Northwind
- Product strategy
- UX/UI design
- Design system
- Remote engagement
- 12 weeks
Northwind sells an AI assistant into enterprises that are, correctly, suspicious of it. The product was accurate and nobody believed it, because it answered like an oracle and offered nothing to check.
We worked alongside Northwind’s research team, who could tell us what the model actually knew about its own confidence. The design problem turned out to be exposing what already existed rather than inventing a trust layer on top.
An assistant that shows its working
A pattern library for AI states
Trust cues tested with real users
Show the working
Every answer carries its sources inline, and the passage that drove the conclusion is highlighted in the source rather than merely linked. Checking an answer costs one click instead of a search.
Say when you are unsure
Low-confidence answers look different — not a hedging sentence, a different component. It is the single change that moved pilot customers from testing the assistant to using it.
A record of what it said
Enterprises need to audit a decision months later. Every exchange is retained with the model version and the sources as they stood at the time, which is what turned a pilot into a procurement conversation.
Sample content
The trust problem was the product problem and we could not see it. Showing the model’s uncertainty as a different component, rather than a softer sentence, changed how our pilot customers behaved within a week.
Priya Raghunathan
