Walmart’s procurement team now runs part of its supplier negotiations through an AI agent. Not a chatbot that drafts a counteroffer for a human to send on — an agent that actually conducts the back-and-forth. Pactum, the vendor behind it, says its agents have handled more than $1 billion in negotiated spend for Walmart, landing an average 3% cost improvement and talking 68% of suppliers into accepting a 35-day payment-term extension. Gartner thinks this is just the opening act: it predicts that by 2030, half of all supply chain management software will run on agentic AI that makes and carries out decisions across the business without a human directing each one.
That’s a sweeping claim dressed in a boring word — “procurement.” But procurement, sourcing, and contract management are exactly where enterprise software is testing agentic AI first, because the decisions are repetitive, well-documented, and easy to measure against a before-and-after baseline. If you want a preview of what AI agents in business software will actually look like once the hype settles, this is the place to look.
What “autonomous negotiation” really means
Despite the marketing, none of today’s procurement agents are freelancing on six-figure supplier contracts. Keelvar’s sourcing agent reportedly automates 85% of sourcing events across its customer base and claims 8% to 12% in extra savings per negotiation round — but it’s working within sourcing events a human has already scoped. Fairmarkit’s agents report an 86% cycle-time reduction and triple the sourcing events per buyer, largely by clearing out tail spend: the low-value, high-volume purchases nobody wants to spend a week negotiating by email.
The pattern holds across the category: agents handle bounded, repeatable deals — renewals, tail spend, sole-source buys within a preset price band — and humans sign off on anything novel, high-stakes, or involving a brand-new supplier. Pactum’s own case studies describe three separate negotiation agents tuned for exactly those three tiers. “Autonomous” here means autonomous within a fence, not autonomous full stop.
The procurement stack is being rebuilt, agent by agent
What makes 2026 different from the pilot-everything era of 2023-24 is that this is now showing up inside the systems analysts already use, not just in standalone point solutions. SAP is rolling eleven named Joule agents into Ariba and Fieldglass this year on a staged schedule:
- A conversational intake agent in Ariba Intake Management (general availability June 2026) that routes requests via chat, email, or Teams and checks policy compliance before a buyer ever sees the request.
- AI-assisted contract drafting in Ariba Contracts, turning natural-language prompts into draft terms linked to downstream pricing and key dates.
- SOW-deliverable drafting in Fieldglass Services Procurement, already live since Q1 2026, with SAP citing a 70% cut in manual drafting time.
- Gate-pass time capture for Fieldglass in asset-heavy industries such as oil, gas, mining, and utilities, planned for June 2026.
Five of the eleven agents land in June, the rest in September, and SAP is giving the Joule Agent Runtime away free through the end of 2026 — a strong signal it wants adoption numbers, not a slow rollout. Rival suites are moving the same direction: Coupa now offers more than twenty named Navi agents built on what it describes as $10 trillion in spend data, and Ivalua has over thirty embedded agents with inherited user permissions. The common thread is that the agent layer is being bolted directly onto the systems of record — the ERP and source-to-pay platforms — rather than sold as a separate AI add-on.
Where the guardrails sit — and why that’s the real story
Kaitlynn Sommers, the Gartner analyst behind the 2030 forecast, frames agentic AI as a genuine break from robotic process automation: RPA follows a script, while an agent completes a task without a fully predefined outcome, adapting as conditions change. That’s exactly why Gartner’s own advice to supply chain leaders is less about turning agents loose and more about containing them — start with narrow use cases that have clean data and verifiable behavior, and set explicit operational parameters so an agent can’t wander outside its scope.
That caution is the part worth taking seriously, because the economics are already real. If an agent can reliably take 86% of the grind out of tail-spend sourcing or cut a six-month RFP down to 27 days — both figures vendors report in production, not in demos — the question for most businesses isn’t whether to adopt this, but how narrowly to scope it on day one.
The takeaway
Agentic AI’s first real foothold in enterprise software isn’t a flashy copilot bolted onto a dashboard — it’s a quiet negotiator working tail spend and renewals inside the ERP you already run. Before rolling anything like this out, map your own spend into the same two buckets vendors already use: bounded, repeatable, low-risk deals an agent can run end to end, and novel or high-stakes contracts that stay with a human. Start the pilot in the first bucket, measure it against last year’s cycle times and savings, and only widen the fence once the agent has earned it.
