From Unstructured Document to Audited Checklist

The work

A workplace safety audit arrives as a dense PDF, and someone has to turn it into who needs which training, by when. The risk is not the reading — it is the inventing. An automation that hands back a finding the source never contained is worse than no automation at all. So the brief set itself: extract only what the document actually says, prove it, and leave the call to a person.

How it works

• The language model extracts each finding by quoting the source word for word — it does not paraphrase, and it does not invent findings.
• Every quote is then checked back against the document: if the cited sentence isn’t really there, it gets flagged. That check is the anti-hallucination lock — a fabricated finding can’t reach a decision.
• Deterministic rules map each verified finding to the right training and a deadline by severity. Same input, same output, every time.
• For anything the rules don’t cover, a RAG layer — retrieval-augmented generation, semantic search over the course catalogue — proposes the closest match. It never decides on its own: it proposes, and the case goes to human review.
• Whatever stays uncertain lands in a review step, for a person to approve before anything is final.

The principle holds across the whole system: the AI quotes, the rules decide, the person approves.

Built to run anywhere

The same workflow, built three ways:

• n8n + Python — n8n orchestrates the flow; Python carries the control logic and the citation check.
• Native n8n — the whole thing inside n8n, no Python, including the RAG agent that retrieves from the catalogue by semantic search.
• Power Platform — the same result rebuilt on Microsoft’s cloud, on an environment set up from scratch.

Different tools, the same result. The judgment was never in the tool.