Sam Stephenson on AI Meetings, Calm Design, and Building with Agents

Granola co-founder Sam Stephenson says a useful meeting assistant begins before the call: a brief brings together relevant context so that participants arrive prepared. Peter Yang and Sam discuss how AI can make meetings less draining and how the company builds its own tools. The episode is sponsored by Google AI Studio, with Linear and Riverside also listed as sponsors.

One popular recipe answers “what did I miss?” after a meeting. Rather than forcing people to read a full transcript, Granola distills the conversation into useful notes, helping a participant return to the parts that matter.

Sam shares Granola’s design principles: be calming, keep the core experience simple, and put the user in control. Meetings already create enough noise, so the product should feel like a handrail rather than another participant competing for attention.

Those principles shape a controversial choice: the transcript is less prominent than the notes. Sam accepts some loss of discoverability in exchange for a simpler experience. A transcript remains useful, but the team wants the default view to focus on what people need next.

The team weighs requests for more active intervention against the atmosphere it wants to create. A real-time bot pointing out disagreements might be impressive, but could also interrupt a conversation; post-meeting analysis can offer feedback without making the assistant the center of the room.

Sam demonstrates using meeting context to solicit feedback in the style of a founder or investor. He asks the assistant to reflect on ideas from the conversation, giving it product principles as context so the response can challenge proposals against the team’s actual priorities.

Granola can draw on shared team meeting notes as well as a person’s own accessible information. The distinction matters: Sam does not claim that a user can search everyone’s private meetings. Shared sales conversations, for example, can help surface recurring customer insights.

Peter asks about scheduled reports. Sam says the team has prototyped automation but wants high-quality outputs; native scheduling is a possible future direction, not a feature he promises is already available. Users can already combine Granola’s API or MCP with external tools. Peter describes sending himself a summary that joins Granola with other sources.

For a designer, the prospect that an agent might bypass the interface brings a pang of sadness. Sam says the company initially resisted opening up, but enterprise customers need tools to work with their internal systems. Granola can collect meeting context while other applications and agents put it to work.

Peter sees particular value in combining meetings with documents and other sources. Sam says the work behind pre-meeting briefs required finding relevant information in email, and sees opportunities to connect other company knowledge. They joke about agents meeting first and involving their owners only when necessary.

The discussion turns to the way Granola works. Peter worries that AI-native companies can normalize relentless hours. Sam describes trying to use the technology to build an effective company while preserving a calmer working culture and making room for considered product work.

Sam says his use of Figma has fallen as coding agents make working prototypes easier. This does not make design judgment obsolete: the team still needs a shared design system, reusable components and clear principles to prevent the speed of generation from producing incoherent interfaces.

Designers start with ideas, sketches and discussions, then explore working code with agents. A shared system gives the agent constraints and lets the team compare real interactions instead of arguing only over static mockups. Sam still uses conventional design tools when they are the best way to think through a problem.

He shows an interactive prototype for pre-meeting briefs. The team can try different placements and behaviors in something resembling the real interface. Getting an imperfect version in colleagues’ hands early exposes problems and supports repeated improvement.

Granola also uses an internal agent called Nacho. Sam describes it as a first stop for company questions: it can draw on internal information and tools, including engineering and analytics resources, to help people investigate issues or answer questions without starting from scratch.

An engineer built the early version, and the wider team has helped improve it. The same pattern applies beyond engineers: people who understand a problem can use agents and shared tools to investigate and contribute. Sam still emphasizes that context, access and reliable answers matter.

The final lesson is that making a first version has become easier, but craft depends on what happens afterward. Teams can release something, observe where it falls short and keep refining it. Sam argues that many people stop improving a product too early; AI should create more room for that work rather than a race to produce endless features.