Generative AI

Most generative tools are built on large language models, and the capability differs in kind from the prior decade of procurement AI: earlier systems could classify a contract clause, while a generative model can rewrite it, compress 60 pages into one, or draft the first version of an RFQ from last year's event plus a new line-item list.

The honest limit is hallucination: the model produces plausible text unsupported by any source. Asked what a supplier's payment terms are, an ungrounded model answers confidently either way. The fix is architectural, not motivational: retrieve the actual documents, generate from them, and cite the source so a person can verify. This is why generation pairs naturally with extraction: structured data in, grounded text out.

Where it changes daily work is volume: summarizing every quote package rather than just the big ones, drafting every RFQ from precedent. Where it does not: the model holds no opinion on whether to single-source, and treating fluent text as analysis is how teams get burned. The shift is far enough along to be worth understanding where the tipping point actually sits. In LightSource, generated quote summaries stay linked to the underlying documents, so a buyer can verify any line in one click.

Frequently asked questions

What is generative AI in procurement?

Generative AI is the class of models that produce new content, such as text, tables, and images, rather than only scoring or classifying existing data. Procurement teams use it to draft RFQ documents, summarize long contracts and quote packages, and answer supplier or stakeholder questions from policy documents. Most generative tools are built on large language models.

How is generative AI different from earlier procurement AI?

Earlier procurement AI could classify things, such as labeling a contract clause, while a generative model can rewrite the clause, compress 60 pages into one, or draft the first version of an RFQ from last year's event plus a new line-item list. The capability differs in kind because generation produces new content instead of scores or categories, which is why it pairs naturally with extraction: structured data in, grounded text out.

What is AI hallucination and how do you prevent it?

Hallucination is when a generative model produces plausible text unsupported by any source; asked what a supplier's payment terms are, an ungrounded model answers confidently either way. The fix is architectural: retrieve the actual documents, generate from them, and cite the source so a person can verify. In one case, an ungrounded assistant claimed a supplier held AS9100 when the certificate library said otherwise, and the team's rule afterward was that every answer must quote the document it came from.

What should generative AI not be used for in procurement?

A generative model holds no opinion on whether to single-source, and treating fluent text as analysis is how teams get burned. Outputs read fluently whether or not they are correct, so grounding in source documents and human review remain part of any serious workflow. The gains sit in volume work, such as summarizing every quote package and drafting every RFQ from precedent, while judgment stays with people.