Somewhere in a London merchant's integration backlog sits a cost line reading anywhere between $0 and $500,000. Same job. Same deliverable. Four orders of magnitude of disagreement. The task — "make your checkout accept payments initiated by an AI agent" — is quoted at effectively zero for a Shopify store, $5,000 to $50,000 for a Stripe shop needing custom work, and up to half a million dollars for an enterprise-built stack. Not one of those figures carries a source citation.
The missing citations are the most honest thing about the entire agentic commerce conversation. Over the past two quarters, 89% of surveyed US and UK merchants said they were preparing for agent-driven payments; 42% said they were already testing. Actual AI-agent transaction share across both markets: 3%.
Thirty to one. That is the ratio between preparation and penetration, and it is the widest narrative-versus-fundamentals divergence I have charted since the yield farms of 2020.
Strip away the framing and here is what happened. AI shopping agents must prove two things to a merchant: that they are a legitimate agent, and that a specific human authorized this specific purchase. Card networks cannot express that with existing rails. A stored card number authorizes everything or nothing. An agent needs to authorize narrowly — this merchant, this amount, this window, this item.
So three frameworks appeared, each solving a different slice of the problem.
Visa's Trusted Agent Protocol performs HTTP-layer signature verification. The merchant fetches public keys from a Visa directory to confirm a request genuinely originated from a registered agent. Mastercard and Google built Verifiable Intent on SD-JWT delegation chains carrying eight types of machine-verifiable constraint — spend ceilings, merchant whitelists, expiry windows. Vouched shipped KYA-OS as a REST API that runs alongside existing fraud infrastructure, then donated the specification to the Decentralized Identity Foundation.
Around them: Ant International on the Asia and cross-border flank, Cloudflare arriving as automation middleware attached to Mastercard, and at least eight competing proposals in the field — UCP, ACP, AP2 among them.
Ant International's presence adds a geopolitical layer the source material flattens. A cross-border Asian payment incumbent entering a US-and-EU identity standard debate inherits the same scrutiny that has shadowed Chinese-adjacent fintech for years. Its regional strength is real; its ability to set global agent-identity norms is constrained by regulators who read the participant list before the specification.
The reason this keeps getting filed under blockchain is thin. Vouched's spec went to DIF, and Coinbase's x402 appears on the protocol list. That is the entire Web3 surface area. And x402 — the one genuinely native settlement path — receives almost no analysis. Code speaks, but culture listens, and right now the culture is listening to the card networks.
Let me be precise, because the imprecision here is load-bearing.
Visa, Mastercard, and Vouched do not compete in the same category. Visa operates a signature verification layer. Mastercard and Google operate an authorization delegation layer. Vouched operates a fraud-bypass integration layer. These stack. They are not mutually exclusive; a merchant could run all three tomorrow.
Calling them "three frameworks fighting for the merchant verification layer" is a category error with commercial consequences. It manufactures the appearance of a winner-take-all race where the architecture permits layering — which means the fragmentation argument ("too many protocols, too much integration cost") and the competition argument ("pick a winner") cannot both hold. The source material uses them interchangeably.
The technical reality is less dramatic than the framing. HTTP signatures are decades old. SD-JWT is a mature format. Nothing here is cryptographically novel. That tells you where competition actually lives: not in mathematics, but in who owns the trust root and who absorbs the integration labor.
Walk through the delegation chain once, because the mechanics matter. A consumer instructs an agent: buy this, up to this amount, from these merchants, before this date. That instruction is encoded as a set of disclosed claims inside an SD-JWT — selectively disclosable, meaning the merchant sees only the constraints relevant to the transaction and never the underlying credential. The issuer signs. The agent presents. The merchant verifies against a published key set. No card number travels. The cryptography is not the hard part; the selective-disclosure UX and the revocation story are.
Look at the anchors. Visa's directory is a single point of trust — merchant security is only as good as Visa's key distribution. Mastercard and Google occupy a semi-centralized position where issuer and platform co-sign the delegation chain. Vouched's DIF donation pushes toward an open specification with no single owner.
Two hidden moves deserve attention. Mastercard's adoption of SD-JWT is regulatory arbitrage in plain sight: the format is native to the W3C verifiable credential stack and the EU's digital identity wallet ecosystem. Adopting it early pre-clears European compliance obligations while competitors improvise. That is an asymmetric advantage dressed as a technical standard.
Vouched's DIF donation is a downgrade attack, not charity. If the verification layer becomes a public good, value migrates upward — to integration services, middleware, trust scoring. Visa and Mastercard want the opposite: keep verification private, keep the merchant locked, keep the fee. The stated goal — one registration, universal transactions — is a marketing narrative. The architecture says the opposite.
Eight competing protocols is not a sign of health. It is a sign that no participant has enough leverage to impose a standard — a vacuum that historically favors whoever ships the best developer experience rather than whoever holds the largest balance sheet. I mapped DeFi's fragmented liquidity layers in 2020 and watched the same vacuum resolve inside eighteen months. Visa and Mastercard hold balance sheets. DIF holds neither balance sheet nor shipping velocity. That asymmetry is the quiet risk in the open-standard path.
Then the economics. The integration bands — zero at Shopify, $5K–$50K for Stripe plus custom work, $50K–$200K for a self-built stack, $150K–$500K+ for enterprise PSP work — are unsourced across the board. My instinct, formed consulting for a Geneva wealth manager on crypto infrastructure theses in 2024, is that these originate in PSP sales decks. Nobody quotes a half-million-dollar figure without wanting you to buy something.
And note the zero-cost row. Shopify absorbs the protocol complexity entirely, becoming a toll gate between merchant and protocol. The merchant that "pays nothing" pays in lock-in. Cost estimates and lock-in exposure run in opposite directions — the cheapest integration is the most expensive dependency.
Cloudflare's position deserves separate attention. By bundling Mastercard's delegation checks into its edge middleware, it captures the value of reducing merchant engineering hours without ever owning the trust root. That is the cleanest arbitrage in the stack: sell the cost reduction, outsource the liability.
Then there is the 40% traffic premium. Merchants running both Visa and Mastercard frameworks reportedly see roughly 40% more agentic traffic. Treat that as a compliance tax collected in visibility rather than currency. It is also uncited — and it produces a perverse result: if the premium is real, agent traffic is concentrating into a handful of protocols, which strengthens network effects and pushes the market toward consolidation, not the permanent fragmentation the source describes.
The consensus holds that adoption stalls on integration cost and consumer trust. I would reorder that entirely.
The primary blocker is liability. When an autonomous agent buys the wrong item, or gets manipulated into an unauthorized purchase, who absorbs the loss? The agent operator? The merchant? The issuer? Chargeback machinery assumes a human pressed a button. It has no vocabulary for a delegated, machine-executed transaction gone wrong. Every merchant conversation I have had on this circles the same question — not "what does integration cost" but "who pays when it breaks."
Liability vacuum beats integration cost beats consumer trust. Integration cost is a symptom, not a cause.
On trust, the numbers are unambiguous: only 14% of consumers trust agent transactions without independent verification, 42% refuse them above $25, and AI ranks sixth or seventh as an information source at 4.51 out of 10. Meanwhile 72% of merchants believe consumer adoption will outpace enterprise readiness. No signature scheme fixes a cultural deficit.
Then the blind spot. Of eight protocols listed, x402 gets a sentence. It revives the dormant HTTP 402 status code and settles in stablecoins, over the web, without touching a card network. No interchange. No directory. No consortium. Another rug pull? Or just another myth? I have watched enough "standards" collapse to stay skeptical — but I also watched modular data availability dismissed in 2022 as a Discord curiosity become Ethereum's scaling thesis inside eighteen months. If x402 compounds while the consortiums argue, this entire war is being fought over a shrinking surface.
Three things to watch across the next six months. Network rule revisions on agent liability — that, not SDK unification, is the unlock. Whether DIF specifications acquire real merchant governance, or remain a public good nobody funds. And x402's developer activity. The Cassandra complex is real, and this time the warning is not about a collapse. It is about a coronation nobody is watching.