AI agents in procurement: why most pilots stall, and what gets them into production

The appetite is real

Procurement leaders are not short of intent. The Hackett Group's 2025 Key Issues study found that 49% of procurement teams had piloted AI use cases, yet only 4% were running them at large scale. BCG's 2026 research across 13 business functions put procurement's AI adoption at 35%, the lowest of the functions studied, behind supply chain and sales at 44% and finance at 40%.

The pressure to close that gap keeps rising. Hackett projects procurement workload growing by about 10% a year against budget growth of about 1%. No team can hire its way through that difference.

Why pilots stall

The data is not ready. An agent that answers supplier questions or builds a sourcing event is only as good as the supplier master, the category structure and the event history behind it. Duplicate vendors, missing tax identifiers and free-text item descriptions do more damage to an AI pilot than any weakness in the model.

The pilot sits outside the process. Suplari's 2026 benchmark of 121 procurement teams found 90% of professionals relying on general-purpose AI tools and only 8% using AI embedded in their procurement systems. If the output of an assistant has to be retyped into the sourcing tool, the time saved in drafting is lost again in data entry.

Nobody owns the decision. The same benchmark found only 17% of teams with enforced AI policies, and Hackett reports that 59% of procurement teams lack the expertise to manage AI deployments. When it is unclear who approves what an agent recommends, the pilot stays a pilot.

Where agents earn their place first

The tasks that reach production soonest share three traits: high volume, clear rules and an outcome a person can check quickly. In procurement that points to four areas.

• Supplier queries. Registration status, payment status and how to respond to an event are the repeated questions a supplier helpdesk answers every day.
• Sourcing event administration. Building the event from a request, inviting suppliers, chasing responses and preparing the comparative statement.
• Supplier data validation. Checking statutory details against official sources at the point of registration, before a bad record reaches the vendor master.
• Approval paperwork. Drafting the note for approval from data the event already holds.

The platform vendors are moving the same way. SAP made Joule agents for intake management and contracts in SAP Ariba generally available in June 2026. As more of this capability arrives inside the platform, the customer's job is to be ready to use it.

Four things to do before the next pilot

1. Pick a task, not a technology. Choose one with volume, clear rules and a baseline you can measure in hours, cycle time or error rate.
2. Fix the data that task depends on. This is usually the supplier master and the category structure, and it is worth doing whether or not the pilot succeeds.
3. Put the agent inside the workflow. It should read from and write to the procurement system directly, with a person approving anything commercial and every action logged.
4. Define production on day one. Agree who owns the agent, what it may do alone and when it must hand over to a person.

The takeaway

The procurement teams getting value from AI are treating it as a process change that happens to use a model. They start narrow, clean the data underneath and keep people in charge of the decisions that matter. That is slower than a demo and far faster than a pilot that never ends.