Your customer's AI will send you a purchase order. Can you read it?

Key Takeaways

• Payment and agent protocols, including AP2, MCP, and UCP, are converging on a shared stack for agentic commerce, but it's built mostly for consumer checkout, not B2B contract pricing.
• Gartner expects more than 40% of agentic AI projects to be canceled by the end of 2027, citing cost, unclear value, and weak risk controls, even as procurement teams keep piloting the technology.
• Industry research puts procurement at a small share of current enterprise AI use cases, well behind sales and customer service.
• OpenAI's in-chat checkout product was shut down within months of launch, while the underlying protocol work continued, a pattern worth remembering before betting on any one standard.
• The practical 2026 reality is assistive: agents extract, route, and draft, and a person still approves before an order posts.
• The readiness test that actually matters is not which protocol you support. It's whether you can ingest an arbitrary machine-written document and validate it against your own rules.

Agentic commerce moved from conference demo to procurement roadmap this year. Payment and agent protocols are converging, and the vendor conversation has shifted from whether AI agents will buy things to when they will. Gartner's answer to that enthusiasm is blunt: it expects more than 40% of agentic AI projects to be canceled by the end of 2027, citing escalating costs, unclear business value, and inadequate risk controls.

Both things are true at once, and that's actually good news for your order desk. You don't need to bet on a winning protocol or build an autonomous purchasing pipeline this quarter. What you need to know is what actually shows up in your inbox when a customer's procurement software starts acting on its own, and it's rarely the sleek API call the demos show.

Agentic commerce, defined without the hype

Agentic commerce runs on a handful of general standards, each owning a different layer of the stack. MCP, which Anthropic open-sourced in November 2024, lets an AI agent connect to tools and data rather than relying on static prompts alone. AP2, which Google introduced in September 2025 with more than 60 partners including Mastercard, PayPal, and Salesforce, adds cryptographically signed "Mandates" so a merchant can verify what a user actually authorized an agent to spend.

UCP, the Universal Commerce Protocol backed by Google, Shopify, Target, and Walmart, among others, covers product discovery and cart building, while OpenAI's ACP handles checkout within a chat interface. None of these protocols compete so much as stack: an agent might use MCP to read a catalog, UCP to build a cart, and AP2 to prove the payment was authorized. That's real infrastructure, built by real companies, with real specifications you can read today.

What's shipping versus what's forecast

What's shipping: the protocols themselves exist, have working implementations, and have signed on major payment networks and retailers. MCP has thousands of live server implementations. UCP became self-serve for any developer in June 2026. What's forecast: the scale everyone quotes. One widely cited Gartner projection puts 90% of B2B buying under AI agent intermediation by 2028, moving more than $15 trillion in spend. That's a forecast for 2028, not a description of today.

The gap between the two shows up in adoption research. Industry surveys tracking enterprise AI use cases consistently place procurement well behind sales and customer service in actual deployment, even though buyer-side interest is high. Gartner has also warned about "agent washing," where vendors rebrand existing chatbots and automation tools as agentic without the underlying capability, and estimates that only a small fraction of self-described agentic AI vendors have the real thing. Meanwhile, OpenAI shut down its flagship in-chat checkout product within months of launch, even as the underlying protocol work continued. Real infrastructure and real hype are shipping side by side, and telling them apart is the actual skill right now.

Why the first wave looks like weird PDFs, not API calls

The part that matters most for a distributor: even where these protocols work well, they're designed for consumer-style checkout, carts, cards, stablecoins, not for negotiated contract pricing, tiered approval chains, and net-30 terms. Nobody has built a mature B2B version of "add to cart" that understands your specific pricing agreement with a specific customer.

So when a customer's procurement agent decides to reorder from you, and you haven't built a UCP endpoint or set up EDI with that customer, the agent still has to get the order to you somehow. The most likely path is the same one a human would take if a portal or API weren't available: it generates a document and sends it, usually as a PDF or an email, sometimes through a customer portal built for people, not agents. That document may format line items in a way your team hasn't seen before, reference SKUs slightly differently than your catalog does, or split a single order across an unusual structure. It's still a purchase order. It just wasn't written by a person.

The readiness checklist that actually matters

The mistake is preparing for agentic commerce by picking a protocol to support. The more useful preparation has nothing to do with which standard wins. It's four capabilities that work regardless of who, or what, generated the document in front of you.

Any format. Your intake process needs to read a PDF, an email body, a portal upload, or a structured feed without caring which it is or who produced it - via business-rule validation. Every incoming order gets checked against your actual pricing, availability, and customer terms, whether a person typed it or software generated it.

Audit trail. You can show exactly what arrived, what was validated, what changed, and when, which matters more once you can't assume a human read the order before it landed in your queue.

Human approval on exceptions. Anything that doesn't match your rules routes to a person instead of posting automatically. That's the boring, unglamorous version of readiness, and it's the version that actually works in 2026.

Where to start this week

• Pull a sample of your inbound orders from last month and check how many arrived as unstructured PDFs or emails rather than via EDI or portal submissions.
• Confirm that your order intake can already handle arbitrary document formats without requiring a person to retype them.
• Set a validation tolerance for pricing and availability that applies consistently regardless of who submitted the order.
• Check whether your current process logs what was received and what was changed, in case a customer disputes an order later.
• Identify which order types are safe to auto-post and which should always be routed to a person, before volume forces that decision under pressure.