AI agents ignite a new battle in the payment industry! Bernstein: Visa (V.US), Mastercard (MA.US), etc. embrace new opportunities

According to Zhitong Finance APP, investment bank Bernstein noted in a recent research report that Meta (META.US)’s launch of Muse has ushered in a new era and competitive round for consumer artificial intelligence (AI) agents. The bank believes that payments could become even more strategically significant as the commercialization of AI agents evolves, potentially heralding the next major battle in the payments industry.

Bernstein pointed out that as the commercialization of AI agents rapidly evolves, payments are keeping pace. Muse is already being used to perform a variety of tasks: shopping (which Meta says is quickly becoming one of its largest usage drivers), tracking prices after booking flights and hotels, making reservations, cleaning up inboxes, and coordinating outfits.

However, the payment experience with agents is still somewhat clunky. During checkout, Muse completes payment by filling out the merchant’s own checkout page via Stripe’s Link or Shop Pay. AI agent startup Instinct stated that 40% of its users shared their personal credit cards within three weeks, though its early users may be more tech-savvy and have higher incomes than the average consumer.

The fundamental question about consumer AI agents in the payments industry is: in an agentized world, who will control payments, liability, and dispute resolution frameworks? What role will payment networks and PSPs (payment service providers such as Stripe and Adyen) play? Will this create incremental revenue opportunities (for example, for V/MA, Adyen, Stripe), or could payment abstraction or tokenization threaten commoditization risk for some payment processors?

Trust remains a major bottleneck in the agentization of commerce—only 23% of U.S. consumers say they trust generative AI to handle their payments. Traditional business models are designed around “human-in-the-loop”. In a world where agents crawl the internet, fill forms, and make purchases, a range of trust issues arise: how can people know if an agent is a "good agent"? How is transaction programmability achieved? How is identity authentication managed (such as Europe’s two-factor authentication, or 2FA)? How is user consent and intent ensured? How should responsibility frameworks be designed? How are disputes handled? How is fraud risk reduced? How are chargebacks managed? How is an audit trail established? How to optimize different payment methods? How to achieve personalization? And how to address standards and technology fragmentation?

The Battle for Control Over Agent Transactions

A tug-of-war is about to unfold over the control of agent transactions and the authority over credentials/payment tokens.

Agents bring significantly more complexity to transactions than before. In the age of agentized commerce, trust and governance are crucial (even more so than in e-commerce), and also highly challenging—entities able to provide agent credentials/tokens and related governance mechanisms stand to gain significantly.

To be fair, agentized commerce currently comprises only a minuscule share of e-commerce, but it’s developing rapidly, and payments companies are preparing for this agentized future.

Both payment networks Visa (V.US) and Mastercard (MA.US) have launched agent tokens (Visa Intelligent Commerce, Mastercard Agent Pay), which embed more data and contextual information into the tokens (such as deeper insights into transaction types, token usage locations, and payer identities), while also adding safeguards (signal-based, leveraging behavioral history and related configuration).

At present, the adoption of these products, and whether issuers are actively participating (e.g., around agent control and liability frameworks), remains unclear. Regardless, 100% of agent transactions will be tokenized (though not necessarily using agent-specific tokens). Meanwhile, Stripe launched Secured Payment Tokens (SPT) for agent transactions a year ago.

It is worth noting that Muse or Instinct services are unlikely to be using Visa and Mastercard’s ‘agent-exclusive’ tokens right now, but are instead using Link or one-time virtual cards—this means payment networks still get these transactions, but not the potentially lucrative ‘agent’ token business, which could in future become a platform for distributing agent-focused services. Muse transactions are simultaneously covered by Link’s purchase protection mechanism.

Interestingly, Visa mentioned in July that it had established a partnership with Meta to use agent tokens within Facebook and Instagram, but no Muse-specific information was found.

Whether payment networks can seize the lead in agent transactions with "agent tokens" may ultimately come down to two words—liability shift. Even if other tokens are just as advanced or more so technically, payment networks play a key role in dispute resolution and liability frameworks, which may become extremely important in the world of agentized commerce.

Learning from History?

The battle for control over agent transactions bears similarities to the token wars of 2014. Twelve years ago, with the launch of Apple Pay, Visa and Mastercard also rolled out their own token services. At that time, fierce competition erupted among payment networks, issuers, and PSPs for control of the tokens.

Today, although multiple tokens coexist, payment network tokens have achieved broad adoption and now account for 60% of e-commerce transactions. Tokenization has also given payment networks a truly leading position, allowing them to distribute their own services via this channel, such as fraud scoring, automatic credential updates, and push provisioning of issuer credentials to merchants. Today, they are even starting to price these services, with much more room for future price increases.

Potential Market Space Could Be Huge

The potential revenue upside could be substantial. For example, Stripe currently charges a $0.0015 (0.15 cent) fee for SPT. If payment network agent tokens achieve widespread adoption, they could use these tokens to distribute their own services. In an optimistic scenario, if agent transactions eventually capture even a mid to high single-digit percentage of all consumer transactions, this could contribute several percentage points of growth for payment networks.

What Should Be Watched?

Key factors to watch include: issuer and merchant system integration (not just default activation); holiday shopping season (and agent adoption); new platform partnerships; whether Google or Meta’s Muse will implement agent tokens; and agent experiences introduced by banks (such as Amex and Chase).

By the way (which is also relevant for platforms), the recently launched agent products do not allow agents to control merchant payments (or assume the Merchant of Record role)—at least, this model is quite complex. Instead, these products are mainly providing consumers with a frictionless way to complete tasks and shop (for example, Muse completing payments via Link or one-time virtual cards).

Merchant system integration faces significant challenges, as does taking over checkout experiences that traditionally belong to merchants, including managing conversion rates, real-time pricing, post-purchase services, synchronization of merchant catalogs and inventory, application of loyalty and discounts, fulfillment, and multiple shopping carts.

Investment Impact

Bernstein believes that payments are the beneficiaries, not the losers, and in the world of consumer AI agents, payments could become even more strategically important.

One possible scenario is that in the future, consumers conduct more transactions via agents (with smaller basket sizes and more discovery of long-tail goods), which would benefit payments companies, including card payment networks.

Moreover, agents could drive further online penetration of complex services as well as certain retail segments. While this might not be as obvious, increased e-commerce penetration generally benefits card networks, since card usage and value-added service (VAS) attachment rates are higher online.

Consumer AI agents may also bring more complexity in terms of agent authentication/verification, capturing consumer intent, dispute handling, agent-exclusive cards/lines of credit, etc.—thus creating differentiation and value-added service opportunities for V/MA and payment processors (like Stripe, Adyen, Checkout).

It is worth noting that 100% of agent transactions will be tokenized (meaning V/MA could earn higher transaction revenue, and it provides them with more service distribution channels). Agent-to-agent commercial transactions may also emerge, which would require a richer ecosystem of services and present further upside.

Bernstein has been thinking deeply about this issue and believes digital wallets (such as PayPal) are in an interesting position in the world of consumer AI agents, possessing strategic options (such as, if agents struggle to find effective monetization models).

On the one hand, if agents handle people’s payments—does that mean consumers will really care about the friction caused by "guest checkout"? On the other hand, bilateral wallets like PayPal (PYPL.US) can significantly expand their value proposition—including trust, protection, fraud prevention, and dispute handling—while also helping consumers dynamically discover savings, rewards, personalized services, and optimize payment choices (like selecting between BNPL, credit card, or debit card).

Buy Now Pay Later (BNPL) providers such as Affirm (AFRM.US) and Klarna (KLAR.US) may also be well positioned, especially in these areas: interest-bearing loans for near-prime and subprime customers (including late fee costs), dynamic credit lines (can they negotiate with agents?), 0% APR loans (for prime customers), and transparency of terms.

In its report, Bernstein rates Visa, Mastercard, Affirm, Adyen (ADYEY.US), Block (XYZ.US), and Toast (TOST.US) as "outperform"; PayPal, FIS (FIS.US), Fiserv (FISV.US), Global Payments (GPN.US), and Klarna as "market perform".