Transforming public administration workflows with multi-agent AI: A human-in-the-loop knowledge orchestration framework

Public administrations face increasing pressure to manage large volumes of institutional knowledge while ensuring transparency, accountability, and timely responses to citizen and council inquiries. This paper presents CARE (Council Agents for Response and Engagement), a human-in-the-loop multi-agent artificial intelligence system designed to support knowledge-intensive workflows within the Autonomous Province of Trento. CARE orchestrates specialized AI agents through a stateful workflow architecture, integrating hybrid retrieval-augmented generation over a corpus of more than 160,000 administrative documents. Unlike conventional automation approaches, the system models existing governance processes, preserves institutional responsibility boundaries, and ensures traceable document grounding in response generation. Deployed in production for six months and used by 30 administrative staff members, CARE achieved a 70% reduction in response preparation time while maintaining human oversight and institutional control. The study contributes a socio-technical architecture for AI-assisted public administration, demonstrating how multi-agent orchestration, human-in-the-loop design, and hybrid knowledge retrieval can enhance institutional knowledge reuse without compromising accountability. Implications for digital transformation, AI governance, and responsible adoption of generative AI in the public sector are discussed.