Agentic commerce does not threaten the rails. It threatens the toll.

The disintermediation worry misreads the threat. The real pressure of agentic commerce falls on pricing power, and the banks hold the weakest hand. 

Zennon Kapron, Director, GL Insight 

On March 2, 2026, Banco Santander and Mastercard completed what they called Europe’s first live agentic payment, executed on Santander’s production infrastructure. Visa expects millions of consumers to let agents check out for them by the 2026 holiday season. The infrastructure question for agentic commerce is largely answered. The harder question for anyone running a payments business is what happens to their economics once the buyer is software. 

The common worry is that AI agents will route around Visa, Mastercard, and the banks, but that reads the threat backwards. The networks are not about to be disintermediated: they moved first, and they will almost certainly stay in the transaction. What is genuinely at risk is the price they can charge for being there. 

A world of choice is a human problem 

The premise behind most agentic-commerce commentary is that buyers face an overabundance of payment options, and that the clutter pressures incumbents. The clutter is real. It is also something agents resolve rather than suffer. 

For a person at checkout, more rails mean more decisions, and the rational response to decision fatigue is a default. People reach for the card they always reach for, and that habit is the quiet foundation of card economics. For an agent, more rails are not confusion but a wider search space. An agent feels no fatigue and keeps no favourite card, so it simply evaluates the options and selects the best one. 

Agentic commerce does not multiply choice in any way that helps incumbents. It collapses choice to the single best option for each transaction, measured on whatever the agent has been told to weigh. The clutter that protected default behaviour does not survive contact with an optimiser. 

It also rewires the merchant side. Today, a merchant optimises checkout for human psychology, building flows around saved cards and the comfort of as few decisions as possible. A merchant that would rather route a payment to a cheaper rail has no clean way to act on that preference, because the customer expects their usual card. An agent turns merchant steering, long a theoretical pressure on card economics, into something software can do at the moment of payment. 

An agent has no loyalty to spend 

Card economics rest, in part, on a behaviour that an agent does not reproduce. In the US, the average interchange fee runs at around 1.81% of the transaction, with Visa’s published rates spanning roughly 1.15% to 2.40% and Mastercard’s reaching as high as 2.90%. Merchants pay it, and consumers tolerate it because it is invisible to them and because switching away from a familiar card takes effort most people never spend. Issuers fund rewards programs out of it, which deepens the habit. The structure leans on inertia, on the customer who does not shop around and the default that never gets re-examined. 

An agent removes that inertia. It will not pay a network close to 2% if a rail that costs a fraction of that clears the same purchase to the same merchant, and it has no top-of-wallet position to protect and no attachment to a brand or a piece of metal. Point an agent at a basket of rails and tell it to minimise cost, and it will. 

There is a second-order effect on the issuing side. Rewards are funded out of interchange, and rewards are what justify a premium card, so when the interchange an agent will pay compresses, the rewards budget compresses with it. The card an agent has no reason to prefer also becomes the card an issuer has less money to make attractive. 

The rail does not have to disappear for the toll to come under pressure. The agent only needs a credible alternative and the instruction to weigh it. This is a different threat from disintermediation, and probably a harder one to manage, because it does not show up as lost transactions but as margin compression on transactions the network still carries. A payments executive modelling agentic volume should be modelling blended take rates falling, not volumes vanishing. 

Why the networks moved first 

The networks understand this, which is why the question of how incumbents stay relevant understates what they are already doing. They are not trying to stay relevant as networks; they are trying to become something harder to skip. 

Visa Intelligent Commerce, launched in April 2025 with partners including Anthropic, OpenAI, and Microsoft, issues scoped, tokenized credentials that an agent can carry, with identity and spending controls attached. Mastercard Agent Pay, announced a day earlier, does the parallel thing through Agentic Tokens that extend Mastercard’s tokenization standard with an agent identifier and a scoped session carrying limits and an expiry. Both networks also sit inside Google’s Agent Payments Protocol, the open standard launched in September 2025 with more than sixty partners, whose core construct is the mandate, a digitally signed statement of what an agent is permitted to do. 

Read together, this is a repositioning play. A mandate is not a settlement rail; it is an authorisation and trust layer. If the network owns identity, permissioning, and dispute resolution for agent payments, then being chosen stops being the question, because the network is no longer competing to be the rail but supplying the layer every rail has to clear through. That is a defensible position, and it is the right move. 

The open question for executives building toward it is whether the trust layer holds its margin or commoditises in turn. Mandates are becoming an open standard, and open standards are good for adoption but not always good for pricing power. 

The rail underneath 

What an optimising agent reaches for is increasingly easy to see. Coinbase’s x402 work, extended into AP2, gives agents programmable, always-on stablecoin settlement at a cost structure cards cannot match, and the Agentic Commerce Protocol from OpenAI and Stripe, launched in September 2025, wires checkout directly into ChatGPT. These are not lab projects; they are shipping. 

The point is sharper outside the US. In markets with cheap, ubiquitous instant rails, the agent has an obvious alternative and no card-shaped gap to fill. India’s UPI processed 228.3 billion transactions in 2025, up roughly a third year on year, and Brazil’s Pix is approaching eight billion transactions a month and cleared on the order of 6.7 trillion dollars in 2025. An agent operating in São Paulo or Bengaluru does not need a card network in the loop, and it does not carry the path dependency that keeps the US consumer on plastic. The US card is sticky in part because account-to-account payments never got cheap and good there, and agents do not inherit that history. They inherit a menu, and the menu looks different in every market. 

For a payments executive, the practical consequence is that the agentic mix will not look the same in any two markets. Expect agents to lean on cards where cards are entrenched, and away from them where instant rails or stablecoins are cheaper and good enough. That also moves a compliance problem into the open, because an agent optimising purely for cost may pick a settlement rail the institution behind it would not have chosen, which is why control has to live in the governance layer rather than the rail. 

The banks hold the weakest hand 

The networks have a repositioning move. Most banks do not. 

When the agent picks the funding source, two bank relationships weaken at once: the deposit account the money sits in and the card the bank issues. The bank becomes a balance sheet under a protocol it does not control, supplying the money rather than making the decision, and for a large issuer, that is the difference between earning interchange and earning a thin spread on settlement. 

The exception shows the rule. JPMorgan Payments signed a global agreement with Mirakl in 2026 to handle payment processing, tokenization, and fraud protection for agentic commerce, and it has been explicit about the logic. In its own words, ‘the differentiator won’t be ‘AI’ - it will be governance: identity, consent, limits’. That is the same move Visa and Mastercard are making, made from a bank, and it is available to a handful of institutions with the scale to build it. The thousands of issuers without that scale have no equivalent, and no bank-led standard at the scale of AP2 or ACP has emerged to give them one. That gap is the story for most of the banking sector. 

For a mid-size issuer, the realistic options are narrow. It can plug into a network’s trust layer on the network’s terms, or join a consortium large enough to build one, but doing neither means competing as commodity funding, which is the one competition a bank cannot win. 

Relevance is the wrong target 

Staying relevant is too easy a test, and clearing it will not protect anyone’s economics. Visa will be in the transaction. Mastercard will be in the transaction. The question that decides the next decade of payments is narrower: who keeps the right to charge for being there. 

Agentic commerce will not kill the incumbents; what it does is closer to an audit. It strips out the inertia that pricing quietly depended on and shows how much of the take rate was paying for friction rather than value. The networks that turn themselves into the trust layer have an answer, and so do the banks that build governance. Everyone still pricing as though a human is choosing should assume an agent is about to find out what that choice is worth. 

Author the author

Zennon Kapron is the founder and CEO of GL Insight, a research and content firm focused on global fintech and payments. He has spent his career analyzing financial technology across Asia and global markets, advising banks, card networks, and technology firms on strategy. A Forbes contributor on fintech, he is a regular commentator and speaker on payments innovation, digital banking, agentic commerce, and the future of money. 

About GL Insight 

GL Insight is a research and content firm specializing in global fintech, payments, and financial services. It produces thought leadership, market research, and strategic content for banks, payment networks, and technology companies, alongside independent analysis published on glinsight.com. Combining primary research with deep sector expertise, GL Insight helps clients understand where the payments and fintech markets are heading, and what that means for their strategy.