What Are Agent Cards?
Agent cards are scoped payment credentials that let an AI agent spend on a consumer's behalf without ever holding the consumer's real card number.
- Visa, Mastercard, Amex, and Stripe all launched their own agent card programs between 2025 and 2026, each with different rules on reusability, approval flow, and fees.
- Most purchases still require human approval as of mid-2026 — agents aren't yet spending fully autonomously.
- Liability, security, and fee structures remain unresolved across the industry, and a separate, more fragmented set of protocols (AP4M, x402, MPP, AP2) is emerging for business-to-business and machine-to-machine payments.

AI agents can now research a purchase, compare options, and negotiate on your behalf — but for most of their history, they've hit a wall at checkout. Someone still had to type in a card number. That changed in 2025 and 2026, as major card networks and fintechs began issuing agent cards: payment credentials designed specifically for an autonomous AI agent to spend on a person's behalf, without ever handing over the person's real card details.
What Problem Do Agent Cards Solve?
Handing an AI agent your actual card number creates three potential problems:
-
Merchant trust: A merchant receiving a checkout request has no way to tell a legitimate AI agent acting on a real customer's behalf from a bot or bad actor impersonating one.
-
Consumer authorization: There's no standard way to prove which agent is authorized, what it's allowed to buy, and how much it's allowed to spend — without a human approving every single transaction.
-
Accountability: If an agent buys the wrong item, buys too much of something, or gets manipulated into an unwanted purchase, there's no established process for who's liable and how the transaction gets reversed.
Agent cards address all three with one mechanism: a credential that's cryptographically scoped to a specific agent, a specific spending policy, and a specific consumer. That scoping is what lets a merchant verify the agent instead of guessing, defines exactly what and how much the agent can buy without a human approving each purchase, and ties every transaction back to a traceable consent record if something goes wrong.
Agent Cards Compared: Fees, Reusability, and Approval
|
Product |
Network/Issuer |
Card type |
Reusable? |
Approval required? |
Fees |
|
Visa Intelligent Commerce / Trusted Agent Protocol |
Visa (via issuing bank); layers onto OpenAI/Stripe's ACP inside ChatGPT |
Wraps whatever Visa credit, debit, or prepaid card the consumer already links |
Reusable, scoped to spending limits and an approved-merchant list |
Yes — consumer gets a notification to approve each purchase in early rollout |
No separate consumer fee disclosed; standard Visa interchange applies. Commercial terms of the ChatGPT integration specifically are undisclosed |
|
Mastercard Agent Pay |
Mastercard (Citi and US Bank were first issuers) |
Wraps an existing Mastercard credit or debit card via a tokenized "Agentic Token" |
Reusable by default — one token per agent, scoped at setup (e.g. "groceries only," a monthly cap); can also be issued single-use or time-limited |
Not a live per-tap prompt — consent and scope are set upfront; a separate "Verifiable Intent" record documents what was asked for |
No separate consumer fee disclosed; standard Mastercard interchange applies |
|
American Express ACE Developer Kit + Agent Purchase Protection™ |
Amex |
Amex credit/charge cards |
Not yet publicly detailed |
Not yet publicly detailed |
No separate consumer fee disclosed |
|
Stripe Issuing for Agents + Link |
Stripe (issued via bank partners, on Visa/Mastercard rails) |
Prepaid/debit-style virtual cards drawing from a funded Issuing balance, not revolving credit |
Supports true single-use cards, capped to one merchant, one transaction, or a set amount |
Configurable by the developer/business issuing the card, not a fixed consumer prompt |
$0.10 per virtual card issued, plus 0.2% + $0.20 per transaction after the first $500K in volume (waived below that) |
|
AgentCard (Alchemy) |
Built on Visa Intelligent Commerce |
Wraps whatever Visa card the consumer links |
Inherits Visa Intelligent Commerce's scoping options |
Inherits Visa's approval flow |
Not publicly disclosed |
|
InFlow |
Built on Visa Intelligent Commerce |
Wraps whatever Visa card the business/consumer links |
Inherits Visa Intelligent Commerce's scoping options |
Inherits Visa's approval flow |
Not publicly disclosed |
Where a fee isn't publicly disclosed, that's because most of these products aren't separate card products with their own pricing — they're a tokenization/authorization layer sitting on top of a card the consumer already has, so the applicable fee is whatever that underlying card's rate already is. Reusability and approval flows vary by issuer configuration rather than being fixed to the product — the table reflects the default or most commonly described setup as of mid-2026.
1. Visa Intelligent Commerce / Trusted Agent Protocol
Tell ChatGPT you want wireless headphones under $150, and it finds a qualifying pair and completes the purchase — that's Visa's own example of what its integration with OpenAI enables. Visa's Trusted Agent Protocol issues a Verified Agent ID alongside a consent record signed by the consumer's card issuer, so a merchant can confirm both that the agent is legitimate and that the consumer actually authorized it. Inside ChatGPT specifically, this sits on top of an existing layer rather than replacing it: ChatGPT's checkout has run on the Agentic Commerce Protocol (ACP) since September 2025, an open standard co-developed by OpenAI and Stripe, in which Stripe issues a Shared Payment Token (SPT) that lets ChatGPT initiate a payment without ever seeing the buyer's card details. Visa's June 2026 integration adds Visa's own tokenization, agent identification, and fraud-monitoring infrastructure on top of that protocol, letting eligible cards use network-level tokenized credentials for agent-initiated purchases inside ChatGPT and OpenAI's Codex. By December 2025, Visa reported hundreds of controlled, real-world agent-initiated transactions and projected millions of consumers would use agent checkout by the 2026 holiday season.
Features
-
Reusable: Yes, within set spending limits and an approved-merchant list, until the consumer revokes access
-
Approval: In the early rollout, the agent sends a notification and the consumer approves each purchase before it's completed; Visa expects this to loosen as trust builds
-
Fees: Rides on standard Visa interchange; the specific commercial terms of the Visa/ChatGPT integration haven't been disclosed
2. Mastercard Agent Pay
Launched in April 2025 with Microsoft, IBM, and Braintree as initial partners, Agent Pay issues an "Agentic Token" — an extension of Mastercard's existing tokenization infrastructure (MDES) — that represents a cardholder's card, scoped to one agent and one commerce policy. Citi and US Bank were the first two issuers to give their cardholders access. Microsoft's Copilot Checkout, built on Agent Pay, launched to consumers in January 2026 in partnership with PayPal.
Features
-
Reusable: Yes by default — one token per agent, scoped at issuance (e.g., category, monthly cap, day-of-week restrictions); can also be configured as single-use
-
Approval: No live per-transaction tap; consent and scope are set upfront, with a separate "Verifiable Intent" record capturing what the consumer actually asked for
-
Fees: No separate consumer fee disclosed; standard Mastercard interchange applies; revocation of a specific agent's access is real-time through the issuer's app
3. American Express Agentic Commerce Experiences (ACE) Developer Kit
Amex released its Agentic Commerce Experiences (ACE)™ Developer Kit in April 2026, a framework covering five integrated services: agent registration (verifying which AI agents are trusted to transact), account enablement (letting a cardmember register a card for agentic use), intent intelligence (capturing purchase intent for authentication and disputes), tokenized payment credentials, and cart-context sharing for validation. Alongside it, Amex announced Amex Agent Purchase Protection™ — a pledge, described as an industry first, to protect cardmembers from charges caused by a registered AI agent's error, provided the agent transmitted authenticated purchase intent at the time of the transaction. This is the clearest explicit liability commitment among the major networks so far; Visa and Mastercard haven't published an equivalent guarantee.
Features
-
Reusable / approval flow: Not yet publicly detailed beyond the five ACE services above
-
Fees: No separate consumer fee disclosed
4. Stripe Issuing for Agents + Link
Rather than creating a new type of credential, Stripe extended its existing card-issuing infrastructure: agents get programmatic access to a consumer's Link wallet, and can be issued one-time-use virtual cards or a Shared Payment Token backed by the cards and bank accounts already saved in that wallet. This is a lower-friction option for merchants already built on Stripe. The trade-off: Visa and Mastercard's frameworks issue an identity credential — a Verified Agent ID or Agentic Token — that any merchant on their network can independently check against the issuer; Stripe's cards are recognized as legitimate within Stripe's own processing relationship with a merchant, not via a shared, cross-network identity standard.
Features
-
Reusable: Configurable — supports genuine single-use cards capped to one merchant, one transaction, or a set dollar amount
-
Approval: Set by whichever business or developer is issuing the card, not a standard consumer-facing prompt
-
Fees: $0.10 per virtual card issued, plus 0.2% + $0.20 per transaction after the first $500K in volume (waived below that) — this is a platform/developer cost, not a disclosed consumer surcharge
5. AgentCard (Alchemy)
Alchemy's AgentCard integrates with Visa Intelligent Commerce to give an agent what the company describes as a complete identity-and-payment stack, provisioned through a single API. Alchemy's stated use cases — an agent booking travel, ordering groceries, or renewing a subscription for a consumer without that consumer touching a checkout screen — are the company's own examples of intended use rather than independently confirmed case studies; no third-party reporting confirms these specific scenarios happening at scale.
Features
-
Reusability / approvals / fees: Inherits whatever Visa Intelligent Commerce offers underneath, since AgentCard is a layer on top rather than a separate credential
6. InFlow
InFlow describes itself as "B2AI" infrastructure — Visa Intelligent Commerce supplies the underlying payment credentials and merchant acceptance, while InFlow layers on identity, multi-currency wallet functionality, and a policy-governed payments engine so a business can safely let agents transact on its behalf.
Features
-
Reusability / approvals / fees: Same as above — inherits Visa Intelligent Commerce's mechanics; InFlow's own pricing isn't publicly disclosed
Are Agent Cards Safe?
Agent cards solve the mechanical problem of letting an agent pay, but several key questions remain unresolved:
-
Liability: Who absorbs the cost when an agent makes an erroneous or unauthorized purchase? Amex is the clearest case so far, with its named Amex Agent Purchase Protection™ commitment; Visa and Mastercard haven't published an equivalent guarantee yet.
-
Security: An agent's payment credential is a new attack surface — credential theft or prompt injection during an automated checkout flow could result in unauthorized spending in ways that don't map cleanly onto existing card-fraud playbooks.
-
Regulation: Consumer protection law was written with a human pressing "buy" in mind. Regulators haven't yet settled how existing rules on authorization, disputes, and chargebacks apply when an AI agent is the one initiating the transaction.
-
Fragmentation: Visa, Mastercard, Amex, and various fintechs are each building their own agent-credential standard. Whether these converge into something interoperable, or remain a patchwork merchants have to integrate with individually, is still an open question.
-
Fee economics: No network currently discloses a separate consumer-facing surcharge for agent-initiated purchases — but merchant-side fees, and merchant appetite for them, have already reshaped one major product. OpenAI began charging a 4% transaction fee to US Shopify merchants using ChatGPT's Instant Checkout starting January 26, 2026 (PYMNTS). By March 2026, OpenAI scaled back in-chat, product-listing checkout — reportedly after only about a dozen of the millions of eligible Shopify merchants had actually integrated, and while OpenAI still lacked its own system for collecting and remitting US sales tax — in favor of routing purchases through integrated retailer apps like Instacart, Target, and Expedia (Forbes). The underlying open standard, the Agentic Commerce Protocol, is still active and still being developed for that app-based model; what's changed is the scope of direct, in-chat checkout, not the protocol itself. The commercial terms of Visa's newer card-network-level integration with ChatGPT remain undisclosed.
What's Next for AI Agent Payments?
Everything above is consumer-facing: a credential tied to one person's spending authority. A separate, faster-moving, and more fragmented landscape is emerging for a different problem: businesses and AI agents paying each other directly, often in stablecoins, at machine speed and often in fractions of a cent. Mastercard's Agent Pay for Machines, Coinbase's x402, Stripe's Machine Payments Protocol, and Google's AP2 are all built by different companies, solve overlapping but distinct pieces of that problem, and aren't interchangeable with one another. None of them are cards — they're payment protocols and settlement rails, and untangling how they actually relate to each other deserves its own explainer.
The Bottom Line
By mid-2026, giving an AI agent a way to pay stopped being a novelty and became a real, if still early-stage, product category. The mechanics — tokenized credentials, issuer-signed consent, spending policies — are converging across networks even as the branding differs. The open questions now are less about whether agents can pay, and more about who's accountable when they get it wrong.