CHAI architecture combines LLMs with rule-based logic for explainable government chatbots

LLMs can draft nuanced answers, but in government settings, hallucinated or opaque responses are unacceptable. To address this, researchers from multiple institutions presented CHAI (Conversational Hybrid AI), an architecture that wraps large language models in a rule-based controller. The controller executes a Dynamic Condition Response (DCR) graph—a symbolic, declarative process-modeling language capable of expressing deontic, defeasible, and temporal logic. This lets the system encode both the rules of law and the steps of legal case management processes, making every chatbot response traceable to a specific regulation.

CHAI was demonstrated with a Covid-19 chatbot built from official government guidelines, showing how citizens can ask natural-language questions while the system grounds answers in an authoritative legal model. An ongoing case applies the same framework to supplementary grant applications for students with disabilities, illustrating how hybrid AI can manage complex casework. The paper, presented at the AIDA2J Workshop at ICAIL 2026, argues that bounded LLMs paired with symbolic reasoning offer a path to accurate, explainable, and accountable conversational agents—a critical step for AI in public administration.