Designing for Trust: Startup Choices When Shipping an LLM-Powered Assistant

AI · 6 min read

LLM-powered assistants can delight users, but they also raise expectations and risk for misinformation, hallucinations, and unexpected behaviors. A recent startup, DraftMate, went through multiple product design iterations to land on an assistant interface that felt powerful but safe. Early prototypes auto-suggested full-paragraph edits and proactive facts; users loved the speed but noted missing provenance and occasional wrong facts.

Designers introduced three control layers: provenance chips that show source snippets for factual claims, an undo timeline for model-generated changes, and a 'confidence' badge that surfaced a model confidence estimate for factual content. The UX also separated creative prompts from fact-driven tasks through distinct modes so users could apply different mental models. Microcopy emphasized the assistant's limitations rather than promising correctness, and a lightweight feedback channel let users flag hallucinations in context.

Post-launch metrics showed increased engagement in creative mode but higher reliance on provenance chips in fact-checking tasks. DraftMate's retention improved when editorial users realized they could quickly verify changes without leaving the editor. The broader lesson for startups is to design with assumption of model error: give users control, surface evidence, and create clear modes so expectations align with capability.