The canonical procurement problem NLP solves: HEX BOLT M8X40 ZN and Bolt, hex head, M8 x 40 mm, zinc plated are the same part to a human and different strings to a database. Matching, extraction, and classification close that gap, which is why NLP sits underneath spend classification, supplier-record deduplication, and contract clause work. Historically each task needed its own statistical model or rule set, one extractor for dates, another for payment terms. Large language models now cover many of these tasks with a single general model, though purpose-trained classifiers survive where volume and cost dominate, like coding millions of invoice lines.

What separates working deployments from demos is measurement: precision and recall per task, on your own documents. Quote PDFs from a 40-person machine shop look nothing like clean sample contracts, and messy real-world documents are where accuracy claims collapse. Good practice is to hold out a labeled test set, measure against it, and route low-confidence outputs to human review rather than forcing an answer.