Designing AI products people can verify

The problem is rarely accuracy. It’s verification.

Why don’t people trust an AI that’s almost always right? Because they can’t check it. A system that’s correct most of the time, with no way to tell which answers are the exceptions, is a system a professional has to redo by hand. The accuracy went up and the work didn’t go down.

That makes trust an interface problem before it’s a model problem. The question worth designing against isn’t “how do we convince people the AI is good” — it’s “how does this person confirm this particular answer, in seconds, without leaving what they’re doing”.

Evidence at the source

Every claim the system makes should point at where it came from, and getting there should cost a click rather than a search.

In TRACE, the AI document review workflow for PwC Document Intelligence, selecting an extracted value highlights the exact place in the contract or invoice it was taken from. Review time per document fell 60% in the pilot — not because the model got better, but because checking it stopped being a second full read of the document.

Orbit applies the same principle at a different scale. Instead of returning a paragraph, it assembles an investigation you can trace back through the documents, conversations and data it was built from.

Uncertainty as priority, not decoration

A confidence score printed next to every field is decoration. The same number used to order the queue — so a reviewer starts with the values worth a second look instead of scanning all of them — is a design decision that changes someone’s day.

Uncertainty should route attention. If the interface can’t say what a person should do differently because a number is low rather than high, the number doesn’t belong on screen.

Corrections that feed back

When someone fixes what the AI got wrong, that correction is the most valuable signal in the system — and in most products it’s discarded the moment it’s saved. Treating the correction as feedback rather than as an edit is what turns a review queue into something that gets lighter over time instead of heavier.

Recommend something. Don’t hand back the data.

Helio reads a building’s consumption, weather, occupancy, tariffs and equipment, and then says: shift battery charging to 2–4am to reduce tomorrow’s energy cost by 18%. It shows the reasoning behind that, lets you simulate alternatives, and puts the financial and environmental impact in front of you before you act.

The dashboard version of that same product displays every one of those numbers and decides nothing. “Don’t show me the data. Tell me what I can do with it” is the whole brief.

a recommendation is a decision you can argue with.

Design the output, not the input box

A text field is the fastest AI interface to ship and rarely the right one. Orbit is less chatbot, more workspace: the output is a structured map you can explore, challenge and follow back to its sources, because “investigate this” isn’t a conversation — it’s a task with a shape.

The question worth spending design time on is what the AI’s output should be. The input usually designs itself once that’s settled.

Prototyped with the people who sign off on the answers

AI interfaces are unusually bad candidates for design-by-discussion, because almost everything that matters depends on how the product feels when the system is wrong — and nobody can imagine that accurately, including the people building it. It has to be made and used.

So the work runs the same way the rest of the studio’s does: a prototype with real interactions, real logic and real data, in front of the reviewers, analysts or operators who would carry the consequences of a bad answer, early enough that the interaction model can still change.