All parties have veto power—why does proxy payment get stuck

AI agent payments remain stuck in a negative feedback loop because intent is neither standardized nor visible across the card ecosystem: 69% of 1,000 U.S. adults surveyed by NMI distrust AI with payments, 70% want to review or veto agent purchases, and only 10% would delegate buying entirely, while a YouGov/ACI survey of 2,080 U.K. adults found 60% would stop using an AI shopping agent after a single mistake. The core failure is that a payment can succeed while violating user intent—agents may misread conditions such as refundability or drift from instructions through hallucination, stale data, or prompt manipulation—yet merchants, PSPs, acquirers, issuers, and card networks each see only a fragment of the transaction, making authorization and liability allocation impossible. Current approaches include raw card-number autofill via vaults (Instinct, Hark Labs) that sit in a gray PCI compliance zone, Agentcard's PCI-compliant encrypted vault with Face ID approval and auto-approve thresholds (e.g., under $50) covering 30 PSPs including Stripe, Shopify, and Worldpay, and agent tokens from Visa, Mastercard, and American Express launched over the past year. OpenAI's Instant Checkout, launched September 2025, had already contracted by March 2026 to product discovery and merchant-owned checkout, illustrating how monetization via merchant commissions and affiliate programs has failed to scale. The article argues the problem persists even if models perfectly understand intent, because intent remains unverifiable by counterparties; what to watch is whether card-network agent token standards gain merchant and issuer adoption and whether any scheme can transmit verifiable intent evidence across all parties.