Calm AI for Crazy Days: Inside Granola's Design Philosophy, with co-founder Sam Stephenson

What It Covers

Sam Stephenson, co-founder and designer at Granola — the AI meeting notes app that raised $125M at a $1.5B valuation — explains the product philosophy behind one of the fastest-growing AI tools on the Ramp spend tracker. He covers viral growth mechanics, inference cost management, privacy architecture, feature restraint, and how AI is reshaping the design-to-ship pipeline at a 60-person company.

Key Questions Answered

• •Extreme-user design framework: Granola targets people in back-to-back meetings all day — not because that's the only audience, but because designing for the most frazzled, system-one-mode user produces a calm, minimal product that works for everyone. This mirrors OXO's strategy of designing kitchen tools for people with disabilities, which consistently produces better products for the general population. Build for the hardest edge case first, and the mainstream experience improves automatically.
• •Single viral mechanism: Granola's growth — ranking second on Ramp's new-customer additions in January, behind only Anthropic — is driven almost entirely by one mechanism: users sharing meeting notes with teammates. Recipients see polished notes appear in Slack within seconds of a meeting ending, ask how it was done, and sign up. One reliable viral loop, executed well, outperforms multiple weak hooks. Teams should identify and optimize their single strongest sharing moment rather than spreading growth efforts thin.
• •Inference cost sequencing: Granola deliberately avoided optimizing inference costs at launch, treating unlimited usage as a product quality investment. Transcription once consumed half the company's total burn rate, but costs dropped significantly as the technology commoditized. The current per-seat pricing works because meeting frequency has natural physical limits. As Granola adds LLM-heavy features like multi-meeting chat and folder auto-classification, usage-based pricing tiers — similar to Cursor or Claude's model — are the likely next step.
• •OS-level audio vs. bot joining: Granola captures audio at the operating system level rather than joining calls as a visible bot participant. This enables recording across every meeting type without setup friction. Critically, raw audio is discarded immediately after real-time transcription via Deepgram and AssemblyAI APIs — only the transcript is stored. Speaker separation quality suffers compared to batch processing, but the privacy and trust tradeoff is considered worthwhile. On-device transcription models are the likely next architectural move for speed and cost.
• •Feature restraint as a design principle: The core meeting notepad is treated as a near-sacred calm space — new features require an exceptionally high bar to appear there, because users are giving Granola roughly 2% of their attention during meetings. More experimental features are deployed in the chat and folder interfaces, where users are in a focused, system-two state and can handle more complexity. This two-tier attention model — 2% during meetings, 80% in post-meeting review — should govern where product complexity is acceptable.
• •Recipes as discovery and marketing: Granola's recipes feature — reusable prompt templates triggered by one click — serves two functions: marketing through social sharing when users publish their recipes publicly, and depth-of-use education for existing customers. Standout use cases include converting messy team discussions directly into documentation updates, drafting job descriptions from unstructured conversations, and running personal coaching analysis across months of meeting history. The emotional response to coaching-style recipes during user interviews was described as among the strongest reactions the team has ever observed.
• •High-fidelity user research over scale: Granola makes core product decisions almost entirely through qualitative observation of a small number of users in real contexts, not A/B testing or large beta cohorts — despite having roughly 10,000 beta users. A key technique: asking users to share their Granola home screen and walking through each past meeting one by one to surface real organizational behavior, rather than asking abstract questions about hypothetical workflows. This grounds interviews in observable facts and prevents users from describing an idealized version of their own behavior.