Challenge

Teams pasted generic chat widgets into existing layouts without clarifying what users should trust, how to recover from incorrect model output, or how to escalate to humans without losing context.

Context

Design systems lacked patterns for citations, partial responses, token streaming, and multi-modal attachments. Product metrics treated assistant success like a web form conversion, missing quality and safety signals.

Approach

• Ran contextual inquiry and task-based usability sessions on realistic jobs-to-be-done, not demo scripts.
• Extended the component library with states for loading, streaming, low confidence, refusal, and human takeover.
• Co-created content guidelines for tone, disclosure, and error copy with legal and brand stakeholders.
• Paired with engineers on latency budgets, optimistic UI risks, and skeleton screens that match final layouts.
• Delivered annotated Figma files, acceptance criteria, and a prioritized backlog for the first two engineering milestones.

Artifacts

Sprints produced journey maps, service blueprints, and a measurement plan tying UX changes to task success rate, time-on-task, and human review volume.

Outcome

Teams entered build with aligned narratives and fewer late surprises; rework in the first release cycle dropped because edge cases were named and designed before code hardened.