Skip to main
Back to Industry insights

Agentic commerce for card issuers: five lessons from Money20/20 USA

What Money20/20 USA 2026 showed about agentic commerce for card issuers: agent identity, trust, fraud signals and agent permissions.

The Thredd Team

Last Updated: November 18, 2026

Money20/20 USA last month had two big themes, and AI agents were one of them. By our count, around 55 sessions on the agenda mentioned AI agents or agentic AI, a few more than mentioned stablecoins, and 15 named agentic commerce. Visa, Mastercard, American Express, Stripe, Plaid, Google and Walmart all spoke on agentic panels.

The sessions we found most useful were less about what agents might do one day and more about what an issuer needs in place before it approves a payment an agent started. Here are the five lessons we took away for card issuers.

1. Agentic payments: trust is the open question

One panel, with speakers from Visa, Stripe and the issuing bank Cross River, was billed as tracing an agent-initiated transaction from start to finish, and asking where consent, control, fraud and liability sit along the way. As its description put it, agents can already discover products and complete purchases on a buyer's behalf. Another panel, "Who Let the Bots Pay?", with Mastercard, Okta and Experian, was framed around the guardrails for agent-driven transactions.

The industry has largely solved how an agent pays. What it is still working out is how everyone in the chain decides to trust that payment.

In our view, agentic commerce is a new payment channel rather than a new payments model. The familiar building blocks still apply: cardholder permission, issuer approval, scheme rules, tokenization, fraud monitoring and disputes. The issuer still approves or declines. It needs more context to do it well.

2. Know Your Agent: identity for AI agents

Know Your Agent (KYA) is the process of checking who an AI agent is, who controls it, what it is allowed to do and whether it has permission to act right now.

A panel called "Know Your Agent", moderated by J.P. Morgan, was billed as covering the protocols and identity frameworks being built for agents that transact on behalf of real people. Another, with Fiserv and Trulioo, introduced the idea of an "identity anchor": a framework for assigning trust, control and responsibility to autonomous agents.

For an issuer, KYA comes down to four questions:

  • Who is the agent?

  • Who controls it?

  • What is it allowed to do?

  • Does it have permission to act right now?

Digital wallets are the closest precedent, because they already let a third party transact with a token in place of the card number. The same principle fits agents. An agent should work through a token tied to defined permissions, never the underlying card details, which makes tokenization central to agentic commerce.

3. Agentic fraud: the signals card issuers need

Plaid's session description framed the shift well. When AI opens an account, moves money or pays a bill, proving the person is who they say they are stops being the whole answer. The question becomes whose agent this is, and whether it is behaving or going rogue. A Stripe session on fraud trends was billed as covering agentic commerce risks alongside first-party abuse.

Most card fraud models were designed around human behavior. With agents in the mix, issuers will want to watch for new patterns:

  • An agent buying something different from what was agreed

  • Unusual speed or volume of purchases

  • Purchases outside the limits the cardholder set

  • Attempts to trick an agent into acting against its owner's wishes

Liability is still open. If an agent buys a hundred packs of toothpaste when the cardholder asked for one, who carries the loss? That is still being worked out across the industry. Until it is, controls set up front reduce the exposure.

4. Agent permissions: what an AI agent may do on a card

A panel with Trustly, Yodlee and the Financial Data and Technology Association was billed as tackling the gap between letting AI read someone's finances and letting it move their money. What should an agent be allowed to start on its own? A Rezolve session with Ines Clark of Visa was billed as covering why mandates, identity, permissions and payment choice need to stay connected.

Permissions can sit at several levels:

  • Browse: the agent searches and compares

  • Recommend: it suggests a purchase

  • Prepare: it gets a payment ready for the cardholder to confirm

  • Complete: it pays within agreed rules

  • Manage: it runs recurring payments within a set limit

Where to draw the line is a product choice for each program. Many of the controls that enforce it, such as spend limits, merchant category limits and velocity rules, are ones issuers already use. Early use cases are likely to be the ones where intent is stable and easy to define, such as subscriptions, restocking and scheduled top-ups.

5. Agentic commerce today: early, and moving fast

Several sessions were billed as separating what is live from what is hype. "Self-Driving Money", with Visa and Chime, put it neatly in its description: copilots suggest, agents transact, and the industry does not agree where it sits between the two. Another panel described agentic AI as reaching payment rails before anyone had finished writing the rules.

That matches what we see. In most flows today, a person still confirms the purchase. Visa and Mastercard are rolling out frameworks for recognizing agents and flagging agent transactions, and those frameworks will shape what issuers need to support.

Our view is that this will feel like the arrival of any new payment channel: new transaction types, new data fields, new token flows and new fraud patterns, built on rules issuers already know.

When the agent pays: Identity, trust and fraud in agent-initiated payments

In this session, Ryan Dew, Chief Product Officer at Thredd, is joined by Anthony Craufurd from Visa, Steve Rayko from Zilch, and Stuart Barclay from Velocity to work through what that means for the network, the rail, the issuer, and the merchant.

Watch now

What card issuers and fintechs should do next

The common thread at Money20/20 USA was trust and control. For issuers, fintechs and program managers weighing agentic payments, four steps follow:

  1. Start with one use case where intent is clear. Subscriptions and scheduled payments are easier to control than open-ended shopping.

  2. Set the agent's permissions. Decide which of the levels above you will allow, and the limits that go with them.

  3. Keep card details out of the agent's hands. Use tokens linked to the agent and its permissions.

  4. Tune fraud and controls for agent behavior. Revisit limits and velocity rules, and track the scheme requirements as they evolve.

Those are the same questions we work through with clients. At Thredd, we treat agentic commerce as an extension of the platform rather than a separate product. Tokenization, digital wallet provisioning, scheme connectivity, real-time authorization and fraud monitoring are already part of what we run for clients.

In July, we joined Visa's Agentic Ready program in Europe, starting with Zilch, so issuers on our platform can prepare for agent-initiated payments without replacing their existing infrastructure. Agents use Visa tokens rather than card details, and a Visa Payment Passkey confirms the cardholder's intent. As the schemes extend their agent frameworks, we track the requirements so new agent signals flow through the platform issuers already use.

Talk to our team about how agentic payments could work within your program

Or download the factsheet to share with your colleagues.