facebook noscript

Agentic Tokens Explained: How AI Agents Pay Without Ever Seeing a Card Number

November 19, 2026

Agentic Tokens Explained: How AI Agents Pay Without Ever Seeing a Card Number

Key takeaways

  • An agentic token is a scoped, AI-ready payment credential that lets an AI agent pay on a user’s behalf without ever seeing or storing the underlying card or bank details.
  • The agentic token credential is associated with the agent’s identity, spend limits (intent and mandate that include, but are not limited to, spend limits, merchant restrictions, and time frame), and user identification and verification (ID&V). Agentic tokens sit between AI agents and existing payment rails, such as end processors or acquirers.
  • Real products already exist: Mastercard Agent Pay, Visa Intelligent Commerce (VIC), American Express Agentic Commerce Experiences (ACE), with Discover coming soon, all of which define how agents carry scoped credentials through a transaction.
  • Merchants and fintechs can use neutral, processor-agnostic vault infrastructure to issue and translate agentic tokens across any agentic protocol without re-architecting their AI agents or locking into a single platform.

What Is an agentic token?

An agentic token is a scoped digital payment credential issued by a provider and the card networks on behalf of a registered AI agent. It represents an underlying card, with embedded rules about how, where, and by which agent it may be used. Agentic tokens are tangible or digital markers of autonomy and empowerment that enable AI agents to execute tasks or transactions within predefined parameters.

An agentic token limits the scope of the credential by binding it to the user’s mandate and consent. It may capture agent identity, spend limits, time windows, and merchant or category restrictions. Agentic payments use scoped credentials instead of typed card numbers, and agentic tokens allow delegating decision-making while preserving user control.

Agentic tokens can reduce friction and anxiety in risk and trust by providing clear boundaries. They also provide a concrete signal that an individual- or, in this case, an agent-has the authority to make decisions within those boundaries.

Here is how agentic tokens compare to related concepts:

  • A network token is issued by a network, such as Visa or Mastercard, typically device- or merchant-bound, and optimized for card-on-file or merchant-initiated transactions.
  • An API key authenticates software to a service. It does not carry per-purchase spend limits or merchant constraints, and it is not a payment instrument.
  • A virtual card is a real card number (PAN) issued with limits, usually single-use or merchant-locked. See the full comparison below.

The actor topology is straightforward: a consumer authorizes a payment method. An AI agent operates in a client or cloud environment. A merchant or aggregator surfaces products. A PSP or acquirer processes the payment. A secure vault sits between them, issuing and managing the agentic token so no party in the AI layer ever touches raw credentials.

Mastercard’s Agent Pay announcement and Visa’s Intelligent Commerce both launched in April 2025, marking the first time major card networks explicitly built tokenization for AI agents.

Agentic token network surrounded by connected signals

Why AI agents need a new kind of payment credential

Picture a travel-planning AI agent that searches six airlines, three hotel chains, and a car rental service on your behalf. To complete those bookings, it needs to make a payment. Handing that agent a raw card number could allow your 16-digit card number (PAN) to leak into prompts, model memory, application logs, vector databases, and observability traces. Every system the agent touches becomes part of your PCI scope.

To avoid this, the industry has created agentic tokens: payment credentials designed for AI agents that serve as stand-ins for the underlying card, allowing the agent to browse, choose, and buy without ever handling sensitive data. Recently, a16z reported that AI-referred e-commerce traffic engages more, spends more time, views more pages, is more likely to convert (and less likely to bounce), and, perhaps most importantly, has a ~50% revenue uplift compared with non-AI-referred traffic. Agentic commerce aims to automate online shopping through AI agents, ranging from consumer shopping agents embedded in various platforms such as ChatGPT, Gemini, Claude, Instinct, or even Grok to production agents that handle enterprise procurement and recurring payments.

Ideally, the infrastructure needs to be flexible enough to adapt to the rapidly evolving agentic ecosystem. This article focuses on agentic tokens as the credential layer for agentic commerce and on the neutral infrastructure that translates those tokens into the formats expected by each PSP, acquirer, or card network.

Two hands meeting to represent trusted agentic payments

How agentic tokens work, step-by-step

Agentic payments involve a four-step lifecycle: discovery, authorization, payment, and reconciliation.

Collect

  • Payment details, including the card, are captured in a secure wallet or vault form running in the browser, app, or SDK.
  • The AI agent never receives the raw PAN or CVV. Only the vault or wallet provider sees the sensitive data.

Tokenize and vault

  • The vault encrypts and stores the original credential, then returns an agentic token that represents the customer’s card and includes agent-specific constraints.
  • Agent-led payments can include identity and authentication data in addition to payment credentials. Mastercard Agent Pay makes agent involvement recognizable, while its Verifiable Intent framework links consumer identity, authorized instructions and constraints, and the resulting purchase.
  • For Visa Intelligent Commerce, the Trusted Agent Protocol uses signed, time-limited agent-to-merchant requests and supports separate signed consumer recognition and payment objects. Consumer recognition can include an ID token; API-based payment flows can carry an encrypted payload containing a payment token and its month and year of expiration, while form-based checkout can use a credential hash.
  • The merchant, or fintech, stores and passes only the agentic token, keeping their systems out of PCI scope for card details.

Authorize with the token

  • When the AI agent needs to pay, it presents the agentic token to the merchant or the PSP’s API rather than a card number.
  • AI agents can make purchases on a user’s behalf under user-defined rules, such as budgets, approved merchants, and time windows. Even when the user isn’t present, these are generally agent-initiated purchases rather than merchant-initiated transactions (MITs). MITs, such as subscription renewals, are submitted by the merchant under an existing customer agreement and require the proper classification, stored-credential agreement, and linkage to the initial transaction. Agentic tokens let AI handle routine payments within those limits.

Pre-model protection: Format-preserving tokens can be applied before data enters an AI model. Even if the agent “remembers” a card-like string, that does not make it a real PAN.

 

Agentic tokens vs. COF network tokens

Card-on-file (COF) network tokens and agentic tokens are both network tokens, meaning they are issued by the card networks as secure substitutes for the underlying card number, but COF tokens are bound to a specific merchant to enable repeat payments the cardholder has already authorized, whereas agentic tokens are scoped for use by an AI agent acting on the user’s behalf within defined limits and permissions.

Agentic Token COF Network Token
Issued by Card networks (Visa, Mastercard, Amex) (Note: Discover agentic tokens are not live yet in market) Card networks (Visa, Mastercard, Amex, Discover)
Scope Card credential plus agent identity, user intent and mandates (e.g., purchase limits), and user ID&V (e.g., passkeys) Encrypted card credential only, bound to a merchant or device
Card cryptogram Yes, a cryptogram validates each transaction Yes, a fresh cryptogram is required for each initial or customer-initiated authorization.
Auto card updates Yes Yes
PCI impact Yes, PCI Compliant Yes, PCI Compliant

Network tokens replace Primary Account Numbers for security. They are issued by card networks for specific merchants, enhance transaction efficiency and reduce fraud risk, and include automatic lifecycle management. Both major networks are extending tokenization for AI agents. Mastercard Agent Pay lets agentic network tokens be tied to agent and session data, and Visa Intelligent Commerce does something similar through its Trusted Agent Protocol.

An agentic token issued through a card network program, such as Mastercard Agent Pay, is itself a network token, not a layer above network tokens. Some vaults and orchestration platforms also offer their own agent-facing alias that sits above network tokens. In that design, one alias for an AI agent may map to multiple underlying network tokens or tokenized credentials across different processors or regions. The vault or orchestration layer decides at runtime which one to use based on routing logic, while the agent only handles the alias. This is a platform design choice, not a universal industry definition. An API key, by contrast, authenticates the agent to a service but does not authorize it to spend from a specific payment method under defined limits and permissions.

Shield protecting agentic payment credentials

Security and compliance: keeping raw credentials away from AI agents

The primary value of agentic tokens lies in reducing the exposure of raw payment data within AI systems.

Agentic tokens keep PANs out of:

  • Model prompts and completions where data might be echoed or memorized
  • Agent memory stores, vector databases, and tool logs
  • Application logs, observability traces, and debugging dashboards used by engineering teams

PCI-DSS implications are significant. To help AI agents, LLM gateways, and agent orchestration layers remain compliant, the industry recommends using agentic tokens for payment to reduce the cardholder data environment footprint.

Every token use can be logged with the agent ID, consumer mandates, and ID&V. This provides security teams with a clear audit trail for incident response and fraud investigations, ultimately supporting authorization and dispute resolution.

 

Frequently asked questions about agentic tokens

  • What is an agentic token?

    An agentic token is a scoped digital payment credential issued for an AI agent. It represents an underlying card, with embedded rules about how, where, and by which agent it may be used. Agentic tokens are tangible or digital markers of autonomy and empowerment that enable AI agents to execute tasks or transactions within predefined parameters.

    An agentic token limits the scope of the credential by binding it to the user’s mandate and consent. It may capture agent identity, spend limits, time windows, and merchant or category restrictions. Agentic payments use scoped credentials instead of typed card numbers, and agentic tokens allow delegating decision-making while preserving user control.

    Agentic tokens can reduce friction and anxiety in risk and trust by providing clear boundaries. They also provide a concrete signal that an individual- or, in this case, an agent-has the authority to make decisions within those boundaries.

  • How is it different from a network token or a vault token?

    Agentic tokens issued through a card network program, such as Mastercard Agent Pay or Visa Intelligent Commerce, are themselves a type of network token. They aren’t a layer above network tokens. A network token is a substitute for the underlying card PAN, subject to network usage controls. Card-on-file (COF) and device tokens are different use cases of the same technology: they are typically bound to a specific merchant or device to support repeat payments the cardholder has already authorized. An agentic token operates within an agentic payment framework that associates the credential with agent identity, user authorization, and purchase constraints through supported records and controls. These don’t need to be embedded in the token itself. For example, passkeys can authenticate the user, but the token doesn’t carry the passkey.

    A vault token is different. It is a provider-specific alias resolved by that vault. It is meaningful within the provider’s environment, and it isn’t a network-issued credential. Some vaults and orchestration platforms offer an agent-facing alias that can map to one or more underlying network tokens, with the platform selecting one at runtime based on routing logic. That is one possible implementation, not a universal design, and it doesn’t mean tokens are portable across processors. Lifecycle updates, such as automatic card updates, depend on issuer and program support.

  • How is an agentic token different from a virtual card?

    A virtual card is a card number (PAN) issued with controls such as a spending limit, an expiration date, or a merchant lock, and it can be single-use or reusable. On its own, it still means the agent handles a card number, which can end up in prompts, agent memory, logs, and observability traces.

    An agentic token is a network token that replaces the underlying PAN. It operates within an agentic payment framework that associates the credential with agent identity, user authorization, and purchase constraints through supported records and controls. Merchants and issuers can then see that an agent is acting, whereas a standard virtual card payment doesn’t signal that. Transaction-specific validation data is provided as required by the network and flow, and token use can be logged with the agent ID and mandate for audit and dispute resolution.

    The two aren’t mutually exclusive. An eligible virtual card can itself be tokenized for agentic use, so virtual card controls and agentic token scope can work together.

  • Is the merchant still merchant of record?

    Generally, yes, but tokenization doesn’t determine merchant of record (MoR). The MoR is the entity that sells the goods and collects payment, typically the merchant or aggregator. The agent acts on the consumer’s behalf, so using an agentic token doesn’t change who the seller is. The exception is when an agent platform or marketplace is itself the seller, in which case it may be the MoR. This depends on the specific program and contracts.

    Tokenization also doesn’t establish PCI compliance on its own. Qualifying EMV tokens can reduce PCI scope, because the merchant stores and passes a token rather than the card PAN. But systems that handle account data, or that are connected to an environment that does, can remain in scope.

  • How does a merchant know that an agent is legitimate and that it actually has the user’s consent?

    These are two separate checks, and agentic commerce relies on three things working together: the token, a verifiable agent identity, and user ID&V.

    • Recognizing the agent. Recognizing a signed agent request indicates that a known agent is making it. It doesn’t prove the user authorized this particular purchase. Visa’s Trusted Agent Protocol (part of Visa Intelligent Commerce) uses signed, time-limited agent-to-merchant requests, and Mastercard Agent Pay makes agent involvement recognizable.
    • Verifying the user’s authorization. This is a separate check. It depends on user ID&V and on records of what the user authorized. Mastercard’s Verifiable Intent framework links consumer identity, the authorized instructions and constraints, and the resulting purchase. Visa’s protocol supports separate signed objects for consumer recognition (which can include an ID token) and for payment.

    Network tokens are also validated at authorization. Applicable Visa cardholder-initiated token authorizations need fresh cryptograms, while properly classified subsequent merchant-initiated transactions omit them. Token use can be logged with the agent ID, consumer mandates, and ID&V, giving security teams an audit trail for fraud investigations and dispute resolution.

  • What are the customer mandates/intent signals associated with an agentic token?

    Mandates and intent are the user’s recorded authorization for what the agent may do. They include, but aren’t limited to:

    • Spend limits
    • Merchant or category restrictions
    • Time frame or time windows
    • Agent identity

    Depending on the network and implementation, these may be linked to the credential through separate signed mandates, intent records, and server-side controls, rather than being embedded in the token itself. The credential operates within an agentic payment framework that associates it with these records, allowing the agent to make decisions while the user retains control.

  • What kind of controls or limits can be placed on a token?

    Depending on the network and implementation, agentic tokens can be scoped by merchant, category, amount, and time window, and can be tied to a specific agent identity and its session data. These controls may be applied through supported records and server-side controls rather than within the token itself. The token can be revoked centrally without affecting the underlying card or account.

    These limits only help if they’re enforced at the payment boundary. If they are, a stolen token is much harder to misuse beyond its intended scope. It may still be usable within its allowed limits, so the limits reduce exposure rather than eliminate it.

  • What happens if an AI agent or its environment is compromised-can a user’s card be compromised?

    Tokens can be scoped by merchant, category, amount, and time, limiting exposure if compromised, and can be revoked centrally without affecting the underlying card or account. Provided limits are enforced at the payment boundary, a stolen token is far harder to misuse beyond its intended scope, though it may still be usable within its allowed limits.

  • Can we use agentic tokens for non-card rails, such as ACH or stablecoins?

    Not directly. The agentic network tokens discussed here represent card credentials and cannot themselves authorize an ACH transfer or stablecoin payment.

    Broader agentic payment frameworks can support non-card methods through separate integrations. ACH requires the appropriate bank-payment authorization and processing flow; stablecoin payments require an authorized wallet or signing service. Identity and consent evidence may be reused where supported, but each rail still needs its own payment credentials and controls.

  • Does VGS have agentic tokens?

    Yes, VGS provisions agentic tokens across three major card networks, including Visa, Mastercard, and American Express. VGS has spent years building that foundation as a token service provider and vault, and it’s why networks and platforms trust us to bring agentic payments to production with agentic tokens. If you’re building an AI shopping agent, an agentic checkout experience, or a merchant platform that agents will transact with, we’re happy to discuss agentic tokens with you.

 

Conclusion and next steps

Agentic tokens are the credential layer that lets AI agents operate safely on top of existing payment infrastructure, without ever touching raw card or account numbers. Trusted agentic commerce ultimately depends on three pillars:

  • Secure payment credentials (agentic tokens) that scope what each agent can spend and where
  • Verifiable agent identity so merchants and issuers can verify who is acting on a user’s behalf
  • User identification and verification (ID&V) so issuers and merchants can confirm the real person behind the agent has authenticated and authorized the agent’s actions

Your next steps: Map your current and planned AI agent use cases and identify where agentic tokens can remove PCI exposure. Engage with a neutral tokenization and vault provider to pilot agentic payments. Map your current and planned AI agent use cases and identify where agentic tokens can reduce PCI exposure. Engage a neutral tokenization and vault provider to pilot agentic payments. Agentic commerce is coming fast, and the businesses that put their credential layer in place today will be first to benefit.

Ready to give your AI agents a safe way to pay?

If you’re building an AI shopping agent, an agentic checkout experience, or a merchant platform that agents will transact with, you don’t need to wait for the ecosystem to catch up. The tokens are live. The infrastructure is ready.

Talk to the VGS team to start provisioning agentic tokens today.

Contact
howard-bio-blog

Howard Xiao

VP of Partnerships

Linkedin Icon
Bassam Chamaa

Bassam Chamaa

Head of Wallets & PSP Partnerships

Linkedin Icon

You Might Also Be Interested In...

VGS Is Live with Agentic Tokens and Already Provisioning Millions
News
VGS Is Live with Agentic Tokens and Already Provisioning Millions

VGS is live with agentic tokens across Visa, Mastercard, and Amex, provisioning 500K+ daily to enable secure AI agent payments.

November 12, 2026
How VGS Helps Gambling and Betting Companies Build a Payments Stack That Keeps Up
Payments
How VGS Helps Gambling and Betting Companies Build a Payments Stack That Keeps Up

Learn how VGS helps gambling and betting operators build a payments stack that owns player credentials, reduces fraud, and keeps payments running.

November 4, 2026
Agent Connect Returns in 2027: The Premier Industry Gathering Shaping the Future of Agentic Commerce
News
Agent Connect Returns in 2027: The Premier Industry Gathering Shaping the Future of Agentic Commerce

Join VGS at Agent Connect 2027 in San Francisco on March 9, 2027. A one-day event on agentic commerce, AI payments, and secure payment infrastructure.

October 27, 2026