Insight
Designing AI Products
When the model is part of the product, not just the tool you built it with, you stop designing screens. You design how a person and a system work together, with the system doing its part out loud.
The old flow: deterministic screens
Click a button, upload the PDF, wait, view the results. A straight pipeline where the software moves data and the person just waits — every step the same input, same path, same output.
The new flow: collaboration
The same task reshaped: the AI starts extracting the moment the file lands, streams its progress, flags the fields it's unsure about, and hands them to the person to fix and approve. The same PDF, and a very different thing to design.
The real design surface: the questions
When should the AI speak, and when stay quiet? How much confidence should it show, and how do you make uncertainty visible? When does a human approve? These interaction decisions — not the screens — are where the design work moved.
The loss-run extraction we built at Ennabl used to be a straight line. A producer clicked the button, picked the PDF, watched a spinner, and came back to a table. Then they checked every row by hand, because the software's only job was to move numbers from the PDF into the system. I designed the screens and the flow between them, and that was the whole job.
Now the model sits inside it. You drop the loss run and it starts reading the second the file lands. Carrier, effective dates, loss amounts, showing up as it finds them. It marks the two or three it's not sure about instead of pretending it's certain. The producer fixes those, approves, exports. Same task, but there's nothing linear left for me to lay out. I'm designing a back-and-forth between a person and a system that's working right next to them.
That back-and-forth is made of decisions I never had to make before. When should it speak up, and when should it stay out of the way? How much confidence should it show, and how do I make "not sure" visible without making it scary? What runs on its own, and what waits for a person to say yes? How does someone get out when the model gets it wrong? None of these are layout questions. The first version showed every field with the same confidence, and people stopped checking any of them.
The easy move is to treat the AI as a feature you add to a screen. A button that summarizes, a panel that suggests. But once the model is doing real work in front of someone, what I'm actually designing is whether they trust it. Whether they believe what they're seeing, know when to step in, and still feel in control of something that's partly out of their hands. Get that wrong and it doesn't matter how good the model is.
Most of what I design now is whether the producer believes the three fields the model flagged. It's a strange job.