Most agentic-commerce commentary obsesses over the bookends: the AI that finds your product, and the one tap that buys it. A recent Nexi Group podcast — "From clicks to conversations: how AI is changing commerce," featuring Nexi's e-commerce EVP Raja Saggi, Google Cloud's Cristina Conti, and Shopware's Michael Pfeiffer — put the spotlight where it actually belongs: the part in between.

Saggi's framing is one of the more useful things said about agentic commerce this year: "It's the middle bit — after they've browsed for something, identified it, and then they want to check it out. That's where… trust, verification, liability — that is the most complex bit."

Take it seriously, because it comes from a payments company, not a model vendor. Discovery is a search problem (mostly solved). Checkout is a UX problem (mostly solved). The messy middle — proving the agent is allowed to spend your money, that it's a real agent and not a fraud bot, and who eats the loss when it goes wrong — is a payments and trust problem. That unglamorous layer is what decides whether any of this ships. Here's what Nexi is actually saying, and where our own data says the middle breaks. Our bias throughout: we saw it in the data first.

#What Nexi is actually saying

Strip the episode to its load-bearing claim and it's this: agentic commerce is a payments problem wearing an AI costume. Nexi doesn't position itself as an AI company — it positions the payment rails as the thing that makes autonomous checkout safe. Saggi's "crawl, walk, run" is explicitly a paytech roadmap: start with an MCP server on the payment gateway exposing a conversational interface internally, move to conversational pay-by-link, and only then to full autonomy with mandates and liability frameworks.

The tell is what he tells merchants to do today: "Merchants should already invest in being discoverable… They should not wait." A payments exec telling retailers their homework is discoverability and being machine-readable — not "wait for the standards to settle" — is the same message we make with data. The rails are coming; the merchants who are agent-legible before the middle is solved are the ones who convert when it is.

#The messy middle, defined

Why is the middle the hard part when the bookends aren't?

• Discovery is a retrieval problem. Agents already search catalogs; the open question is whether your catalog is parseable (often it isn't — but that's a fixable data problem).
• Checkout is a UX problem. One tap, or one complete call. Mechanically simple.
• The middle is a trust-and-money problem with no clean owner. Three unsolved questions live here:
  1. Authorization — did the human actually approve this purchase, at this price, within this budget?
  2. Identity — is the thing calling your checkout a legitimate agent, or a bot committing fraud?
  3. Liability — when an agent buys the wrong thing, or a fraudulent one does, who is on the hook?

None of those are AI problems. They're payments problems — which is precisely why a company like Nexi, not a model lab, is the one with a credible plan for them.

#The trust stack being built for the middle

The podcast laid out the emerging machinery for each gap:

• Intent mandates for authorization. Saggi: "for the first time you have… digitally signed cryptographic evidence that this agent had this level of mandate or approval." This is the AP2 idea — a signed, auditable record that the human consented to a bounded action (budget, category, scope).
• Know Your Agent (KYA) for identity. Saggi puts it bluntly: "How does the merchant know that the agents that engage with it… are in fact real agents and not bots… trying to perpetrate fraud?" Visa and Mastercard are building agent-validation services to answer exactly that.
• Human-in-the-loop and merchant guardrails for liability containment. Conti: "the human in the loop… remains critical at so many steps." Pfeiffer: "we want to ensure the merchant is in control… defining the guard rails within which the agents can behave autonomously."

Notice the pattern: the entire stack exists to make the middle safe enough to automate. The bookends needed none of it.

#Why Europe is a year behind

The most concrete prediction in the episode is also the most Europe-relevant. Saggi: "In the US you will see a completely autonomous transaction happening this year. In the EU… you will have to wait until next year."

That gap isn't timidity — it's structure. European payments carry strong-customer-authentication and multi-party settlement requirements a US card flow doesn't. An EU autonomous transaction has to reconcile an agent's mandate with SCA, with the merchant's acquirer, with cross-border settlement — more parties in the middle, each needing to trust the others. The messy middle is simply messier in Europe. For European merchants that's not a reason to wait; it's a reason to finish the parts you control — discoverability, a clean machine-readable catalog, conformant endpoints — now, so you're ready when the trust layer lands.

#Where the protocols fit

The middle is also why three protocols keep coming up instead of one — AP2, UCP and MCP do different jobs, and the messy middle needs all three. AP2 carries the mandate and credentials (the authorization layer). UCP is the commerce contract — catalog, cart, checkout (the rails the transaction runs on). MCP is the agent-to-tool plumbing underneath. (If that stack is fuzzy, we break down how MCP, UCP and AP2 fit together here.) Nexi's own pilots start at the MCP layer on the gateway and build up — which is exactly the order the stack composes in.

#What it means — and where the middle actually breaks

Here's where we can put data behind Nexi's thesis: we've watched agents fail in the middle, repeatedly.

• Checkout completion is the real bottleneck. Getting an agent from "found it" to "bought it" is where conversion leaks — our readiness data across thousands of merchants shows the gap between stores that declare a checkout capability and stores an agent can actually complete a purchase against.
• Variant mismatches are the single biggest checkout failure mode we see — the agent selects an option that never resolves to a buyable SKU. That's a middle failure: discovery worked, the buy didn't.
• Time-sensitive inventory breaks agents outright — we ran five frontier models against expiring offers and none respected an offer's clock. A signed mandate means nothing if the thing it authorized expired mid-session.

So the takeaways, by audience:

• Merchants: do exactly what Saggi said — be discoverable and machine-readable now. The trust layer is someone else's job; a clean, conformant, parseable store is yours, and it's the prerequisite for everything in the middle. Check whether an agent can actually transact with yours.
• Paytechs and platforms: the middle is the land grab. Whoever owns mandate verification, KYA and liability for agentic payments owns the most defensible layer in the stack.
• Everyone: stop demoing the bookends. Browse-and-buy is the easy demo; the trust-and-liability middle is the actual product, and it's mostly unbuilt.

#What we're watching

Whether AP2 mandates and KYA converge into something interoperable or fragment per vendor; whether the EU's later timeline holds or slips on SCA and settlement complexity; and — the part we measure — whether agents stop failing in the middle (variant resolution, offer expiry, checkout completion). We track the ecosystem monthly in the State of Agentic Commerce census, and we report what the data shows.

#Sources

• Nexi Group — From clicks to conversations: how AI is changing commerce (Raja Saggi, Nexi; Cristina Conti, Google Cloud; Michael Pfeiffer, Agentic Commerce Alliance / Shopware)
• Our analysis — MCP vs UCP vs AP2 and Agentic Commerce Optimization

#About UCP Checker

UCP Checker is the independent validation and monitoring layer for the Universal Commerce Protocol. We crawl, validate and grade every public UCP manifest we can find, run the merchant directory, the UCP Score and live adoption stats, and test how real AI agents behave against real stores — including where they break in the messy middle.

• Check your store: ucpchecker.com/check
• Grade it (UCP Score): ucpchecker.com/score
• See capability & payment coverage: ucpchecker.com/capabilities
• Browse the directory: ucpchecker.com/directory

Discovery and checkout get the demos. The middle gets the merchants who actually convert. We'll be measuring it.