Large language model (LLM)

Before LLMs, software computed on structured fields: part number, quantity, price. Everything else, the reasoning in an email thread, the conditions buried in a quote's notes, the indexation clause on page 31, was unstructured data that systems could store but not understand. LLMs read it, which moves the bulk of procurement's information inside the boundary of what software can work with.

Practically, LLMs are the engine behind most current generative AI, have replaced many task-specific NLP pipelines with one general model, and become agents when given tools and goals. For a buyer, the visible difference is asking questions in plain language and getting answers drawn from actual documents.

Working limits: outputs are probabilistic, so extracted data must be validated against the source, ideally with confidence scores routing uncertain lines to review. Long documents get processed in sections. And capability scales with cost, so well-built systems send simple tasks to small, cheap models and reserve large ones for hard reading. None of this is exotic; it is the same accuracy engineering procurement already applies to any data pipeline.

Frequently asked questions

What is a large language model?

A large language model (LLM) is a neural network trained on massive volumes of text that can read, interpret, and generate language. LLMs are the engine behind most current generative AI, have replaced many task-specific NLP pipelines with one general model, and become agents when given tools and goals.

Why do LLMs matter for procurement?

LLMs matter to procurement because most of the function's knowledge, including quotes, contracts, emails, specifications, and negotiation notes, exists as language rather than database rows. Earlier software computed only on structured fields like part number, quantity, and price, while an LLM can read the reasoning in an email thread, the conditions in a quote's notes, or the indexation clause on page 31, making that material searchable and comparable for the first time.

What are the working limits of LLMs?

LLM outputs are probabilistic, so extracted data must be validated against the source, ideally with confidence scores routing uncertain lines to review. Long documents get processed in sections, and capability scales with cost, so well-built systems send simple tasks to small, cheap models and reserve large ones for hard reading. These are the same accuracy engineering practices procurement already applies to any data pipeline.

What can an LLM do with procurement documents in practice?

In practice an LLM can extract both a quote's line items and the note that changes the price, such as a 30-day validity window with a 4% adder below 5,000 units, where a template parser would capture the table and miss the note. Portfolio questions also become possible: asked which of 85 supply agreements include raw material pass-through, a model flagged 12 with clause citations, review confirmed 11, and the one false positive cost two minutes to dismiss.