AI-ASSISTED DESIGN PRACTICE
Rebuilding an Enterprise Workflow with Claude
What happened when an experienced design leader used AI to move from problem definition to working prototype — and where design judgment still mattered.
Alex Thanasenaris - 9 min read
Claude
AI Prototyping
Product Design
Interaction Design
The problem
Enterprise customer-care teams don't have a message problem. They have a judgment-at-scale problem. Every shortcut AI introduces — auto-classify, auto-route, auto-respond — is also a place where a wrong guess costs real time. The interesting design problem isn't whether to use AI here. It's where, exactly, to draw the line between what AI decides and what a person decides, and how to make that line visible to the person doing the work.
The system
The first AI output
I gave Claude the workflow and asked for a first pass at the agent workspace. What came back was fast, plausible, and wrong in instructive ways — it stacked six competing AI modules, surfaced a 97% confidence score with no explanation, and had the AI auto-send a reply and close the case with no human review.
What I changed and why
Two decisions stand out. The interface itself, and the smaller-but-sharper decision on the case-creation form's Owner field — where I left the AI's guess out entirely rather than risk a wrong route.
Iteration
The gap between the flawed first draft and the final experience took a genuine middle step: automation removed, Accept and Dismiss added — but still no Edit option, no source citations, and a reply that read like a form letter.
The final experience
Six screens make up the finished system.
What AI accelerated
Structural directions for the inbox and workspace, realistic message content across channels, fast component variation, and the repetitive parts of the design system. None of it replaced design judgment. All of it removed the friction that normally keeps a design leader from staying hands-on at this level of detail.
Where human jugdement mattered
Deciding where the automation boundary sits and defending it with a specific principle. Recognizing that a fully-confident AI panel is a trust problem, not a feature. Choosing to leave a field blank when guessing wrong costs more than asking. And rejecting the AI's first visual direction because it read as a demo, not software a Fortune 500 support team would trust.