Most requests we get for AI features describe a chat interface. Most problems those clients actually have would be better solved by something with no chat interface at all.
The pattern that pays for itself
The pattern that reliably pays for itself is unglamorous: somewhere in the business, a person is moving information between two systems by hand. Reading an invoice and typing it into an accounting tool. Categorising support tickets. Extracting fields from a document. These are narrow, high-volume, tolerant of a review step — and that combination is where current models are genuinely strong.
The pattern that reliably disappoints is the open-ended assistant bolted onto a product, where the model must be right about anything a user might ask, and there is no cheap way to verify that it was.
The question that decides it
The engineering question is always the same: what happens when it is wrong? If a wrong answer costs a click to correct, automate aggressively. If a wrong answer silently corrupts financial data, you need a human in the loop regardless of how good the benchmark looked.
We would rather ship a narrow feature that quietly removes an hour of manual work every day than a broad one that impresses in a demo and gets switched off in a month.