AI Shopping Agents Force Payments Giants To Rebuild The Wallet

Forty-eight percent of US consumers say they’d let an artificial intelligence agent do their grocery shopping. The same share would hand off subscription management to an autonomous assistant. The payment networks heard them, and they aren’t waiting for the survey ink to dry.

That headline number sits inside the April 2026 Payments Innovation Tracker from PYMNTS Intelligence and Paymentology, centered on a single question: what happens when machines start spending money on behalf of humans? The report’s answer is that the payments stack itself becomes the new battleground. Authorization rules, programmable spending caps, identity verification, and dispute liability all move upstream of the checkout button. Visa, Mastercard, American Express, Stripe, OpenAI, and Google are already racing to define those rules before anyone else does.

What Forty-Eight Percent Of Americans Want Their AI To Buy

The Paymentology-backed survey shows delegated purchasing is no longer a fringe sentiment. Three categories cluster at the top of consumer interest, and they share an obvious trait. They’re boring.

• 48% would let an AI agent do their grocery shopping or plan their meals.
• 48% would let an autonomous assistant manage their subscriptions.
• 44% are open to using an AI for buying gifts.

The pattern is consistent across the data. Routine, high-frequency purchases score the highest delegation interest. Discretionary, high-stakes buys score lower. Consumers want machines to handle the cart they already half-ignore, not the wedding ring or the next car.

Why The Payment Layer Just Became The Real Battlefield

Card networks already lost the recommendation engine fight. Amazon, Google, and Meta took that round. The next round decides who owns the rails an AI agent uses to spend, and the incumbents are spending heavily to make sure it’s them.

The shift is structural, not cosmetic. Traditional e-commerce ran on influencing human intent at the moment of checkout. Agentic commerce moves the human upstream into a policy-setter role, codifying preferences, defining spending limits, and extending trust to a system that transacts continuously. The checkout page collapses into a back-end protocol call. Cart abandonment as a metric becomes meaningless when no human is holding the cart.

Visa’s Chief Product and Strategy Officer Jack Forestell put the stakes plainly in a March 2026 interview with Digital Commerce 360.

I have not stared into a bigger growth opportunity than what we have ahead of us in the development of the agentic web broadly, that then will turn into agentic commerce, turn into agentic payments.

That conviction explains the corporate scramble. Visa’s Agentic Ready Program global expansion announcement in 2026 confirmed pilot programs across Asia Pacific and Europe, with partners including Anthropic, OpenAI, Microsoft, Perplexity, and Samsung. Mastercard, OpenAI, Stripe, and Google have each shipped their own competing protocol stack in the past nine months.

Inside The Protocol Wars Defining How Agents Get Paid

Three competing standards now define how an AI agent identifies itself, proves it has permission, and pushes a card through a merchant’s gateway. None of them is fully interoperable with the others yet.

Stripe’s Agentic Commerce Protocol announcement with OpenAI introduced Shared Payment Tokens, a primitive that lets an agent initiate a payment using a buyer’s permission without exposing the underlying card credential. Google’s AP2 protocol launch on Google Cloud took a different route. AP2 leans on tamper-proof cryptographically signed Mandates that serve as verifiable proof of a user’s instructions, backed by partners including Adyen, American Express, Coinbase, JCB, Mastercard, PayPal, and Worldpay.

Mastercard CEO Michael Miebach told analysts on the company’s October 30, 2025 earnings call that the first agentic transaction had taken place on Mastercard’s network during the third quarter. U.S. Bank and Citibank cardholders piloted Agent Pay first, with a broader US issuer rollout in November 2025 and global expansion announced for early 2026.

Amex’s Quiet Pledge To Pay When The Agent Buys Wrong

While the protocols slug it out, American Express made the most concrete commitment so far. On April 14, 2026, Amex released its Agentic Commerce Experiences (ACE) Developer Kit and Agent Purchase Protection program. The kit covers agent registration, account enablement, intent intelligence, payment credentials, and cart context.

The protection piece matters more than the developer kit. Amex committed to covering eligible cardholders for charges tied to AI agent error, provided the cardholder authorized the agent and the agent sent Amex an authenticated purchase intent. It’s the first time a major issuer has explicitly underwritten the risk that an autonomous system buys the wrong thing.

Who Eats The Loss When The Bot Buys Wrong

Liability is the silent fault line under every press release about agentic commerce. The protocols handle authentication. They don’t yet handle what happens when an agent misreads a prompt and orders six cases of canned tuna instead of one.

Early merchant data tracked by chargeback firms shows agentic checkouts already trend toward lower cart abandonment but higher dispute rates in some categories. The Consumer Bankers Association’s January 2026 symposium white paper on agentic AI flagged the same pattern. Customers dispute purchases their agent made because they don’t remember granting that level of approval. The merchant can’t easily prove intent because the human never touched the checkout flow.

Friendly fraud already accounts for 75% of all chargebacks in conventional e-commerce. Layer an autonomous purchasing agent on top and the abstraction grows. Palo Alto Networks’ Unit 42 documented adversarial uses of agentic capability throughout 2025, including agent-driven credential stuffing and agent-mediated mule recruitment, where compromised agents become unwitting accomplices.

That’s why the protocol wars have a registration component. Mastercard’s Agent Pay requires every agent on the network to be registered and verified, with identity and authorization data passed through each transaction for traceability. Mastercard’s Agent Pay press release from April 2025 framed the registration requirement as a fraud control. It’s also a liability sorting mechanism. An unregistered agent is the merchant’s problem.

The Solana network and Google Cloud went a step further with stablecoin-denominated agent payments, a separate track that bypasses card rails entirely. Coverage of that move in our Pay.sh launch report on Solana and Google Cloud enabling USDC payments for AI agents sketched the parallel rail forming on crypto infrastructure.

From Cart Abandonment To Machine-To-Machine Latency

The metrics that mattered for e-commerce are losing their grip. New ones replace them, and they look more like network engineering benchmarks than marketing dashboards.

• 48% interest among US consumers in delegating grocery shopping or meal planning to AI agents, per the April 2026 PYMNTS-Paymentology Tracker.
• 60+ organizations signed onto Google’s AP2 protocol at launch, including Adyen, American Express, Coinbase, Mastercard, and PayPal.
• $175 million raised by Paymentology in May 2026 to scale real-time processing solutions, much of it earmarked for agentic transaction infrastructure.
• 75% of all chargebacks already classified as friendly fraud before AI agents enter the dispute pipeline.

Patrick Collison, Stripe’s co-founder and CEO, summarized the moment on X when ACP launched in September 2025, saying it’s clear that internet purchasing modalities are going to change a lot. The shift away from human checkout flows toward machine-to-machine settlement reframes what payments companies sell. They no longer sell conversion. They sell the rules of engagement between a customer’s stated preferences and the algorithm spending the customer’s money.

The April 2026 Tracker frames the prize bluntly. Whoever sets the framework AI agents use to transact safely captures the next decade of digital commerce, and that capture happens at the credential layer rather than the storefront. The companies still writing checkout-conversion case studies are reading from last year’s playbook. The ones already filing protocol specifications and underwriting agent error are writing the next one.