Artificial intelligence (AI)
Artificial intelligence (AI) is software performing tasks that historically required human judgment: reading documents, classifying spend, predicting prices, drafting text, or executing multi-step work. In procurement the label matters less than the breakdown: classification, extraction, prediction, generation, and agents. A system marketed as AI may be any of these, or just a rules engine, so the working question is which task it performs and how its accuracy is measured.
Examples
Extraction: Six suppliers quote a 120-line RFQ in six different PDF formats. An extraction model converts all six into one normalized table in under ten minutes, and the buyer spot-checks the 14 lines flagged as low confidence instead of retyping 720 prices.
Prediction: A model trained on three years of resin purchases expects a molded part at $2.10 to $2.25 given current resin indices. A renewal quote arrives at $2.48, gets flagged, and the buyer finds an outdated material surcharge baked in.
The rules-engine test: A tool advertised as AI classifies invoices by matching supplier names against a fixed list. It never improves and breaks on every new supplier. Useful software, but a directory, not AI.
Definition
AI is an umbrella, and in procurement it covers five working capabilities. Classification sorts things into categories, like coding 50,000 PO lines to a taxonomy. Extraction pulls structured fields out of documents, like line items from a quote PDF. Prediction estimates what comes next, the basis of predictive procurement. Generation drafts new text, the domain of generative AI. And agents chain these together into multi-step work like chasing down missing quotes.
Most of these run on machine learning and language models rather than hand-written rules, which gives you an honest test for the marketing version of the term: does the system improve as it sees more data, and does it handle inputs it was never explicitly programmed for? A lookup table with a chat window fails that test.
The practical posture: AI output is probabilistic. A system that extracts quote lines at 97% accuracy is enormously useful and still wrong on three lines in a hundred, so workflows need review steps wherever errors are expensive.
The right evaluation for any tool is which stage of the sourcing lifecycle it serves and what it actually does there, which is worth mapping stage by stage. LightSource, an AI-native procurement platform for direct materials, applies extraction and anomaly detection to supplier quotes so buyers review exceptions instead of retyping PDFs.
Frequently asked questions
What is artificial intelligence in procurement?
In procurement, artificial intelligence is software performing tasks that historically required human judgment: reading documents, classifying spend, predicting prices, drafting text, or executing multi-step work. The label matters less than the breakdown into five working capabilities: classification, extraction, prediction, generation, and agents.
What are the main types of AI used in procurement?
AI in procurement breaks down into five working capabilities. Classification sorts things into categories, like coding 50,000 PO lines to a taxonomy; extraction pulls structured fields out of documents, like line items from a quote PDF; prediction estimates what comes next; generation drafts new text; and agents chain these together into multi-step work like chasing down missing quotes.
How can you tell real AI from a rules engine?
An honest test separates real AI from a rules engine: does the system improve as it sees more data, and does it handle inputs it was never explicitly programmed for? A lookup table with a chat window fails that test. A tool that classifies invoices by matching supplier names against a fixed list never improves and breaks on every new supplier, which makes it a directory rather than AI.
How accurate is AI in procurement workflows?
AI output is probabilistic rather than guaranteed. A system that extracts quote lines at 97% accuracy is enormously useful and still wrong on three lines in a hundred, so workflows need review steps wherever errors are expensive. In practice that means buyers spot-check the lines flagged as low confidence instead of retyping entire documents.
How should you evaluate an AI procurement tool?
The right evaluation for any tool marketed as AI is which stage of the sourcing lifecycle it serves and what it actually does there. A system sold as AI may be classification, extraction, prediction, generation, or agents, or just a rules engine, so the working questions are which task it performs and how its accuracy is measured.
Previous