Ledger Drift

Subscribe to get high-signal insights on how modern fintech is built.

analysis

Anthropic Built the Shopping Brain and Skipped the Wallet

Anthropic, Visa, and Mastercard announced a commerce agent spec. It handles the shopping. It doesn't handle the money.

By Alex Kugell ·

Anthropic just unveiled a commerce agent spec with Visa and Mastercard. The pitch: AI agents that can browse products, compare options, and make purchases on behalf of consumers. Shopping, automated.

The spec describes how an agent selects a product. It does not describe how that agent proves it has permission to spend $400 of your money.

What the spec covers

The agent side is well-specified. An AI agent receives a shopping intent from a user, browses merchant catalogs, evaluates options against preferences and constraints, and presents a recommendation or executes a purchase. Anthropic's Claude handles the reasoning. Visa and Mastercard provide the merchant network and transaction infrastructure.

This is a real capability. An agent that can comparison-shop across twenty merchants, filter by your dietary restrictions, check your purchase history to avoid duplicates, and negotiate bundle pricing would save hours of human time. The AI works.

What it doesn't cover

The spec says nothing about how the agent actually pays. When the agent is ready to buy, it needs to authorize a transaction. Today, that means one of two things: the agent stores your card number and submits it through a checkout flow designed for humans, or the agent triggers a notification asking you to approve the charge manually.

The first option is a security problem. An AI agent holding a card number, CVV, and billing address is a credential storage risk that Visa and Mastercard have spent decades building PCI-DSS to prevent.

The second option defeats the purpose. If the agent has to ask you every time it wants to buy something, it's a recommendation engine with an extra step.

The missing piece is scoped authorization. An agent needs a token that says "you can spend up to $200 at grocery merchants this week" without ever touching the underlying card credentials. Merchant-category-scoped, amount-limited, time-bounded, and revocable.

Card networks have the building blocks for this in tokenization and merchant category codes. The spec doesn't assemble them.

The Catena comparison

Catena Labs shipped an answer to this problem in 2025 with the Agent Commerce Kit. The agent gets a cryptographic identity, trust scopes that limit what it can do, and a receipt layer so every transaction is auditable.

The agent never holds card credentials. It holds a scoped authorization token that the merchant can verify independently.

That architecture was purpose-built for agents. The Anthropic spec takes the opposite approach: make the agent smart enough to navigate infrastructure built for humans. One designs new rails for a new actor. The other forces a new actor onto old rails.

Where this goes

Three-quarters of consumers surveyed by Visa don't trust AI agents to make payments autonomously. The commerce agent spec doesn't solve that. The agent can decide what to buy. The payment infrastructure has no native way to let it pay.

This gap will close. The question is whether it closes by retrofitting card rails with agent-scoped tokens, or by building agent-native payment protocols from scratch.

The card networks will push the first approach because it keeps them at the center. Startups like Catena, and potentially Stripe with OpenRouter's Machine Payments Protocol, will push the second.

Both will probably coexist. The card networks will handle agent commerce that looks like human commerce with a proxy. The new protocols will handle agent-to-agent transactions where no human is involved at all.

The boundary between those two categories will shift as agents get more autonomous. Whoever defines that boundary defines the next generation of payment infrastructure.

Update: Sep 11

Four days after I published this piece, Visa, Mastercard, and Ant International announced a joint AI agent trust framework. The card networks are building the authentication layer I described as missing. Meanwhile, Poland's BLIK payment system completed a pilot transaction executed independently by an AI agent, no human in the loop. The gap between "agents can shop" and "agents can pay" is closing fast.

Sources

Alex Kugell
Alex KugellAI/ML Engineer, Fintech Infrastructure

Writes Ledger Drift. Engineer building AI systems and financial infrastructure. Open source contributor to x402, ACK, and MPP.

Frequently Asked Questions

Why can't Anthropic's commerce agent complete a purchase on its own?
The spec covers how an AI agent reasons about shopping, comparing products, evaluating preferences, and making selections. It does not specify how the agent authorizes or executes a payment. Without a payment mechanism, the agent either needs to store card credentials, which violates decades of PCI-DSS security architecture, or interrupt the user for manual approval on every transaction, which defeats the automation.
What would agent-scoped payment authorization look like?
A token that restricts the agent to specific merchant categories, a spending ceiling, and a time window, without the agent ever handling card numbers or CVVs. Card networks already have the primitives for this in tokenization and merchant category codes. The gap is assembling them into a credential format that agents can present and merchants can verify without human intervention.
How do purpose-built agent payment protocols differ from the card network approach?
Catena Labs built infrastructure from scratch where agents carry cryptographic identity, operate within declared trust scopes, and produce machine-readable receipts. The Anthropic spec takes the opposite approach: make the agent smart enough to navigate checkout flows designed for humans. One designs new rails for a new actor. The other retrofits the actor onto existing infrastructure. Both will likely coexist, with card networks handling agent-as-human-proxy scenarios and new protocols handling fully autonomous agent-to-agent transactions.
Do consumers trust AI agents to make payments?
Visa's survey found 75% of consumers don't trust agents with autonomous payments. When Visa's brand is involved, 61% said they would trust the process. The card networks are positioning their brand recognition and consumer protections as the bridge between agent capability and consumer confidence, which gives them leverage over agent-native protocols that lack comparable consumer trust.

Built by Trio, a fintech-native engineering partner helping teams build the next generation of financial technology and infrastructure.

Subscribe to Ledger Drift for high-signal insights into how modern fintech is built, from systems to code to teams.

Keep reading

fintechCash Flow Infrastructure Is a Platform DecisionCash flow starts as a product feature and becomes infrastructure the moment two systems need the same balance.
ai & mlAML's Real Bottleneck Is the NarrativeTransaction monitoring generates the alerts, but the four-hour SAR narrative is where compliance teams actually lose their week.
anatomySpade and the Payments Data LayerCard authorization messages give issuers 40 bytes to identify a merchant, so Spade built the side channel that carries the rest.
View more ›