Designing trust: UX decisions for integrating generative AI into a customer support product

AI · 7 min read

SupportFlow added a generative composer to help agents draft replies. Early pilots showed faster response times but raised problems: over-reliance on AI drafts, hallucinated facts, and agents unsure how to correct the model. The design team adopted four UX patterns: transparent provenance (source snippets and confidence scores), inline edit affordances, staged automation (suggest vs autocomplete), and escalation paths with one-click citation insertion.

The composer UI surfaced the model’s source fragments and allowed agents to toggle visibility. Rather than an all-or-nothing autocomplete, drafts were offered as editable suggestions with clear “update source” and “verify” actions. The product also added soft policy nudges—inline reminders for sensitive topics—and a mode for high-risk tickets that disabled autogenerated content until a supervisor verified the reply.

After rollout, average first-response time dropped 28% while the number of agent edits per suggested draft stayed at a healthy 1.4, indicating agents treated drafts as starting points rather than black-box answers. Incident reports of hallucinated facts decreased after the provenance UI and verification mode were introduced. The case underlines that integrating generative AI requires UX controls that keep humans accountable and informed rather than invisible automation.