The Interface Is Where the Model Meets a Person

I designed conversational systems before the current wave. Teneo, Amazon Lex, Dialogflow, the Microsoft Bot Framework. Every path a user could take had to be written by hand, and if you had not written it, the assistant fell over. It was slow, unglamorous work, and it taught me something that the last two years have made more important rather than less: the interface is where the model meets a person, and the person does not care how clever the model is.

Large language models removed the hand-authoring. They did not remove the design. If anything they moved more of the design into the interface, because the model will now say almost anything with confidence, and the screen around it is the only thing that decides whether that confidence is safe.

What the model cannot know

Earlier this year I built a demonstration for a large US insurance carrier. A fictional New Jersey company, a real product set, real state regulation, one policyholder with a real-looking record. The assistant answers from her policy: whether her basement is covered if the storm floods it, why a burst pipe is different from rising water, which endorsement she does not hold. It is good at that, and it is good because of what surrounds the model rather than the model itself.

The model does not know that it must never compare carriers. It does not know that New Jersey has two tort thresholds and that recommending one is legal advice. It does not know that when a woman says a tree has come through her roof and water is coming in, the right response is a claim started and a named agent notified, not a paragraph. All of that is design. It is written as boundaries, as a handoff pattern, as a compliance layer that educates and never advises, and it is tested the way a user interface is tested, by putting the wrong thing in and checking what comes out.

When she says she would rather talk to her agent, the assistant does not argue. It sends the transcript and her contact details to Michael Torres at the Princeton office by name and tells her what to do in the meantime. That is an interface decision, and it took longer to get right than the model integration did.

The failure state is the product

Most of the AI interfaces I am asked to look at have been designed for the happy path. The demo works. The answer is good. What they have not designed is the moment the answer is wrong, or partly wrong, or right but unwelcome, and that moment is where trust is won or lost.

In the intent-model days this was explicit. You designed the fallback because the fallback fired constantly. Now the fallback fires rarely, which makes it more dangerous, because nobody has looked at it. A good AI interface shows its sources, states its limits, offers a person, and keeps the conversation when it hands off so nobody repeats themselves. Those four things are not model features. They are UX and UI decisions, and they are the ones I spend my time on.

What changed and what did not

What changed is the cost of producing a screen. I can go from a sketch to working code in an afternoon, and I do. What did not change is that someone has to decide what the screen is for, who it is for, what it must never do, and how you will know whether it worked. The tools got faster. The judgment did not get automated.

I hesitate to call the older technology unrealistic. It worked, within its limits, and the limits were the point. The people who did that work know something the newcomers are learning the hard way: a conversation is a designed thing, and the design is mostly about the edges.