The AI no one trusted
The system could produce grades in seconds.
Upload an image, and machine vision would analyse quality, detect defects, and assign a grade faster and more consistently than any human. On paper, it was powerful. On the ground, it was ignored.
Operators would glance at the result and go back to manual checks.
“Not sure if I trust this,” was the common reaction. The AI worked, but the experience didn’t.
The trust gap
The problem wasn’t accuracy; it was understanding.
The system gave answers, but no reasoning:
A grade appeared, but why?
A defect was flagged, but where?
A decision was made, but could it be trusted?
For users in the field, warehouse operators, graders, and suppliers, this wasn’t just data. It impacted real money. Designers saw outputs. Users needed confidence.
From output to explanation
I stopped asking:
“How do we show AI results?”
And started asking:
“What would make someone trust this decision?”
The answer wasn’t more data.
It was clarity.
I redesigned the experience to make AI decisions explainable:
Visual highlights showing exactly where defects were detected
Confidence indicators instead of absolute answers
Step-by-step breakdowns of how grades were derived
The system started showing its thinking, not just its conclusions.
Designing for the real world
This wasn’t a lab environment. Users worked in fast-paced, low-attention conditions often outdoors, with time pressure and variable lighting.
So we simplified everything:
Clear, glanceable results
Minimal steps to capture and analyse
Interfaces that worked reliably in imperfect conditions
The goal wasn’t sophistication, but usability under pressure.
What changed
The UI didn’t just look better.
It changed behaviour. Before, the system gave answers. Afterwards, it built trust. The shift wasn’t in the model, it was in the experience around it. Because in the real world, AI isn’t useful when it’s smart. It’s useful when it’s understood.
