Two years into the chat-window gold rush, we kept noticing the same thing inside customer accounts: the AI features that actually got used didn't open a new modal. They didn't make anyone "talk to the assistant." They simply did the next obvious thing — pre-filled the cart, drafted the reply, surfaced the right CPT code — and got out of the way.
The chat window is a tax
Asking a human to switch contexts, open a panel, and form an English sentence is a real cost. For internal operations work, it's almost always the wrong unit of interaction. The agent should already know which customer you're on, which queue you're staring at, and what "send it" means in this context.
That's not anti-LLM. It's the opposite. It means trusting the model enough to embed it directly in the workflow surface, with the right tools wired in, and only escalating to a conversation when the user wants one. The chat box is the fallback, not the front door.
What we ship instead
Our most-loved features have three traits in common: they're triggered by user intent, not by typing; they show their work inline; and they're a single click away from being reverted. That's the agent UI we believe in.