病院のAI導入はなぜ止まるのか｜RAG・NotebookLM以前に必要な「ルールの明文化」｜規程・ローカルルール・例外処理をどうAI化するか

This document outlines the problems encountered when using RAG or NotebookLM in healthcare settings to enable AI-powered searching and answering of internal regulations, manuals, and application forms. While it may seem convenient at first glance, actual hospital rules are not that simple. Official rules documented as regulations, local rules specific to each department, and exceptions such as nighttime, emergency, and staffing shortages. If these three layers are not properly organized before feeding the AI ​​with the information, it becomes ambiguous—even before considering the AI's performance—as to determine which document is the latest version, who approved it, and who makes decisions regarding exceptions. The most dangerous aspect isn't just AI hallucination. The AI ​​might pick up outdated but genuine documents as evidence. This video explains the concepts of document management, metadata, document ownership, exception handling, and periodic testing necessary before implementing AI-powered internal rules in hospitals and general companies. Instead of aiming for a "decision-making AI" from the start, begin with a "reception AI" that indicates the location of documents, submission destinations, deadlines, and supporting regulations. What's needed for that isn't a perfect set of rules, but rather a clear documentation of principles, standard procedures, and exception handling.