The Bottleneck of Expert Knowledge Elicitation

Constructing Bayesian Networks (BNs) for operational decision support typically requires extensive, time-consuming interviews with domain experts to define causal relationships and conditional probability tables. This process is often the primary bottleneck in deploying robust decision-support systems. The authors propose a 'virtual survey' approach to mitigate this by leveraging Large Language Models (LLMs) to act as proxies for human experts, automating the initial stages of network structure and parameter estimation.

The Virtual Survey Methodology

Instead of relying solely on human availability, the framework utilizes LLMs to process domain-specific documentation and simulate expert reasoning. The process involves:

• Causal Discovery: Using LLMs to extract potential causal variables and relationships from unstructured operational data or technical reports.
• Expert Simulation: Prompting the LLM to provide probabilistic estimates based on specific operational scenarios, effectively conducting a 'virtual survey' that mimics how a human expert would quantify uncertainty.
• Network Refinement: Integrating these AI-generated structures with human-in-the-loop validation, ensuring that the resulting Bayesian Network remains grounded in real-world operational logic while benefiting from the speed of automated elicitation.

Operational Impact and Trade-offs

This approach shifts the role of the human expert from 'primary source' to 'validator.' By automating the heavy lifting of initial network construction, organizations can iterate on decision models much faster. However, the authors emphasize that the quality of the resulting network is highly dependent on the quality of the source data provided to the LLM. The primary trade-off is the risk of 'hallucinated' causal links; therefore, the framework mandates that AI-generated networks must undergo rigorous sensitivity analysis and human review before being deployed in high-stakes operational environments.