THE MOVING FRONTIER · JANUARY 2025 RETROSPECTIVE
The wrong unit of design
Software has traditionally been designed around screens. We map a workflow, divide it into pages, assign actions to buttons, and ask users to navigate from one state to the next. The screen becomes our organizing metaphor because it is the part we can see.
But customers do not buy screens. They buy a way to get something done. The manager approving an expense wants a defensible decision. The analyst preparing a briefing wants a reliable conclusion. The operator resolving an exception wants the issue closed without creating another one.
As models become capable of interpreting intent, retrieving context, and performing bounded actions, the screen becomes one possible interface rather than the center of the product. That sounds obvious; its consequences are not.
Three layers of a useful product
I find it helpful to separate the product into three layers. Intent is what the person is trying to accomplish, including constraints they may not initially articulate. Execution is the work required to produce the result. Evidence lets the person understand, verify, and take responsibility for it.
Traditional software makes people assemble those layers themselves. They translate a goal into a sequence of clicks, perform the steps, then reconstruct evidence from whatever the system recorded. An AI-native product can assemble more of that journey.
Imagine someone asking, ‘Which customer renewals are at risk, and why?’ A conventional product offers filters, dashboards, and account pages. A stronger product gathers relevant changes, explains drivers, links to records, and proposes next actions. The interface has not disappeared. It has moved closer to the decision.
Where the screen still matters
The temptation is to replace every workflow with a chat box. That is often wrong. A financial model, a comparison across dozens of items, or a high-stakes approval may demand a structured visual surface. People are good at scanning, comparing, and spotting anomalies when information is laid out spatially.
The question is not ‘chat or UI?’ It is ‘what should the system do, and what should the user be able to inspect or change?’ Some work is best delegated. Some belongs in a table. Some should stop and request permission. The best experience may combine conversation for intent, automation for execution, and structured views for review.
Measure the outcome
If a team measures only screen engagement, it risks optimizing the wrong thing. More clicks may mean deeper usage—or more friction. Fewer clicks may mean better automation—or less visibility into mistakes.
I would begin with time to a confident outcome. How long to finish the task? How much output is accepted? How often must a user reconstruct the reasoning? How costly is correction?
The next generation of durable products will not simply eliminate screens. It will make work simpler without making responsibility disappear.
Editorial note: This is a retrospective essay written in November 2026. January 2025 describes the period explored, not its original publication date.
