Trust, Transparency, and Prompts: Redesigning AI Explanations in a Legal Research Tool

AI · 6 min read

The tool initially surfaced AI-suggested case law with a single line justification that often omitted scope and confidence, leading lawyers to accept matches without critical evaluation. User research with legal professionals uncovered two needs: succinct signal of model confidence and an easy path to inspect source rationale. The team launched a layered explanation model to meet both needs.

The redesigned card showed a confidence badge, three-line rationale, and an ‘expand for provenance’ link that revealed citations, model reasoning steps, and a quick way to flag inaccuracies. Designers also introduced adjustable prompt sliders enabling users to prioritize recall or precision in results. These controls were intentionally lightweight and reversible to avoid burdening expert users.

After deployment, adoption of the explanation panel grew to 54% among power users, time-to-first-valid-citation decreased by 21%, and the number of flagged erroneous citations dropped 30% as users could correct the model earlier in the workflow. This case highlights how transparency and user control over AI prompts can align model outputs with professional standards without slowing expert workflows.