid comments that people would never put in a shared document. A system that preserves those conversations for future use must decide not only how to retrieve them, but who can retrieve them and for what purpose.

Making transcription visible

Granola’s decision to capture audio locally came from a product goal, Pedregal says: the app should work wherever a conversation happens. Third-party bots can be awkward in meetings and may fail to join certain calls. Capturing system audio offers a consistent way to transcribe a video meeting, a huddle or an in-person conversation. Pedregal says the company did not choose this approach to hide the notetaker, but acknowledges that its initial design gave other people little indication that recording was active.

The main video-conferencing services do not provide Granola with a general way to ask every participant for consent. They also promote their own meeting note features. So Granola built an optional watermark that appears on the user’s video feed. It works through a virtual camera: the app inserts the mark into the outgoing video before it reaches services such as Zoom or Google Meet. A company deploying Granola can turn the feature on, and Pedregal says organizations commonly want that visibility.

The workaround illustrates a gap in how meeting software handles third-party transcription. A bot joining a call can be conspicuous, but it may be technically unreliable or disruptive. Capturing audio from one participant’s computer can work more broadly, but leaves that participant responsible for telling everyone else. A watermark makes the activity easier to notice, yet it depends on video being used and the feature being enabled.

Pedregal expects workplace norms to change. He thinks people may come to assume that work meetings are transcribed unless someone asks otherwise. For social conversations, he hopes the default remains the reverse: transcription should require an affirmative request. That distinction reflects the different expectations people bring to professional and personal settings, though it also leaves open who sets those norms and how a participant can object when an employer has chosen the tool.

Granola is built for work conversations, Pedregal says, rather than for continuously recording someone’s personal life with a wearable device. Asked about a story involving someone using the app to record dates and send transcripts to another AI service for advice, he said that was not a recommended use. He added that recording should ideally happen with everyone’s agreement. The example underscores how easily a transcription product can move beyond its intended setting once it can capture audio and pass information to other systems.

Keeping the record under the user’s control

Granola says it has never stored audio recordings. The company can process the audio to produce notes, but chose not to keep the recordings because Pedregal wanted the product to feel like a notepad. The original plan was to delete transcripts after generating notes. The team changed course because transcripts preserve source material that users may need when revising notes or generating a different summary.

Now Granola is researching whether transcripts should be deleted automatically after a period of time. A verbatim account of a meeting from years earlier may be more detail than a user needs, while the decisions and context from that meeting could still matter. The company is considering whether it can compress the transcript into a shorter memory that retains useful information without keeping every word.

That trade-off is central to the product. Long-lived transcripts offer flexibility: users can check what was said, rebuild notes or ask new questions later. They also preserve material that may be sensitive, incomplete or easy to misinterpret when removed from its original setting. Compression could reduce the amount of raw conversation retained, but it introduces another question: what information does the system preserve, and what does it discard?

Pedregal says Granola notes are private by default. A user is the only person with access unless they choose to share them. The company says it maintains that rule even when an employer pays for the service, despite some customers wanting broad access to the records. A manager cannot simply browse an employee’s notes or run an agent across them, Pedregal says.

That boundary shapes Granola’s position in the workplace. The product can help a person remember what happened in a one-on-one, prepare for a follow-up or reflect on how a meeting went. Those uses can be valuable precisely because the notes are private. If the same records become visible to a manager, employees may speak differently or avoid recording sensitive conversations. The tool’s usefulness could change along with the audience.

There are trade-offs. Private notes make it harder for an employer to search every meeting record or assemble a shared account of work. Pedregal says companies can set up a culture where employees share notes, provided they understand that expectation. Granola does not offer a hidden route for managers to access private material. He argues that a backdoor would let people with power use accumulated context to make judgments about employees.

Privacy also raises questions about what should be shared when a user does choose to distribute notes. Granola is exploring ways to let people share a modified version rather than the full record. For a sales conversation, colleagues may need the discussion of a client’s needs, but not personal details about someone’s health or family. A user might want to remember a contact’s child’s name without making that detail available across a work team.

Pedregal sees these choices as a matter of human judgment, not a simple rule about whether a conversation should be transcribed. A meeting may contain both work information and personal context. The right sharing boundary can depend on the relationship, the purpose of the record and the expectations people have set with one another. Tools that treat every sentence as equally shareable risk flattening those differences.

From transcript search to useful memory

Granola’s current connections to other AI systems let users ask questions about their notes, including through Claude. But a large archive is not automatically useful context. Pedregal says an agent may retrieve a limited set of relevant meetings and still miss the wider picture. What he wants instead is a compact, continuously updated representation of what matters about a person: their work, goals, projects and ongoing concerns.

That compressed representation could then be supplied to different tools and agents. The aim is not simply to search more transcripts. It is to transform a growing record into context that can travel with the user, so an agent has a better sense of what the person is trying to do. For Pedregal, the gap between an archive and a usable memory is one of Granola’s main product problems.

Granola has already introduced pre-meeting briefs that prepare a user for a conversation with an external person. The feature draws on prior context to summarize what the user may want to know before the meeting. Pedregal says it addresses a task that technically minded users can assemble with general-purpose AI tools, but that takes effort to set up and maintain. Turning a useful workflow into a product is harder than demonstrating that a prompt can produce an answer.

That distinction matters for AI adoption. Many possible workflows require users to gather the right information, write a good prompt, connect systems and remember to run the process. A product can remove those steps, but it has to deliver useful results without creating more work. Pedregal says Granola’s brief feature still needs improvement, while arguing that common AI tasks should become proactive product features rather than repeated exercises in chat.

Proactive assistance has its own failure mode: another stream of notifications and tasks. People already manage email, work chat, personal lists and other feeds. An assistant that surfaces every possible action can add pressure instead of saving time. Pedregal says the company has found that users do not want another inbox filled with things they ought to do.

He also thinks AI may be more useful on small tasks than on the biggest decisions. People are usually already focused on their most important work. A system that drafts a thoughtful thank-you note after a meeting may help with something the person would otherwise forget. The product challenge is to surface those opportunities at the right moment, without distraction or an unmanageable queue.

More personal forms of assistance are possible as well. Before the interview, the host asked Granola to assess his meeting style. It suggested that he sometimes lets others set the agenda too much. He found the observation useful, but questioned whether a user should have to open an app and ask for that kind of feedback. Pedregal agreed that the more promising version would offer relevant help proactively, while keeping the experience calm and optional.

Context for agents, and a fight over interfaces

Granola is testing “enhanced transcripts” built for AI agents. The idea is to add information that makes a conversation easier for software to interpret. If a speaker uses an internal project name, the transcript could explain what the project refers to. If someone mentions a first name shared by several colleagues, the record could identify the intended person and connect that reference to relevant contact information.

Such a document would be more than a cleaned-up transcript. It would attach context to the conversation so an agent can reason about it with fewer ambiguities. The approach could make meeting records more useful as part of an organization’s knowledge base, while raising the same questions about access and sharing that apply to the original notes. The more information a transcript carries, the more consequential its privacy settings become.

Granola has connectors to Claude and ChatGPT, allowing users to ask questions about their meeting history through those tools. Pedregal says he was unsure whether the company should make its context so readily available to other agents. The information is valuable, and sharing it could help a user do more. But a system that supplies context to another company’s agent risks making that agent the place where the user actually works.

He says his view shifted after Claude Code drew strong interest late in the prior year. Users were finding new ways to apply agents, and meeting context could improve many of those workflows. Granola should make that context available, he concluded, while building its own products where meeting-specific features can offer a better experience.

One example is meeting preparation. A user had set up a custom process in Claude to create briefs before meetings, then stopped using it after trying Granola’s version. Another example came from a salesperson who used meeting transcripts to create client-specific websites, incorporating the client’s branding and interactive charts. Pedregal does not think Granola should compete to build those websites. Its role in that workflow is to supply the conversation context that informs them.

That division offers a possible strategy for a crowded category. General agents can use meeting records for many tasks, while a dedicated product can focus on preparing for conversations, extracting decisions and helping users follow up. The question is whether those focused features will be valuable enough to keep users in Granola for meeting work, even as they use other systems for everything around it.

Pedregal says transcription and meeting notes are an early, useful application of a much larger shift in workplace software. He compares the current moment to the early personal-computer era, when word processing was useful but the familiar range of software had yet to emerge. In his view, the more important competition is over the small set of tools people will use throughout their workday as AI agents become more capable.

He expects the familiar set of work apps—documents, spreadsheets, email and chat—to change. He is unsure whether users will do most of their work inside a general-purpose chatbot or rely on a different set of specialized products. Granola wants to be among the interfaces people return to, especially for work built around meetings. The company’s position will depend on whether it can turn context into dependable actions, rather than simply storing and retrieving conversations.

What AI changes about work

Granola has around 70 to 80 employees, according to Pedregal. He can imagine the company growing to roughly 100 or 120 people, though he does not say AI is the reason for that limit. He has long favored smaller teams and says that preference sometimes kept the company from hiring when demand was growing. The company is now hiring.

Pedregal rejects the idea that new AI tools will make large companies or human work unnecessary. Technologies change the tasks people do and the jobs that exist, he says, but expected gains in leisure do not automatically follow. He points to the personal computer as an example of technology that many expected to reduce work but did not lead people to work less.

AI agents might increase the amount of work a person can handle. That could reduce the need for some tasks, but it could also make it worthwhile to hire another person who can manage more agents. Pedregal does not claim to know which outcome will dominate. His view is that work will change while people remain central to it.

The promise of productivity gains also comes with a risk: organizations may use AI to generate more briefs, messages, slides and summaries without giving people more time. Pedregal argues that the pressure to work long hours is often cultural and organizational, rather than a direct result of the tools people use. Better software could shift time away from rote activity toward deeper thought, even if the workday does not shrink.

He expects meetings to remain a place where people decide what matters and what they believe. Some can feel routine; others help a team make choices that an agent cannot settle on its own. That leaves a role for products that make meetings easier to navigate, but it does not guarantee that automation will reduce the number of meetings or the demands placed on workers.

Pedregal does see a potential benefit in helping people manage interruptions. Notifications and apps have made it harder to focus, he says. AI might help individuals handle that stream, though he offers no settled design for how. A personal assistant could filter distractions, but it could also become another source of alerts and work. Granola’s own effort to avoid creating an extra task inbox reflects how narrow that design problem can be.

The underlying question is who experiences AI as their tool. Pedregal says the same technology can feel empowering when an individual controls it and oppressive when an employer uses it to monitor or direct them. Granola’s default privacy settings are one attempt to preserve the individual’s control, even though businesses pay for many workplace tools and may want access to their output.

Meeting transcription is a practical starting point for this debate because it turns ordinary conversation into a durable record. That record can help someone remember a decision, prepare for a customer call or understand their own habits. It can also reveal personal details, capture remarks without clear notice or give a powerful person information they would not otherwise have. How Granola handles sharing, retention and agent access will shape whether its archive feels like a private aid or workplace surveillance.

Pedregal’s product vision depends on getting both sides right: making conversation context useful enough to support new work, while keeping people in control of where that context goes. The next step is not just better transcription. It is deciding what to retain, what to summarize, what an agent should be allowed to see, and when the software should act without adding another burden to the user.