AI: A straight A student who knows the model answer but cannot read the room.

I said I would share AI limitations and use in my adaptation to hospital and equipment planning. This one is about a strength that becomes a shortcoming.

When you give an LLM a source document, it knows to prioritise what is in it. A notable tool for this is NotebookLM by Google, which uses RAG (Retrieval Augmented Generation) to fish out relevant content from a huge amount of text by mathematical calculation. It answers only from the documents you upload.

I once used it to summarise service descriptions from a department functional brief.

Clean, controlled, no hallucination.

But there were one or two services with no written description in the documents, simple things like who the housekeeping department serves.

The easy answer is the every department in the hospital, as everywhere needs to be cleaned regularly. NotebookLM kept rearranging the words it had. It could not fill in what was not written.

A generic LLM answered that instantly.

The strength of staying within the document became a limitation the moment the document had a gap.

There is no perfect tool. It is accurate but limited. A generic LLM is broader but harder to trust. Knowing which one to use and when is the real skill.

I use NotebookLM, Claude chatbot and custom built AI agents depending on the use case.

After all, I found NotebookLM or other RAG tool a good office admin companion as there are generally numerous manuals and documents already existed but no one had time to search for a needle in a sea.

I am also testing how to use RAG to accurately extract compliance requirements from numerous regulations documents. It is OK but not passable to my standard. I will share when I figure it out.