A Modular Low-Code Architecture for Trustworthy Conversational Agents with Governance and Clinical System Integration
The rapid adoption of large language models (LLMs) has accelerated the use of conversational agents in digital health. However, in regulated and safety-critical environments, challenges related to trust, governance, and controlled integration with clinical information systems continue to limit their practical deployment. In particular, reliance on model-level behavior alone is insufficient to ensure operational safety and compliance.This paper proposes a modular architectural pattern for governed orchestration of LLM-based conversational agents in healthcare. The approach integrates (i) explicit separation between institutional knowledge, structured clinical data, and session-level context, (ii) a configurable governance layer that decouples prompts, model parameters, and operational constraints from workflow implementation, and (iii) controlled interaction with clinical systems through rule-based, tool-restricted operations.The architecture was implemented using a low-code orchestration platform and evaluated in a test healthcare environment through controlled scenario-based testing. The evaluation included repeated execution of operational scenarios, inspection of workflow logs, and analysis of system behavior. Observations indicate consistent enforcement of architectural constraints, absence of unsupported data generation in tested scenarios, and response latency ranging from 5 to 10 seconds depending on the type of request.The results suggest that embedding governance and control mechanisms at the architectural level can improve transparency, traceability, and operational reliability of conversational agents in healthcare. The proposed approach highlights the importance of system-level design in supporting trustworthy AI beyond model-centric optimization.