AI first is a research posture, not a feature

The pattern is easy to recognise once you have shipped it. A product exists, a model becomes available, and somebody adds a box in the corner where the user can type a question. The box is real work. It ships. It demos well. Then you watch the analytics and almost nobody opens it, and the people who do ask it to describe the screen they are already looking at.

Why the box does not work

A bolted on assistant asks the user to already know what to ask. That is a strange request, because the reason someone opens an enterprise tool is usually that they do not yet know the shape of their own question. It also leaves the original flow untouched. Every step that existed before still exists. So the assistant becomes a parallel route to the same screen, and a slower one, because now there is a sentence to compose.

Adding a model to a flow is a design decision. Asking whether the flow should still have that shape is a research one.

What we found across three personas

On an AI companion built for three very different users, the assumption going in was that they needed three interfaces. What they actually needed was three different jobs done, and the interface mattered far less than we thought. The analyst wanted the working shown, because she would be asked to defend the number. The manager wanted the number and nothing else. The person new to the tool wanted to know which question was worth asking, which is not a thing you can type into a box. One model, three jobs, and only one of them was well served by a place to type.

The posture

AI first does not mean the model goes in front. It means you ask what the job becomes when the model is assumed, before you decide what to build. That is uncomfortable, because it produces no artefact for several weeks. You cannot show a chat box in a review. You can only show that the five steps somebody used to take are now two, and that the two remaining ones are the ones they were paid to do.