Beyond the Perimeter of Today's Agentic Banking Thesis

Agentic banking isn’t limited to what we’re seeing in the market. There is the ground a bank already owns, which I argue the bank will keep if it stays sharp, and the ground it can win if it makes the right moves in the near future. We'll define both, how they fit in agentic consumer finance and where the puck is moving.

Agentic consumer finance conversations are addressing only two of the three critical parts. There is agentic banking, where a bank uses its own data and systems to act for the customer inside its own walls, doing what used to take a stack of forms or a trip to the branch to do. Then there is agentic payments, how an AI agent buys something at a merchant. Both are being built in public, with names and pilot programs we’re all seeing and reading about.

There is a third part that has no agreed name and almost nobody is mapping it directly. It’s the largest of the three. It has to do with taking action on a consumers behalf - originating from the bank or fintech - to help a consumer be better with their money directly with the merchants, brands and services consumers use on a daily basis.

The first two are where the noise is and the answers there mostly already exist. You can guess who wins them, and I will get to why. The third is the part still genuinely open, and it is the reason for this article.

Draw the wall

It helps to draw this whole thing out. The great banks and fintechs likely come out on top when it comes to understanding and acting on data within their four walls. Inside that boundary are the deposit account and the ledger, the mobile app and the customer data. That is the home of agentic banking as it works today - agentic banking 1.0.

There’s a perimeter around agentic banking 1.0. That’s agentic payments. A network is not a place you visit; it is the door your card goes through on the way out, and Visa, Mastercard and Discover have sat on that line since long before anyone said the word agent. The far side of the wall is the third region. It has no name in today’s parlance, so I will call it the frontier. When a bank crosses its own wall to act in the frontier, that is agentic banking 2.0.

Put all three together and you get what the industry now calls agentic consumer finance. That umbrella is fine until people use it to mean only the two regions they can already see - agentic banking 1.0 and agentic payments.

The frontier

To understand the frontier, take a brief walk through how a person interacts with the world via the lens of their banking relationship. Think about it as spend and income. A customer signs up for Netflix and puts a card on file, and the bank does not see it happen. Payment credentials are now sitting in Netflix's vault. The first the bank hears of it is a charge for $24.99 that clears a few weeks later. If that customer later replaces that card with a different account to chase points, the bank learns about it only when its own charge stops arriving, and it is never told why. The bank's first hint that a customer is leaving is usually a charge that fails to show up, if the bank notices. But they don't know if the customer cancelled the subscription or changed the card-on-file.

The same thing happens up and down the frontier. Subscriptions are cancelled. Confusion about transactions lead to chargebacks. A direct deposit sits in one account for years and then moves without warning. Somewhere in the pile, almost always, is a retirement account from a job two employers back, a few thousand dollars the customer has not looked at since they left that job years ago.

None of this resolves cleanly at the bank. It happens off-bank. The data that would explain the difference lives inside the merchant, payroll system, tax prep provider or biller. That data sits behind a login the bank does not have access to.

The blindness gets worse behind the biggest merchants, because that is where mountains of personalization pile up. The average Amazon Prime member in the US runs more than a thousand dollars a year through Amazon - with many spending tens of thousands of dollars - and behind that single merchant name sits the Prime membership, the Subscribe & Save reorders, the Prime Video channels, and purchases on other cards, whether their own or a competitor's. On-bank transactions that reach the bank may be identified as Amazon transactions with an amount and date of spend, but that’s pretty limited. Off-bank transactions are entirely unknown to the financial institution today.

Apple has the same shape on the subscription side. More than a billion paid subscriptions now run through Apple's platform, its own services and other companies' apps billed side by side, and the bank sees one merchant. One charge is not the thing to look at. What matters is everything stacked behind the merchant's login, and the bank is on the wrong side of it.

Wealth splits along the same line, which settles a question that fits awkwardly into payments or banking. The advisory core, the discretionary management and the planning, stays inside an institution for the licensing reasons it always has. The coordination does not. Tracking down held-away accounts and consolidating money across firms the customer has no relationship with is frontier work. It sits right next to the cancelled subscriptions, and it is worth a great deal more.

Agentic banking 1.0

Mercury recently launched Command, and it is one of the cleanest pictures yet of where agentic banking starts. A customer talks to the account the way they would talk to a person. The example Mercury's Alexey Likuev gave was setting up a checking account, creating an auto-transfer rule, and spinning up a card for a daughter's summer trip, all in one conversation, with the customer approving each step before it ran.It runs on OpenAI's GPT-5.5, and Mercury is blunt that it treats the model as a commodity and swaps in whatever system tests best — notable given the company recently received conditional approval from the OCC for a banking charter.

Look at where every one of those actions happens. Opening the account, setting the rule, issuing the card, answering a question from the customer's own data, all of it sits inside Mercury's walls. This is agentic banking 1.0: the things a bank has always done, now done by talking or chatting, instead of tapping through menus, having to speak with someone and fill out a bunch of forms providing data the bank already has. Elements of agentic banking 1.0 are popping up all over right now. Bank of America logged more than thirty billion client digital interactions last year across its platforms, with Erica alone fielding roughly 700 million of them, and JPMorgan's internal AI now runs across more than two hundred thousand employees. None of it required the bank to leave home.

Agentic banking 2.0

The move that matters more is the one where the bank crosses its own wall. Agentic banking 2.0 is the bank acting on the frontier, reaching into the accounts a customer holds off-bank, on the customer's behalf and with their permission. The permissioned connections into those accounts are the bridge; version 2.0 is the bank crossing it.

Example - The bank identifies (with bank data) that a customer seems to be paying for two subscriptions at the same merchant, reaches out to the customer about it and asks if the customer wants to cancel one of them. Then the bank - permissioned by the customer - operates on their behalf to cancel the unwanted subscription.

It can appear in other ways: moving the card on file at a merchant over to the one that earns better points, sending a direct deposit somewhere new, or pulling SKU-level data behind a charge the customer is squinting at and considering filing a dispute for.

This sounds like new ground but banks have been reaching across the perimeter for years, just narrowly and one rail at a time. Bill pay sends money to a landlord or a utility the bank has no other relationship with. A chargeback is the bank stepping into a fight with a merchant on the customer's side, clawing money back under rules written for exactly that. Credit score is pulled from the large bureaus and embedded in financial services apps. It's one of the most-used features in these apps. And when a card gets reissued, Visa's Account Updater and Mastercard's equivalent quietly reach into every merchant holding that card on file and swap in the new number so the subscriptions do not lapse. That last one is fascinating to me. The card networks already reach into merchant card-on-file vaults and change the credential for you. They only do it when a card is reissued, on the network's schedule, never because a different card would serve the customer better. Version two is that same motion, opened up and handed to the customer.

Whether a bank can actually pull any of that off comes down to three things, and only one of them gets talked about. The first is usable data. Plenty of banks sit on rich transaction data they struggle to use, and insights stranded in a warehouse never reach the customer. The second, and the harder one, is the experience. Banks and fintechs are figuring out how to put on-bank and the off-bank data into one view and act on both from there. The final element is the bridge that needs to be built into the frontier merchants, payroll and tax prep systems, and billers. The banks that manage pulling these three things together will win big, because they stop being the customer's primary banking relationship and become the primary banking interface, the place a person goes to run their whole financial life.

The reasons this stays with the bank, rather than drifting to an outside agent, were covered in AI won’t embed banking. Banking will embed AI.

What it does for the bottom line

The two versions pay off in different currencies. Banking 1.0 mostly lowers the cost of serving a customer you already have. Servicing moves from a phone queue to a chat, and the customer who can do everything by asking tends to stay longer. Those are real savings and they land early. They are also capped, because the ceiling is doing what you already do for less. The intelligence behind it is rented, which Mercury says out loud, so it is not where one bank pulls ahead of the bank across the street.

Agentic banking 2.0 changes how much a customer is worth in the first place. When the bank's agent moves its card into first position across a customer's subscriptions, every recurring charge starts earning interchange at a minimum. Capturing the direct deposit and consolidating the old retirement accounts brings balances and assets onto the bank's books - a direct line to NIM (net interest margin). A customer whose whole off-bank financial life runs through one bank's agent does not shop around, and attrition is where retail banking quietly bleeds value. The off-bank and line-item data that comes back sharpens underwriting and reduces costs related to fraud. The bank has become this customer's primary banking interface.

Agentic banking 1.0 is a necessary foundation and it produces the on-bank intelligence that agentic banking 2.0 runs on. Just do not mistake the on-ramp for the destination. The cost savings are the easier part and I believe every bank that makes it through the next 3-5 years will have them eventually. But the bottom line grows upward when the bank crosses the wall, and that is the ground the model-agents are and will be fighting for in the future because who controls that can make a great bid for becoming the primary banking interface.

Where a bank should start, and what it's worth

Not every move onto the frontier is equal. We see fintechs and banks entering in a few different ways: subscription management, payment switching, card on file management, SKU-level data. No one way is right but a few things we’ve seen might be helpful. Subscription visibility and cancellation is slowly being pushed by card issuers. Visa and Mastercard have begun mandating subscription management and clearer recurring-payment transparency from the banks that issue their cards, starting in Europe, Southeast Asia, and the Middle East, with the US expected to follow. The direction of travel is set. A bank that builds this now gets ahead of the mandate instead of scrambling when it lands. It also wins the public opinion poll. Payment switching is a clear primacy play and if you offer a product people want it can be both engaging and a great entry point to establishing a bridge to providers that make subscription management and SKU-level data much easier. SKU-level data as a way of reducing the volume of disputes has also been highly effective.

Direct deposit switching, tax prep and consolidating held-away retirement accounts are all pieces of the frontier that banks and fintechs have engaged with in the last few years. These are not a single build path. They reach into different systems, so a bank gets to choose its order. The order I would choose starts where it is small and valued, earns the bank a right to win future elements of the frontier, then takes on the harder, higher-value ground.

Sizing this precisely is impossible. That said, we know that winning the US consumer-banking, payments, and wealth revenue pool runs well in excess of $1 trillion a year. The new money the frontier itself opens runs to tens of billions: $40 to $50 billion of recurring-card interchange sits in positions an agent can move, the $2.1 trillion in forgotten 401(k)s alone can convert into deposits and fees, and households waste more than $25 billion a year on subscriptions nobody remembers signing up for. Capitalize the per-customer uplift across the country's banked households and the value of being the interface, rather than the backend, runs into the hundreds of billions. The exact figure resolves only against a given bank's own book. The shape does not. Whoever wins this wins the next phase of consumer banking.

The open question is whether the pull toward agentic gets strong enough to make the merchants and platforms open real APIs, or whether reaching the frontier stays what it is today, user-permissioned and reverse-engineered. Either way, whoever can deliver that access durably owns the moat.

The wall, and who holds it

A word to the wall - agentic payments - because that is where most of the noise has been. The consumer-checkout version of agentic payments is going to belong to the networks and not for reasons of branding. A merchant never trusted the shopper anyway. What it trusts is the network standing behind the card, and that trust is exactly what an agent needs to borrow at checkout. The machinery for letting an unknown party be trusted already exists and already belongs to Visa and Mastercard (I think Discover has a strong play here, as well, which I'll get to). Mastercard shipped Agent Pay in April 2025. It binds a tokenized card to a specific agent and a set of rules about what that agent may buy, built on the same MDES tokenization that already runs contactless and stored card-on-file. Citi and US Bank cardholders were among the first turned on. Visa has its own version, Intelligent Commerce, and spent its June 2026 payments forum bringing OpenAI in and standing up a directory of agents it has verified as real. The chargeback and the liability shift a cardholder never thinks about have no equivalent on the open internet, and a merchant letting an agent check out is going to want them.

The more underrated move is Capital One's. It closed its purchase of Discover in May 2025, a deal worth more than $35 billion that handed it the Discover network and PULSE. Set the cards aside and look at the company that came out the other side. Capital One can now issue the card, run the network it rides on and operate the deposit account. It sits on every side of its own transaction. In the ordinary four-party model an agentic payment has to gather consent and settle liability across an issuer, a network, an acquirer, and a merchant, each pulling in its own direction. Capital One can fold most of that into one company. American Express has quietly taken this approach for decades, and the closed-loop players may end up moving faster on agentic for exactly that reason: there is no one else to convince.

Will stablecoins become a critical part of the wall? The networks haven’t yet locked in one element that’s coming. A kind of payment is showing up that cards were never built for: agents paying other agents, a fraction of a cent at a time for a slice of data or a model call, with no human in the loop and nothing to dispute. Interchange and a chargeback window make no sense there, so it is being built on stablecoins. Coinbase's x402 protocol has already cleared more than a hundred million such payments on Base at no protocol fee. None of this threatens Visa at the grocery store. It is a new market forming beside the old one, and the real question for a network is whether it owns the part that still matters when value settles in stablecoins: the agent's identity and the signed record of what it was allowed to do. Visa seems to know it, and put stablecoin settlement on the table at the same forum where it pitched its agent directory.

Summary

None of this is hard to see once you map it out clearly. Agentic banking 1.0 and agentic payments already have owners, and I don’t see those key players getting dislodged. Banks and fintechs providing excellent banking experiences will continue to win in agentic banking 1.0. The networks hold the wall, and they’re likely to continue holding it. The frontier is wide open, and it happens to be where the customer actually spends their financial life. It is not the kind of ground a license or a token can hand you, but those assets can increase your right to win. The real winners will be those that actually help people be better with their money. Banks and fintechs have generally done a good job keeping people's money safe, and the best have already garnered some rights to act on the customer's behalf. In the agentic age, these players can win the deeper trust required to operate across the frontier, the trust that turns a good banking experience into the primary banking interface for their customers.

Sources

• Mercury, "Mercury Launches Command, a New Way to Run Your Business Through Banking," June 16, 2026. https://www.businesswire.com/news/home/20260616596362/en/Mercury-Launches-Command-a-New-Way-to-Run-Your-Business-Through-Banking
• Mastercard, "Mastercard unveils Agent Pay," April 29, 2025. https://www.mastercard.com/us/en/news-and-trends/press/2025/april/mastercard-unveils-agent-pay-pioneering-agentic-payments-technology-to-power-commerce-in-the-age-of-ai.html
• Visa, "New AI, Stablecoin and Token Innovations at Visa Payments Forum," 2026. https://usa.visa.com/about-visa/newsroom/press-releases.releaseId.22491.html
• Visa, "Visa Account Updater overview," developer documentation. https://developer.visa.com/capabilities/vau
• Capital One, "Capital One Completes Acquisition of Discover," May 18, 2025. https://investor.capitalone.com/news-releases/news-release-details/capital-one-completes-acquisition-discover
• Amazon Prime member average annual spend. RedStag Fulfillment, "Average Annual Spend of an Amazon Prime Member," 2026. https://redstagfulfillment.com/average-annual-spend-of-an-amazon-prime-member/
• Apple paid subscriptions surpassing one billion. PYMNTS, "Apple Tops a Billion Subscribers as Services Climb to Record Revenue," January 2025. https://www.pymnts.com/earnings/2025/apple-tops-a-billion-subscribers-as-services-climb-to-record-revenue/
• Chainalysis, "Inside x402: 100M Agentic Payments on Base," 2026. https://www.chainalysis.com/blog/x402-agentic-payments-adoption/
• Atomic, "AI Won't Embed Banking. Banking Will Embed AI," June 1, 2026. https://atomic.financial/insights/ai-wont-embed-banking-banking-will-embed-ai/
• Card interchange. Nilson Report, US card swipe fees in 2024 totaled $187.2 billion ($148.5 billion credit, $38.7 billion debit). https://nilsonreport.com/articles/merchant-processing-fees-in-the-united-states-2024/
• Interchange math. Issuer interchange is the bulk of that swipe-fee total, roughly $150 billion. Card-on-file and recurring positions an agent can switch are an estimated 25 to 30 percent of consumer card spend, so $150 billion times about 30 percent is $40 to $50 billion. This is contestable interchange, the share that moves to whoever holds first position, not net-new revenue.
• Forgotten subscriptions. Self Financial and C+R Research, 2024: about $32.84 a month, or $394 a year, per person on unused subscriptions; other surveys put the figure closer to $200 a year. https://www.self.inc/info/cost-of-unused-paid-subscriptions/
• Subscription math. $200 to $394 per household per year across roughly 130 million US households is $26 to $51 billion a year. The article cites the conservative floor of $25 billion.
• Forgotten 401(k) balances. Capitalize, "Forgotten 401(k)s Grow to Record $2.1 Trillion," 2025. https://www.businesswire.com/news/home/20250930788421/en/