Agentic Tokens: How Visa and Mastercard Let AI Agents Pay by Card

Short answer: An agentic token is a scoped, single-purpose version of a payment card that an AI agent uses to pay, so the agent never touches the real card number. Both Visa and Mastercard now issue them: Visa through Visa Intelligent Commerce and Mastercard through Mastercard Agent Pay. The token is limited to a merchant, an amount, or a time window, which makes it far safer than pasting a corporate card into an agent config. What an agentic token does not do is decide whether the agent was allowed to make the purchase, enforce a budget across a fleet of agents, or pause for a human when spend looks wrong.

Last updated July 2026.

What an agentic token is

Tokenization is not new. When you add a card to a phone wallet, the merchant never stores your real number; they store a token that stands in for it and only works in a specific context. An agentic token applies the same idea to an AI agent. Instead of giving an autonomous agent a real 16-digit card, the network mints a token that represents the card but carries built-in limits: which merchant it can be used with, how much it can charge, and how long it stays valid.

The practical effect is a credential with a small blast radius. If a normal card number leaks, an attacker can use it anywhere until you cancel it. If an agentic token leaks, it is close to useless, because it is bound to one purchase context rather than being a reusable credential. That is the core security argument for letting agents pay this way, and it holds up.

How Visa and Mastercard implement them

The two big card networks arrived at agentic tokens on similar timelines but with different emphasis. Visa Intelligent Commerce issues Visa-tokenized, merchant-specific credentials with scoped permissions, so a token minted to book a flight only works with that airline for that trip. Mastercard Agent Pay uses Mastercard Agentic Tokens, built on the same tokenization that already powers contactless and secure card-on-file, with Citi and US Bank among the first issuing banks. Both networks added machine-to-machine payments in June 2026, letting agents pay other agents and services directly.

If you are choosing between them, the differences are real but narrower than the marketing suggests. We walk through them in detail in our comparison of Visa Intelligent Commerce vs Mastercard Agent Pay. For most teams the network is a secondary decision; the primary decision is how you govern what the agent spends once it can pay at all.

Why an agentic token beats a shared corporate card

The honest comparison is not Visa versus Mastercard. It is an agentic token versus what most teams actually do today, which is drop a shared corporate card number into an agent so it can buy things. On that comparison the token wins on every axis that matters.

A team that moves from a shared card to agentic tokens genuinely reduces risk. The mistake is thinking that is the whole job.

What an agentic token does not do

A token scoped to a merchant is not the same thing as a budget. It caps what one credential can do at one seller. It does not cap what a fleet of ten agents can spend across forty sellers in a day. It does not know that an agent renewing a SaaS seat for the fourth time this month is looping on a failed retry. It does not pause when an agent, steered by a prompt injection buried in a web page, tries to buy something no policy would ever allow. Every one of those transactions can be perfectly tokenized, fully network-compliant, and still be money you never meant to spend.

This is not a flaw in Visa or Mastercard. It is out of scope by design. The networks answer how an agent pays. They leave which agent may spend, up to what total, with whom, how often, and who approves above a threshold to you, because those are your policies, not theirs. The gap gets expensive precisely because the token makes the payment feel handled when the decision behind it was never checked.

How to actually control AI agent card spend

Governing agent spend takes a layer that sits above the token and enforces policy before any credential exists. In practice that means four things working together. First, a funded wallet per agent with a hard limit, so no single agent can reach beyond its own balance. Second, spend controls checked at request time: a per-transaction ceiling, a budget over a rolling window, a merchant allowlist, and velocity rules that catch runaway loops. Third, a human approval gate, so spend above a threshold you set pauses for a person instead of going through. Fourth, an immutable audit trail that ties every charge back to the agent, its human owner, the intent, and the policy verdict that allowed it.

Alerting sits alongside all of that. Even with limits in place, finance and operations want to know the moment an agent's spend crosses a line, and a real-time budget alert the instant spend crosses a threshold turns a month-end surprise into a same-minute heads-up. The token controls the credential; the policy layer controls the decision; the alert controls how fast you find out. You need all three before you let software spend on your behalf.

The bottom line

Agentic tokens from Visa and Mastercard are a genuine upgrade over handing an agent a real card, and if you are letting agents pay at all, you should be using them rather than a shared number. Just do not mistake a safer credential for a governed one. The token limits what a single payment can do if it goes wrong. It does not decide whether the payment should happen. That decision, a budget an agent cannot exceed, an approval a human signs, and a record you can audit, is the part you own, and it is the part that keeps autonomous spend accountable as your agents start buying more on their own.