AI in procurement: what works today and what is still hype

Every procurement software vendor now says "AI-powered". For a buyer, that phrase says almost nothing. What matters is which tasks a model can do reliably today, and which still need a person to decide.

Where AI already earns its keep

The best results come from work that is repetitive, text-heavy and easy to check.

Classification. Sorting thousands of invoice lines and purchase orders into a category tree (such as UNSPSC, NAICS or national codes) used to take analysts weeks. Language models now do a first pass in minutes, and people review only the low-confidence lines.
Extraction. Pulling dates, prices, payment terms and liability clauses out of contracts, tender documents and supplier certificates.
Search and summary. Asking "which of our contracts renew in the next six months and have no price cap?" instead of opening folders.
Drafting. First versions of clarification answers, supplier emails, evaluation reports and market-research notes.

In each case the output can be verified quickly. That is the test to apply to any AI proposal: how fast can a person check the answer?

Where it needs a human

Anything that commits money, or that a supplier could challenge, still needs an accountable person.

Award decisions. In public procurement, scoring must follow published criteria and be explainable to a bidder who lost. A model can prepare the evidence. It should not be the decision-maker.
Negotiation strategy. Models are good at summarising what is known about a market. They are poor at judging a supplier's real appetite for your business.
Anything with thin data. A model trained on general text knows nothing about your specific supplier's delivery record unless you give it that data.

The rules differ by market

There is no single global rulebook. The EU AI Act sorts AI systems by risk and is phasing in over several years. In the US, requirements come from a mix of federal guidance, agency rules and state laws. China has specific rules for generative AI services, and several Gulf states and Asian countries have national AI strategies with their own data and governance requirements. Most procurement tools, such as classification and search, are low risk everywhere. But a system that evaluates people or makes decisions with significant effects on them can attract stricter rules. Wherever you operate, ask every vendor two questions: what data does the model use, and what does it log about each decision? If they cannot answer clearly, treat that as a warning sign.

A simple way to start

Pick one painful, checkable task. Spend classification is a common first choice.
Run the AI alongside the current process for a month and compare results.
Measure three numbers: time saved, error rate, and how often a human overrides the model.
Write down the rule for when a human must review. Then scale up.

What to ask a vendor

Where does our data go, and is it used to train your models?
Can we see why the system gave a particular answer?
What happens when the model is unsure?
Can we export all our data if we leave?

The bottom line

AI is already a good assistant for procurement and a poor replacement for it. Teams that treat it as a fast, tireless junior analyst, one whose work is always reviewed, get real savings. Teams that treat it as an oracle get confident mistakes. Start small, measure honestly, and keep a person accountable for every decision that matters.