By Charlie
Over the past week, several developments landed in finance that, on the surface, did not look related.
On Aug. 25, Google Cloud launched Gemini Enterprise for Financial Services, pushing AI into workflows closer to the core of financial institutions, including KYC, credit analysis, portfolio monitoring, bond issuance and financial research.
Two days later, MoonPay said it had integrated PayBox with Kamino on Solana, allowing users to initiate lending and access yield through conversational interfaces such as Claude and ChatGPT.
At the same time, the Federal Reserve Bank of Dallas published research examining what could happen to deposit stickiness if deposits can move 24/7 in real time and AI can automatically compare yields for businesses and consumers.
A day later, the Bank for International Settlements, or BIS, in a discussion of stablecoins and tokenized deposits, again referred to Project Agorá, which had just completed a live-money test.
Viewed one by one, each item fits a familiar category. Google belongs to enterprise AI. MoonPay sits in crypto finance. The Dallas Fed paper is banking research. Project Agorá is part of work by central banks and commercial banks on next-generation cross-border payment infrastructure.
Placed side by side, though, they point to a line running through all of them.
The real threshold is not whether AI understands finance better
For the past two or three years, the industry has kept asking when AI will truly change finance. The article argues that the question is not precise enough. The real dividing line is not when AI becomes better at understanding finance, but when it is allowed to make financial decisions on behalf of people and actually move money.
The gap between those two things may be larger than the jump from search engines to ChatGPT.
A few years ago, when the intersection of generative AI and fintech first drew attention, the practical use cases were still relatively conservative: customer service, report writing, fraud prevention, expense management, internal knowledge retrieval and tools that lift employee productivity. Those uses mattered, but they largely kept AI standing next to money rather than touching it.
The reason is straightforward. If a model writes a flawed piece of marketing copy, it can be revised. If a model sends $5 million to the wrong place on behalf of a company, that is not something fixed by pressing undo. Financial institutions have far less tolerance for black-box behavior than ordinary software companies. Accuracy, permissions, auditability, accountability and regulatory requirements all have to be addressed. If any one of those pieces is unresolved, a model remains an assistant no matter how capable it looks.
The market is not leaping from manual finance to full autonomy
What has changed over recent months, in the author’s view, is not that the industry suddenly believes fully automated AI wealth management is right around the corner. It is that more companies are breaking the idea of letting AI move money into smaller, more controllable layers of authorization.
When Robinhood introduced Agentic Trading this year, it did not hand a user’s entire portfolio to a model. It created a separate account that AI was allowed to operate. By the end of the second quarter, those accounts were approaching 100,000, with more than $100 million in assets under management.
Ramp has taken a less flashy route, but one that may be closer to how real corporate finance evolves. Its AI does not replace the CFO outright. It reads company reimbursement, payment and accounting policies, handles the cases where confidence is high, and sends exceptions back to humans.
Rocket Money’s recently introduced Rowan points in the same direction. It no longer just shows users a chart of monthly spending. It can detect subscriptions in the background, send reminders, help cancel services, negotiate bills and even complete savings transfers under preset rules.
Put together, these products suggest the market is not jumping directly from human operation to machine autonomy. It is moving along a more practical sequence: observe, recommend, prepare an action, obtain human approval, then slowly hand over the low-risk tasks with clear rules to the machine.
That is why the article argues that the most important concept in agentic finance is not automation. It is authorization.
Authorization systems may matter more than a smarter model
Today, when an employee carries a corporate credit card, that already represents a form of financial authorization. The employee cannot spend all the money in the company’s bank account, but can make purchases within predefined merchant categories, limits and approval rules. Fund managers, corporate treasurers and procurement managers operate under the same principle. They manage someone else’s money inside an authorization framework set in advance.
What an AI agent needs to do is not invent a new financial order. It needs to turn this familiar relationship into rules that a machine can parse and execute.
A company could tell an AI treasury agent, for example, to keep at least 30 days of operating cash on hand; place excess funds only with designated banks or high-rated products; avoid foreign-exchange risk; require human approval for any transfer above $500,000; switch only if the yield gap is more than 20 basis points; and force manual review of any newly introduced counterparty.
Once machines can directly interpret such rules, the nature of the system changes. AI does not need unlimited power. It does not need to become an unregulated robot hedge fund. It only needs room to act inside boundaries.
The article says that this is the increasingly visible common direction in agent payments work at Google, MoonPay, Visa and Mastercard. On the surface, the conversation is about how AI can make payments. In practice, much of the effort is going into permissions: who gives the mandate, how much authority is granted, when it expires, which merchants and accounts are in scope, under what conditions the system must stop and ask for confirmation again, and how a complete audit trail is preserved if something goes wrong.
From that angle, one of the most important products in agentic finance may not be a more intelligent large language model. It may be a reliable financial authorization system that can translate a vague instruction such as “manage my company’s idle cash” into hundreds of machine-enforceable rules.
MoonPay’s recent moves look different when read as a sequence
Seen through that lens, MoonPay’s actions over the past six months become more interesting than the Kamino partnership alone.
In March, MoonPay launched Open Wallet Standard, focusing on how AI can use wallets safely and obtain limited signing permissions without touching private keys. In May, it acquired DFlow to add trade execution infrastructure. In June, it bought Entendre to add reconciliation, finance, treasury management and reporting capabilities. In July, it launched PayBox, connecting cards and wallets directly to interfaces such as Claude and ChatGPT. In August, it added lending and yield management through Kamino.
None of those steps looks earthshaking in isolation. Together, they resemble the assembly of a financial base layer for AI agents: identity and permissions first, execution next, financial operations after that, then credit, lending and asset allocation.
That is also where the line between agentic payments and agentic finance becomes clearer.
The first asks: how can AI pay on my behalf?
The second asks: how can AI manage my balance sheet?
Once the question shifts to the second one, payment becomes only one action in a larger stack. Where cash sits, when to borrow, what assets to pledge, how much yield idle funds should earn, when to rebalance, and how liquidity and risk should be managed are all decisions that have traditionally fallen to treasury teams, bankers and investment managers. Those judgments may start moving into the machine’s execution range.
Without machine-speed money, agentic finance hits infrastructure limits
There is still a basic infrastructure tension underneath all of this.
If AI can compute an optimal choice every second, but money itself still moves at the pace of banking hours, correspondent banking, batch settlement and manual reconciliation, then even a highly capable agent will have limited effect. The article compares that mismatch to a high-performance sports car forced onto a rough dirt road.
That is why Project Agorá belongs in this discussion.
Project Agorá is not an AI project. It is an experiment jointly advanced by the BIS Innovation Hub and the Institute of International Finance to study how tokenized commercial bank deposits and central bank reserves could be placed on a programmable, multicurrency platform to improve wholesale cross-border payments.
Its connection to agentic finance is not that BIS has started doing AI. The connection is that the project is exploring another piece of infrastructure AI may ultimately need to call: whether money itself can become easier for machines to read, compose and execute against.
In July this year, 28 financial institutions and central banks completed live-money testing in a controlled environment involving the Swiss franc, euro, pound sterling, yen, won and U.S. dollar. The test was small and nowhere near production scale. Even so, it verified at least one thing: commercial bank deposits and central bank money are not naturally confined to the slow and fragmented systems used today. They can also exist in an environment that supports conditional triggers, atomic settlement and automated rule execution.
BIS remains cautious. Interoperability between systems, legal definitions of final settlement, responsibility when smart contracts fail, cybersecurity, governance and migration from decades-old systems all remain unresolved. Describing Agorá as a completed next-generation global financial system would clearly overstate the case.
Still, the experiment highlights a point that is often obscured by crypto narratives: programmable money does not necessarily mean stablecoins, and it does not necessarily mean banks are bypassed. Traditional bank deposits themselves could become financial assets that software can call.
The Dallas Fed focuses on the value of deposit stickiness
That is where the more interesting part begins.
Banks are spending heavily to remove friction from money movement, but some of that friction has long been part of the banking business model.
This is the key insight in the Dallas Fed paper, according to the article.
Why are bank deposits valuable? There are many formal answers. When a business keeps $50 million at JPMorgan, the decision is not simply about the inconvenience of opening another account. It may be tied to credit lines, foreign exchange, payroll, cash management, custody and years of customer relationship. Those links do not disappear just because another bank offers a rate that is 10 basis points higher.
But there is a less elevated reason that is just as real: moving money is a hassle.
If a company has $2 million sitting idle, Bank A offers 4.20% and Bank B offers 4.35%, does the finance team really recheck that every morning for 15 basis points? Is it worth logging in again, comparing counterparty risk, initiating the transfer, considering settlement timing, and then doing another round of reconciliation, while making sure payroll and supplier payments the next day are not affected?
Most of the time, the answer is no.
The money stays where it is.
The article calls this one of the system’s under-discussed assets: human inertia.
The Dallas Fed paper says operational deposits are sticky in part because real-world friction prevents funds from being quickly reallocated. If instant settlement, tokenized deposits and AI agents all mature together, then customers seeking better yield could, in theory, move funds with almost no manual effort.
AI will not decide that 15 basis points is too much trouble. It does not have a Monday morning meeting. It does not forget. It does not postpone the task because the day got busy. For an agent, the marginal attention cost of continuously comparing financial products is close to zero.
That does not mean all corporate deposits will behave like hot money and switch banks every five minutes. Customer relationships, credit lines, regulation and risk management do not disappear, and switching will never become truly costless.
But the point is that the system does not need to reach that extreme to affect banks.
Using roughly $17 trillion in relevant deposits in the U.S. banking system, the Dallas Fed paper ran a sensitivity analysis. If the average holding period of deposits falls by 10%, the system’s ability to bear maturity transformation could shrink by about $580 billion in 10-year equivalent risk exposure. If deposit rates become 10% more sensitive to changes in market rates, banks’ capacity to bear duration risk could decline by about $700 billion.
The article stresses that this does not mean AI would suddenly cut U.S. bank lending by $700 billion. Reading that number as a direct drop in lending capacity would be inaccurate.
What it does show is something else: deposit stickiness has enormous economic value in its own right.
If programmable money and AI agents weaken that stickiness even modestly, banks’ funding costs and liability duration structure could change with it.
Competition may move from distribution to routing
In earlier discussions of programmable money, attention usually centered on the benefits: faster settlement, less reconciliation, lower cost, round-the-clock operation and the ability to embed payment conditions directly into transaction logic.
There is another side that receives less emphasis. Once money becomes easier to move, the person or firm holding it gains more choice.
For asset holders, that is efficiency. For institutions that depend on those assets as a stable funding source, it means sharper competition.
If banks want money to operate more like software, they may also have to accept that money starts comparing prices at software speed.
The article then pushes the argument one step further. For years, finance has looked like a distribution business.
Banks compete to be where salaries land. Card companies compete to become the most-used card in a consumer’s wallet. Brokerages compete for space on the investor’s phone screen. Wealth managers compete for customer relationships. Fintech companies have spent tens of billions of dollars on customer acquisition in order to control the same thing: the position between the user and the financial product.
Whoever owns distribution gets a chance to decide which product is seen.
AI agents could gradually shift that competition from distribution to routing.
The old question was: which bank do I like best?
A more important future question may be: which banks are on my treasury agent’s acceptable counterparty list?
Fund managers used to fight for a money market fund’s position on an app’s home page.
In the future, an agent may simply decide that if the yield spread exceeds 18 basis points, liquidity requirements are met and counterparty risk stays below a threshold, funds should flow automatically to a given destination.
Lenders used to compete on advertising, brand and sales channels.
An agent may instead compare actual borrowing rates, collateral requirements, prepayment terms and a company’s current cash-flow profile, then determine whose capital to call.
At that stage, what matters is no longer just whether a user downloaded your app. What matters is whether you are inside the machine’s default rule set.
Whoever is written into the authorization scope earns the right to compete. Whoever wins the routing decision wins the flow of funds.
In the article’s framing, agentic commerce changes where demand flows. Agentic finance changes where capital flows.
The longer-term question for banks is about the next dollar
If the speaker today were a bank CEO, AI’s ability to cut the time analysts spend writing reports or automate KYC and anti-fraud investigations would obviously matter. Those are practical efficiency gains.
But the longer-term question may be another one entirely: when a corporate client has a treasury agent working 24 hours a day, why should the bank still get that client’s next dollar?
The article suggests the smartest banks may not resist this change. They may be the first to put agents inside their own systems. A client’s operational cash can remain in checking accounts, idle balances can move automatically into higher-yield deposits or funds, and liquidity can be brought back when needed. Foreign exchange, credit lines, collateral and payments can be optimized in the same framework. In that setup, even if an agent keeps searching for a better answer, the money may still stay inside the same bank’s ecosystem.
From that perspective, experiments such as Project Agorá may not weaken banks. They may even help regulated commercial bank money retain a central role in the age of stablecoins.
So this is not a simple story in which AI and crypto together eliminate banks. The business models facing real danger may be those that have long treated customer reluctance to compare, switch or deal with hassle as a moat.
In the past, one major advantage for a bank was that leaving was inconvenient. In the next phase, the strongest banks may need to prove the opposite: customers can leave at any time, but after the AI runs the numbers, it still keeps the money here.
For more than a decade, fintech’s fiercest battles played out on phone screens. Companies fought over daily active users, primary accounts, the most-used card and proximity to the customer.
The next round of major financial battles may have no interface at all.
They may play out in a set of authorization rules running in the background, in an approved counterparty list, in a yield gap measured in basis points, and in the instant when AI decides where a given dollar should sit.
That is why the real question in agentic finance may never have been when AI will manage money for us.
It may be this instead: as more money starts searching automatically for its best destination according to rules machines can understand, who still gets to decide where that money ultimately stays?