Last February, I spent three days watching a support team at a logistics company in Rotterdam try to use the chatbot interface we had shipped them. They typed questions into a white rectangle and received paragraphs of text that were technically correct and practically useless. The model was doing its job. We were not.
That week changed how our team thinks about product. We had spent fourteen months obsessing over latency benchmarks and retrieval accuracy, and almost no time on the shape of the conversation itself — how information unfolds, when to stop talking, what a human actually wants when they ask a question at 4pm on a Tuesday.
The Uncanny Valley of Product Design
Most enterprise AI interfaces today sit in an awkward middle ground. They are too verbose to feel like tools and too sparse to feel like colleagues. The output reads like a Wikipedia summary — accurate, encyclopedic, and strangely impersonal. Users learn to tolerate them rather than trust them.
The output reads like a Wikipedia summary — accurate, encyclopedic, and strangely impersonal. Users learn to tolerate them rather than trust them.
We started prototyping a different model: short, structured answers with clear provenance — linking back to the exact document, the exact page, the exact paragraph the model drew from. We cut default response length by forty percent and watched task-completion rates climb from 61% to 83% in a single sprint.