Took an AI diagnostic tool from zero to 32,8% completion.
- 32,8%Completion, from zero↑
- 6Design iterations
- 3User groups, one flow
The client didn't want user testing. One workshop changed their mind.
01the problem/
Breakdowns came without warning.
Customers only called Alfa Laval once a heat exchanger was already down. Alfa Laval wanted them to see maintenance needs earlier, and to open the door to a premium preventive service.
02the users/
Users first, screens later.
Before any design work I pushed for clarity on who it was for. Repair technicians, support agents and account managers each had a different idea of useful. Naming that early kept the tool from serving only one of them.
03discovery/
Empathy before solutions.
I ran a workshop where stakeholders stepped into the shoes of technicians and customers. It surfaced unclear handoffs, even inside Alfa Laval, and the visible gaps built buy-in for the direction.
04the flow/
Scan, report, act.
One flow: open the app, add thermal images from the camera, get an AI health report, and contact support if needed. A traffic light sorts results into no faults, potential faults and critical faults, so the decision is easy to make and the service easy to explain.
05testing/
Making the case for testing.
The client didn't see a need for user testing. After the workshop I asked again, and we tested with support agents standing in for technicians. The flow held up. Camera connection failures and where to take photos did not.
06iterations/
Six rounds, from real feedback.
A serial number fallback for units not in the database, a products page with owned units, checkups and reminders, a support fallback to cut drop-off, and simpler unit checks. Alfa Laval's engineers then took it into development.
32,8%completion rate, from zero
07looking back/
What worked, what didn't.
- what worked/Testing with support agents. They see where things break across product, customers and field work.
- what didn't/Discovery had to earn its space. Arguments for more time didn't move the client, a concrete workshop output did.
- what I'd change/Book testing into the plan from day one, and test with real technicians as well as their proxies.

