How Building AI Products is Different | Notion After Hours

Building AI products at Notion hinges on a hard, practical tradeoff: deliver real value in messy, real-world inputs without over-constraining the model. The team’s core focus isn’t just getting a model to generate the “right” text—it’s designing the surrounding harness, permissions, steerability, and UI so the system can act safely and usefully inside enterprise workflows.

Several engineers and AI leaders trace that philosophy to long-horizon, stateful problems. Driving taught them that a model can’t be evaluated only in a controlled demo; what matters is what happens after step two or step three when the environment changes. That same combinatorial explosion shows up in chat: earlier messages reshape what’s possible later. In Notion’s framing, AI agents are essentially “driving” through enterprise knowledge workspaces—where context, prior actions, and permissions determine whether the next step is correct.

The conversation also draws a clear line between model labs and agent labs. Model capability matters, but the differentiator is expertise in integration and product layers: enterprise permissions, interpretability, undoable actions, and interfaces that meet customer expectations. Early on, Notion had to be explicit about what the system could do because users didn’t know how to use it. Over time, adoption improved as the product shifted from instruction-heavy interactions to goal-based setup—users specify an outcome (like configuring a relational database or generating meeting notes), while scaffolding and setup happen behind the scenes.

Decision-making for what to build follows a portfolio mindset rather than a single bet. Work is split across maintaining and improving what exists, tackling user pain points that demand new capabilities, and pushing capability boundaries. The team emphasizes that surprises are normal—early versions of features like meeting notes were “bad,” and some approaches (such as fine-tuning or tool-calling strategies) didn’t pan out. When assumptions fail, the response is not incremental patching but sometimes tearing down infrastructure and rebuilding.

Speed has become a competitive advantage. Demos are now far easier to produce than in earlier AI eras, but the real challenge is turning a demo into something valuable under real inputs. The team credits faster feedback loops—coding agents and rapid iteration—to the ability to reinvent parts of custom agent setup only weeks before launch. That requires low ego engineering: deleting code, revisiting assumptions, and letting the best idea win even when it’s someone’s own.

Finally, the group connects product building to everyday use. People describe using Notion AI to coordinate real-life tasks (like drafting a dog-care plan from messy group messages) and to manage personal workflows (credit card rewards tracking). Those stories reinforce the central message: the product’s success depends on harness design that makes models reliable enough to be useful, while still leaving room for the model to explore what it can do.