An AI feature in an enterprise asset management (EAM) platform that finds similar past failures, so maintenance teams can diagnose faults without asking the experts. It shipped and is live.
Placed in the existing sidebar, after the current options
The feature sits in the record's existing sidebar, after the options users already know (widgets, images, documents). It sits after a divider and has an animated icon.
| Position | What I found | Outcome |
|---|---|---|
| At the top | People would click it by mistake when reaching for widgets. | Rejected |
| At the bottom | It wasn't prominent enough. | Rejected |
| After a divider, below the existing options | Tested internally with stakeholders. | Chosen |
Analysis starts on a single click, not automatically
Users start the search for similar failures with one Start analysis button. Running an LLM is expensive, so a deliberate click means the model doesn't run every time someone moves between records. The trade-off is one extra click before results appear.
Dedicated AI colours, always paired with icons and words
AI-generated content gets its own colour treatment, so it is clear at a glance that the feature is AI and can make mistakes. Colour is never the only signal: an AI icon and a warning that results may contain mistakes carry the same meaning. Colour pairs were checked with contrast checkers against WCAG 2.2.
Protanopia
Deuteranopia
Tritanopia
AchromatopsiaKnowing it's AI
Colour and icons, plus a warning that AI-generated results may contain mistakes.
When the AI is unsure
Only results with a relevance score are shown.
Human approval
The user decides which past issue is most relevant and generates work instructions from that card.
Undo and correct
The user chooses whether to move forward through the suggestions and whether to generate work instructions, and can stop or correct at any point.
The assistant offers help when it spots similar failures
While the user works in a work order, a notification beside the assistant icon says similar issues have been reported. Nothing runs until they choose Begin Analysis.
Out of 5 results, the best suggestion is shown first
The panel opens on a Best Match card: the issue, its likely cause, the recommended solution and chips linking to previous failures by work order and date. From here, the user can press Generate work instruction. The four related symptoms sit below, along with a reminder that AI can make mistakes.
Generated steps reach the work order only once the user confirms them
Generate work instruction opens a modal marked AI generated, with five maintenance steps from preparing the equipment to operational testing. The user can edit them or confirm. Confirmed steps fill the Work Instructions card in the work order, and the assistant stays docked on the right, so past failures and related symptoms are still in reach.
Let’s connect
I’m always excited to discuss new opportunities and ideas. Drop me a line, and let’s start a conversation about your next project.