Three vendors, three different products
"Agentic AI" has become the default label across procurement software, but three vendors that shipped agent products in the first half of 2026 built genuinely different things under that name.
Coupa launched Coupa Compose at its Inspire 2026 user conference on May 12, an "Agentic-as-a-Service" bundle for building and orchestrating a catalog of specialized AI agents inside its platform (Coupa, 2026). Ivalua launched IVA Studio on June 11, built around a single agent called IVA rather than a growing catalog (Ivalua, 2026). And on January 23, Zip was named a Visionary in Gartner's 2026 Magic Quadrant for Source-to-Pay Suites, the only vendor on that list built as an orchestration layer that can sit across other companies' procurement systems rather than as a suite of its own (Zip, 2026; Business Wire, 2026).
Three real product decisions, three different answers to the same question: what should an AI agent in procurement actually be.
What "agentic" is supposed to mean
The baseline distinction vendors are drawing is between AI that answers questions and AI that takes action. A chatbot that summarizes a contract is not agentic. A system that reads the contract, flags the clause that conflicts with a new supplier risk event, drafts an amendment and routes it for approval, is what vendors mean by the term.
Ivalua's press material describes this directly with a supplier risk scenario: an event triggers IVA to surface every affected contract, open a purchase order, draft a mitigation plan and propose qualified alternative suppliers, with a human still approving the outcome (Ivalua, 2026). Coupa's version ships as named, purpose-built agents, including ones for sourcing event creation, bid comparison, risk sentinel monitoring and sanctions screening, that a buyer assembles rather than a single generalist agent (Coupa, 2026).
Where the three bets diverge
Coupa's architecture is a marketplace: more than 20 persona-based agents that customers configure through a no-code interface, with the vendor explicitly betting that procurement work is varied enough to need many narrow agents rather than one broad one (Coupa, 2026).
Ivalua's architecture is the opposite bet. Its own press release frames the difference explicitly: users "do not need to build, orchestrate, or select from a growing catalog of agents." Instead one agent, IVA, has full access to the platform's data and creates temporary sub-agents only for complex, multi-step work, with permissions inherited from the human user it's acting for and every action logged (Ivalua, 2026).
Zip's bet is different again. Rather than compete as a full source-to-pay suite, it built an orchestration layer that connects into a company's existing procurement, finance and ERP systems, which is the basis Gartner cited for the Visionary placement (Zip, 2026). Zip states its platform has driven more than $6 billion in customer savings, a company-reported figure, not an independently audited one (Zip, 2026).
What to check before signing
The label "agentic" does not tell a buyer much on its own. A few questions do:
What can the agent actually execute, versus draft for approval? Ivalua and Coupa both keep a human-in-the-loop step for consequential actions like issuing a purchase order or approving a supplier change. Ask a vendor to show that step, not describe it.
Whose permissions does the agent run under? Ivalua's governance model explicitly ties IVA's permissions to the human user it's acting on behalf of, so it can't exceed what that person is already authorized to do (Ivalua, 2026). That's a concrete, checkable design decision. A vague answer about "role-based access" is not the same thing.
Does it require migrating off your current system, or sit on top of it? Zip's pitch depends on working across systems a company already runs. Coupa's and Ivalua's depend on the company running on their platform. That changes the implementation timeline and the exit cost if the agent underdelivers.
Is the model locked to the vendor? Ivalua describes IVA Studio as model-agnostic, letting customers use Ivalua's models or bring their own through Model Context Protocol connections (Ivalua, 2026). That's worth asking every vendor about directly, since the answer affects both cost and long-term flexibility.
Where it commonly breaks
The failure mode procurement teams report most often has nothing to do with the AI model. It's data. An agent that's supposed to draft a mitigation plan for a supplier risk event needs clean, current contract and supplier data to work from. Vendors pitching agentic AI are, in effect, also pitching a data cleanup project, whether they say so upfront or not. Buyers evaluating any of these three products, or others following the same pattern, should ask what happens when the underlying data is incomplete, not just what the agent does when it isn't.