System Integration versus App Distribution
Sam Hall, CEO of Wafer, on AI agent form factors
The real prize is not a better chat window, it is control of the layer where user behavior becomes usable training data. Going lower in the stack gives an AI product direct access to signals like microphone activity, screen state, app switching, calendars, and system context, which makes it far better at understanding intent. The cost is that distribution gets much harder because installs, OEM deals, and default status are far scarcer than a simple app download.
-
Granola is a clear example of why companies accept that trade. Its desktop agent sits in front of the microphone, watches for calendar invites and active call context across Zoom, Meet, Slack, and Teams, then captures the meeting and turns it into structured notes. That workflow would be much weaker as a normal web app.
-
Perplexity shows the opposite side of the trade. It has strong consumer distribution and reached $100M ARR in March 2025, but its assistant still depends on AppIntents, which means it can only do actions that app developers explicitly expose. That creates a narrower product than an OS level system can offer.
-
The strategic tension also comes from platform economics. Google and Apple benefit when users enter apps through app stores and pay inside those apps. An assistant that routes around apps and reads across them threatens that model, which is why startups and OEM partners have more room to push deeper system integration than incumbents do.
This pushes the market toward a split structure. Broad AI assistants will keep winning easy distribution at the app layer, while a smaller set of companies will chase the OS, launcher, assistant, and device layers to own richer context. Over time, the winners are likely to be the ones that turn raw system access into dependable actions, then secure distribution through hardware bundles and default placement.