Stripe and Amazon show why agentic commerce will fail at scale
By Nikhil Raghavan · Reporting from San Francisco ·
As Terry Angelos and Alex Rampell tout automated shopping, brittle payment workarounds and retail pushback prove the technology cannot scale without clear statutory liability.
When an early tester tried to dispute an AI agent's purchase at Walmart, American Express could not find the retailer on the ledger. It only saw Stripe, which had secretly bought the item on the user's behalf with a disposable virtual card. Amex handed the customer a phone number that rang directly on the personal cell phone of a Stripe software engineer.
On the latest episode of the "a16z" podcast, venture partner Alex Rampell and Terry Angelos, chief strategy officer at data-vaulting firm Very Good Security (VGS), laid out why Silicon Valley is suddenly obsessed with "agentic commerce." As Angelos put it, AI models have transitioned from spectators to actors because they now possess arms and legs: payment credentials. VGS stores over 10 billion credentials across online commerce. It reports provisioning new "agentic tokens" per week, up from zero three months ago.
The Virtual Card Hack and the 3 A.M. Pager
Strip away the venture jargon, and what is happening under the hood is a classic infrastructure scramble. To let an autonomous agent buy outdoor lights or book a piano tutor, engineers currently route around the standard payments stack using a hack. The platform issues a single-use virtual card, pays the merchant as an intermediary, and settles with the consumer in a separate transaction.
It is sellable, but it is not shippable at scale. The mechanism severs the basic plumbing of commercial law. When Angelos bought chandeliers through Walmart via an agent, the retailer sent a follow-up offer to credit shipping fees toward a Walmart Plus subscription. The upsell failed instantly because the underlying card had evaporated. More dangerously, dispute mechanisms collapse. If a model hallucinates a reservation date or orders the wrong SKU, who answers the chargeback? Today, merchant acquirers absorb friendly fraud by pushing liability back onto the cardholder or eating sub-$10 claims. When transactions are executed through synthetic cards and autonomous scripts, the engineers pushing code to production are the ones getting paged when the reconciliation ledger breaks.
The Platform Defense Against Headless Browsers
The strongest case for agentic shopping is consumer efficiency. Why navigate dark patterns, expiring promo codes, and multi-tiered checkout funnels when an LLM can scan options and execute the optimal transaction via an API? Proponents argue that open standards for delegated authority will force retail monopolies to compete purely on price and fulfillment.
History suggests otherwise. We have seen this cycle play out in screen scraping during the early days of open banking, when aggregators like Plaid scraped bank interfaces until JPMorgan Chase forced everyone into tightly controlled, toll-gated APIs. The platforms will not surrender margin quietly. Amazon already moved to block the shopping agent Muse. The reason is not technical conservatism; it is basic accounting. Amazon is an advertising conglomerate whose highest-margin SKU is the human eyeball scanning search results. An agent running headless browser sessions generates server load without viewing sponsored ads. Retailers will insert verification hurdles, throttle bot traffic, or launch captive agents programmed with house bias.
Nobody Has Written the Spec for Hallucination Fraud
The missing layer here is not another token vault. It is a clear statutory and network definition of delegated authority. In a traditional transaction, authentication relies on three gates: card validity, cardholder ownership, and commercial intent. An autonomous agent blows through the third gate entirely. If an autonomous system misinterprets an ambiguous user prompt and drops $500 on non-refundable tickets, neither the Truth in Lending Act nor Visa's dispute rulebook clearly dictates whether that transaction constitutes authorized use.
Unless the card networks and federal regulators codify exactly where model hallucination ends and consumer authorization begins, agentic commerce will remain trapped in proprietary walled gardens. If an autonomous agent cannot be legally distinguished from a compromised account, the banks will protect their margins the only way they know how: by blocking the transactions.