AI in procurement: what works in 2026 and what does not

Published 26 August 2026 · 6 min read

AI in procurement covers a lot of ground, and the gap between the demo and the deployment varies enormously by use case. This is a sorted list: what is working in production today, what is promising but immature, and what is still a slide.

Working now

Spend classification and supplier normalisation.The clearest win. It is a language task on messy text, which is precisely what large language models are good at, and the output is checkable — you can look at a category and know whether it is right. See how it works.
Contract data extraction. Pulling renewal dates, notice periods, price-review clauses and liability caps out of PDFs. Accuracy is high enough to be useful, and errors are cheap because a human confirms before acting.
Drafting. RFQ documents, supplier emails, scoring rubrics, internal summaries. Modest but immediate time savings.

Promising, not yet dependable

Demand forecasting. Works where you have long, clean history and stable demand. Most procurement data is neither.
Supplier risk monitoring. Good at surfacing news and filings, weak at judging materiality. Expect a lot of noise per genuine alert.
Should-cost modelling. Plausible outputs, hard to validate. Useful as a challenge in negotiation, dangerous as a number you commit to.

Still a slide

Autonomous negotiation. Demos well on simple repeat-buy categories. Real negotiations involve relationships, leverage and context no model has.
End-to-end “autonomous procurement”. A category strategy is a set of commercial judgements about risk and relationships. Prediction is not judgement.

Why classification comes first

Nearly every other application depends on it. You cannot forecast demand by category, monitor risk by category, or model should-cost without categorised spend. Teams that buy an AI analytics platform before classifying their data end up with an expensive dashboard over an uncategorised transaction log.

It is also the use case where the technology is most mature and the output easiest to check — which makes it the sensible place to find out whether these tools work for you at all.

How to evaluate any procurement AI tool

Ask what happens when it is wrong. A tool with a review step and retained corrections is built by people who have deployed one. A tool with no error path is not.
Test on your worst data, not your best. Everything works on the top 100 suppliers. Ask it to handle the tail.
Check reproducibility. Run the same input twice. Different output means you cannot compare periods.
Establish where your data goes. Which sub-processors, what retention, whether it trains models. See the security checklist.
Insist on a trial with your own file. Vendor demo data is chosen to work.

A reasonable expectation

AI will not replace a procurement team. It removes the weeks of manual categorisation that stop the team doing the work only they can do — understanding the market, and negotiating. That is a smaller claim than most vendors make and a much more reliable one.