At two in the morning, while your customer sleeps, her AI agent reviews the interest rate on her savings account, compares it against forty other institutions, initiates a transfer to the best offer, and updates her budget to reflect the change. She finds out over breakfast, in a one line summary she may not even read. She has been your customer for eleven years. Her agent has no such loyalty, and it never will.
This is not a scenario from a strategy offsite. The rails for it are being laid right now, in public, by the largest payment companies in the world. In April 2025, Mastercard announced Agent Pay, a program built on agentic tokens that lets AI agents transact autonomously, developed with Microsoft and IBM. Visa followed with Intelligent Commerce, a portfolio that embeds payment credentials, authentication, and controls directly into AI initiated transactions, with OpenAI as a named partner. In September 2025, OpenAI and Stripe published the Agentic Commerce Protocol, an open standard for agent led purchasing, and switched on Instant Checkout inside ChatGPT with Etsy and Shopify merchants. PayPal has since brought its wallet into ChatGPT as well.
Notice what these announcements have in common. Every one of them treats the AI agent, not the human, as the transacting party. The infrastructure of commerce is being rebuilt around a new kind of customer, and that customer is software.
I have spent over two decades building and integrating technology inside the banking industry, which means I have lived through every previous prophecy of the industry's disintermediation. Internet banking was going to kill the branch. Peer to peer lenders were going to kill the loan book. Neobanks were going to kill the incumbents. Open banking was going to kill the account relationship. Each wave changed banking, sometimes profoundly, and each time the incumbents survived by doing some combination of three things: copying the innovation, buying the innovator, or waiting for the economics of customer acquisition to crush the attacker.
That playbook worked because every previous wave attacked either the interface or the product. The relationship, the durable, inert, profitable connection between a person and their primary bank, stayed put. Customers did not move because moving was tedious and the perceived differences were small. Banking strategists have a polite name for this: the primacy of the primary relationship. An economist would call it monetized inertia.
Agentic AI does not attack the interface or the product. It attacks the inertia itself.
The economics of inertia, and what dissolves them
McKinsey's Global Banking Annual Review 2026 puts numbers on what is at stake. Global banking earned roughly 1.3 trillion dollars in net income in 2025, and deposit margin, the profit made possible largely because customers do not move idle cash, has historically driven on the order of 60 percent of retail banking revenue. The same report describes exactly how agentic systems threaten it: agents that monitor balances in real time, compare returns across institutions, and sweep idle cash into higher yield accounts.
Read that sentence again as a retail banking executive. An agent that sweeps idle cash is not a feature. It is a machine for deleting deposit margin. Every basis point you earn from a customer not paying attention is a basis point an agent exists to claw back.
The speed of the shift matters as much as its direction. McKinsey notes that generative AI reached 45 percent of the US working age population within about two years of ChatGPT's launch, an adoption milestone that digital banking took fifteen years to hit. Banks calibrated their transformation clocks to the fifteen year version of disruption. The two year version is now the operating environment, and as the same review warns, there will be no grace period for banks this time.
Meanwhile the competitive field has already tilted. The top 1,000 fintechs have grown their share of banking revenues from 10 percent in 2021 to 17 percent, compounding at 22 percent annually against roughly 5 percent for traditional banks, and digital native players such as Nubank and Revolut serve 131 million and 69 million customers respectively at returns on equity incumbents rarely touch. These are the firms best positioned to hand every customer an agent, because their unit economics improve when customers optimize, while an incumbent's deteriorate.
Why the old defenses fail against a machine
The uncomfortable exercise is to walk through each traditional moat and ask how it performs when the counterparty is software.
Brand? An agent does not feel reassured by a 170 year old name on a branch. It reads disclosures, parses fee schedules, and scores reliability from data. Brand still matters, but it matters as a machine readable trust signal, not as a feeling.
Experience? Banks have spent two decades and extraordinary sums making apps pleasant for human eyes and thumbs. An agent calls an API. The award winning onboarding journey, the carefully tuned push notification, the personalization engine: an agent routes around all of it. This is the quiet writedown nobody has booked yet, the depreciation of customer experience assets that assumed a human on the other end.
Cross selling? The mortgage conversation that starts because a banker notices a customer's savings pattern does not happen when the savings live wherever the agent parked them this week. Distribution built on proximity to the customer's attention collapses when the customer stops paying attention, because paying attention is now delegated.
Switching costs? These were always partly artificial: forms, waiting periods, direct debit reissuance. Regulators have spent a decade dismantling them through open banking regimes precisely so that third parties can act on customers' behalf. The plumbing that was meant to empower consumers empowers their agents just as well.
American Banker has framed the resulting question bluntly, asking whether AI agents create disintermediation risk for banks. The honest answer from inside the industry is that they create something sharper than disintermediation. Previous disruptors inserted themselves between the bank and the customer. The agent does not insert itself anywhere. It is appointed by the customer, carries the customer's mandate, and takes the relationship with it wherever it goes. You cannot acquire your way out of that, because you cannot acquire your customer's agent.
Utility, agent, or arbiter: the three postures
Strategy discussions about agentic banking tend to wobble between denial and panic. It is more useful to recognize that a bank facing an agent mediated market has three coherent postures available, and that choosing none of them is itself a choice, the worst one.
The utility. Accept that the agent owns the relationship and compete to be what the agent selects: the best priced, most reliable, most API complete manufacturer of deposits, credit, and payments. This is a real business with real returns, but it is a scale and efficiency game with commodity margins. It is also where every bank drifts by default if it does nothing, except that drifting there means arriving with a cost base built for a relationship business.
The agent. Fight for the principal role by becoming the trusted agent yourself. Banks hold advantages here that the technology industry quietly envies: regulated status, deposit insurance, decades of transaction history, and a legal duty of care. A bank issued agent that optimizes across a customer's whole financial life, including moving money to competitors when that serves the customer, is a radical proposition, because it requires cannibalizing your own margin before someone else's agent does it for you. Early moves in this direction are visible in assistants such as the one Lloyds is deploying in its mobile app, though the test of seriousness will be whether any bank lets its assistant recommend a competitor.
The arbiter. Own the trust layer that agentic finance cannot function without. Someone must verify that an agent is authorized, that its mandate is genuine, that the human it claims to represent actually exists and consented, and someone must stand behind the transaction when the agent is deceived. The card networks are racing to occupy this ground, but banks, with their know your customer obligations and their balance sheets, have a legitimate claim to it. The Bank for International Settlements' vision of a tokenized financial system built on central bank money and commercial bank money, with integrity safeguards as a core design test, leaves considerable room for regulated institutions to anchor exactly this role.
These postures are not mutually exclusive, but they demand different investments, different talent, and different pricing discipline, and a bank that funds all three timidly will lose to banks that choose one boldly.
What Monday morning looks like
For executives who find the framing persuasive but the horizon abstract, the near term agenda is concrete. Make your products agent readable: pricing, terms, and fees exposed through documented APIs, because products an agent cannot parse are products an agent cannot choose. Build agent identity and mandate verification now, before fraud teams meet their first wave of synthetic agents claiming to represent real customers; the deepfake economics of that attack are already visible in today's scams. Reprice deliberately for machine shopped products, because rates set on the assumption of customer inattention will be systematically exploited. And re examine every metric that assumes human attention: app engagement, net promoter scores, time in app. In an agent mediated market, the healthiest customer relationship might show zero logins.
Banking has survived every disintermediation wave of my career, and the industry's confidence is earned. But each survival came from the same underlying fact: the customer stayed put while the technology churned around them. The agentic wave inverts this. The technology will sit still, embedded in protocols and payment rails, while customers, through their agents, finally move as fast as the market always assumed they would.
The banks that thrive will be the ones that stop asking how to make customers love them and start asking a stranger question: what does it take to be chosen, over and over, by a machine that cannot love anything and compares everything? Answering it is uncomfortable. Every alternative is worse.