Procurement and supply chain are not exempt from the AI cost shock. Procurement is about cost management, and rising AI costs need to be professionally managed, too. Urgently!

Most loud AI billing comes from software engineering. It is tempting for procurement leaders to assume this is someone else’s problem. It is not.

The upside is real. Walmart’s work with the negotiation platform Pactum, as reported in a 2022 Harvard Business Review article and since expanded, is one of the better verified examples in procurement. Walmart used an AI chatbot to negotiate terms with tail-end suppliers, the long list too numerous for human buyers to handle individually. In the original Canadian pilot, the system closed agreements with 64 percent of the 100 suppliers invited, against a 20 percent target, and extended payment terms by 35 days. As the program expanded to the US, Chile, and South Africa, the close rate reached 68 percent. That is agentic AI doing in days what human buyers could not do at all. And costs have increased.

The risk side deserves equal attention. In January 2026, supplier data firm Apexanalytix published a report titled Procurement AI has a Hallucination Problem.

Its argument: most procurement organizations run AI on supplier data that is fragmented, duplicated, and often outdated. When AI reasons over records like this, it infers and guesses rather than stopping to ask. The output looks confident. It is not necessarily correct, and you paid lots of money for no or poor results!

That matters more in procurement than most back-office work. A sourcing agent recommending a supplier from stale financial data, or an onboarding agent approving a vendor against an outdated sanctions list, can create exposure that surfaces months later in a stockout or compliance finding.

Deloitte’s 2025 Global CPO Survey found 94 percent of procurement executives use generative AI weekly. Separate tracking puts real, governed, full-scale deployment at closer to 4 percent.

That gap is where the token bill and data risk first surface.

Do:

set a hard usage cap before the first agent goes live.

Clean supplier data first. Start with contained, low-risk spend.

Measure cost per task from week one.

Scope context deliberately. Keep a human in the loop for payment or contract terms.

Do not:

roll out agentic AI tenant-wide before a contained pilot.

Let leaderboards reward token volume over correct outcomes.

Point a frontier model at routine tasks a cheap model handles.

Trust a recommendation without checking the data behind it.

Assume a falling token price means a falling bill.

It’s about time that procurement organizations ramp up capabilities to manage AI costs and contracts. It’s a totally different animal than buying SaaS solutions

targetP evolving procurement. More than 35 digitalization projects in Procurement speak for themselves. So do decades of Procurement experience

The post Procurement AI Cost Shock appeared first on targetP.