On Wednesday, American Express and Fanatics launched a co-branded credit card designed for sports fans. Tickets, merchandise, and fan experiences are all tied to a rewards program built on the assumption that the cardholder cares about points and watches games.

On the same day, Nvidia reported $82 billion in quarterly revenue, calling out agentic AI as a major driver. The contrast is the story. Loyalty programs are built for humans who remember their balance. The shopper that will use those cards next year is increasingly a machine that does not.

Loyalty programs were built on the assumption the customer remembers and cares. Agents skip the entire layer. Operators are not yet pricing the gap.

How loyalty actually works

A loyalty program is not really a discount mechanism. It is a behavioral memory system. The customer remembers their points balance. They activate an offer in the app. They check the balance before checkout. They wait for double-points weekends. Every step of that loop is a human noticing, remembering, and acting.

The economics work because the noticing changes behavior. Promoted offers convert at higher rates than unpromoted ones. Card-linked offers shift spend toward partner merchants. Tiered status drives sustained spend toward a single brand.

That model has been the workhorse of the cards industry for 30 years. Amex built a global business on it. Mastercard and Visa each operate vast loyalty platforms for their issuers. Bilt turned rent into points. Co-brand deals are bigger now than ever, and the Fanatics partnership extends the format into sports, with Amex contributing the rails and the cardholder relationship.

The machine does not notice

Agentic checkout breaks the loop at the noticing layer. A purchase initiated by an agent inside ChatGPT, Claude, Gemini, or Perplexity does not pause at the loyalty app. The agent does not check the points balance. The agent does not see the double-points banner. The agent does not care.

This is not theoretical. OpenAI shipped Instant Checkout late last year, with Stripe providing the rails. Anthropic ran its Project Deal experiment in April, with agents acting as both buyers and sellers inside a closed marketplace. Practical Ecommerce covered a second wave of agentic storefront tooling this month. Each of these surfaces strips the loyalty interaction layer out by default.

What remains visible to the agent is what was sent in the structured product feed or the merchant API. Card-linked offers and points multipliers are not typically in those payloads. The agent does not factor them in. The cardholder might still earn the points after the fact. They simply did not influence the decision.

If the loyalty program does not influence the decision, the loyalty program does not work.

Three structural questions for loyalty operators

Are points portable to agents? The first question for every operator is whether the agent can read the cardholder's earned points and redemption preferences at checkout. Today the answer is almost universally no. Loyalty data sits inside the issuer's app or the merchant's customer record, not in a structured field the agent can call. This is solvable. It is not solved.

Does brand affinity survive the agent layer? Sports loyalty in particular trades on emotional pull. Fanatics works because someone wants their team's jersey, not the next algorithm-ranked equivalent. If the agent ranks by price plus fulfillment speed, the affinity is silent. The Amex/Fanatics card may still earn points on every Fanatics purchase. But if the cardholder asks an agent to buy a jersey and the agent picks a cheaper third-party seller, the partnership generates no incremental volume.

Who gets attribution? Co-brand economics depend on attribution. If an agent intermediates the purchase, who counts as the converter? The agent operator. The merchant. The card. The current attribution model assumes the cardholder makes the choice. When an agent makes the choice, attribution becomes contested, and the partner who used to pay for shelf placement may stop.

What incumbents are doing

Not much, yet. The card networks have spent the past 18 months building agentic identity protocols, not agentic loyalty protocols. Visa Trusted Agent Protocol authenticates the agent. Google AP2 carries signed intent. OpenAI ACP scopes payment tokens to single carts. None of these touch loyalty.

We covered Visa's agentic payments rollout with Santander in five LATAM markets, where the focus was on transaction settlement, not rewards. The loyalty side of the equation has not made it into the public protocol layer. There is no published standard for an agent to read a cardholder's points balance, evaluate a rewards multiplier, or honor a co-brand offer.

This is the gap. The first network to publish a loyalty-readable agent standard sets the rules for the next decade of co-brand economics. The first issuer to bake it into a co-brand partnership owns the example.

The recommendation poisoning angle

If the agent makes the choice, the integrity of the ranking matters more than the integrity of the offer. We wrote earlier about recommendation poisoning, where the merchant or agent provider can manipulate which product the agent surfaces. In a loyalty context the same vulnerability becomes a loyalty kill. If a cardholder asks their agent to buy a Yankees jersey and the agent surfaces a third-party seller paying for placement, the Fanatics deal collapses at the moment of intent.

The defense is structural. Agents need verified product feeds with surfaced loyalty signals. Card networks need to publish the schema. Co-brand partners need to demand the integration.

What to watch

The first co-brand to publish agentic redemption. Whether it is Amex/Fanatics, JPMorgan Chase/United Airlines, or Apple/Goldman Sachs, the first co-brand partnership that names an agent SDK and supports points redemption at agent checkout sets the format.

Card network loyalty schema announcements. Watch for Mastercard or Visa to publish an extension to their respective agentic commerce frameworks that includes loyalty payload fields. The schema is the bottleneck.

Agent platform pricing changes. If OpenAI, Anthropic, or Google start monetizing structured commerce feeds, expect loyalty placement to move there next, with the same dynamics that played out in search advertising 20 years ago.

Charlie Major is a Product Development Manager at Mastercard. The views and opinions expressed in Major Matters are his own and do not represent those of Mastercard.

Sources

If an agent intermediates every purchase, what does the loyalty industry actually sell, and to whom?

Charlie Major is a Product Development Manager at Mastercard. The views and opinions expressed in Major Matters are his own and do not represent those of Mastercard.