Designing for agentic AI: six UX patterns for control, consent, and accountability¶
Insight: Six UX patterns for the agentic AI lifecycle: Pre-Action (Intent Preview, Autonomy Dial), In-Action (Explainable Rationale, Confidence Signal), Post-Action (Action Audit & Undo, Escalation Pathway). The Autonomy Dial provides four levels of progressive authorization: Observe & Suggest → Plan & Propose → Act with Confirmation → Act Autonomously. Metrics include acceptance ratio (target >85%) and override frequency (>10% triggers model review).
Detail: The Intent Preview is the "conversational pause before action" — showing a complete plan (cancel flight, rebook, update hotel, email itinerary) with three control paths: Proceed, Edit Plan, Handle it Myself. The Autonomy Dial should be per-task-type (e.g., scheduling meetings vs sending emails). The framework applies across domains from travel assistants to DevOps release agents. Core principle: "Autonomy is an output of a technical system. Trustworthiness is an output of a design process."
Sources
Related: existing entry "AI interaction patterns emerging in design" in external/ai-assisted-design.md — COMPLEMENTS