CoPilot-Finance: Graduated Autonomy for Human-AI Collaboration in Investment Advisory
Abstract
AI-powered robo-advisors promise democratized financial guidance, yet adoption remains limited by trust deficits and opaque recommendation processes. We propose CoPilot-Finance, a human-AI collaboration framework featuring graduated autonomy that dynamically adapts the AI's decision-making role based on three measurable factors: decision complexity, advisor expertise, and market risk context. Unlike existing binary accept/reject interfaces, our framework defines a five-level autonomy spectrum—from passive information display to conditional full autonomy—governed by a formal selection algorithm. We contribute (1) a principled autonomy selection algorithm grounded in Sheridan's supervisory control taxonomy, (2) an expertise-adaptive explanation mechanism that conditions justification detail on assessed user proficiency, and (3) a proposed evaluation protocol with preliminary simulated results using 48 synthetic advisor profiles. Simulation estimates suggest 28% improvement in decision quality and 35% reduction in decision time. Scope: All quantitative results are simulation-based; human participant validation is planned as future work pending IRB approval.