For most of the history of digital banking, the customer remained the decision-maker.
Technology changed how instructions were given. Banking became faster, cheaper and more convenient. Branch visits became web sessions. Web sessions became mobile taps. Paper forms became APIs. Payments that once took days began to happen almost instantly.
But the underlying relationship remained remarkably familiar: the customer decided what should happen; the system executed the instruction.
A person chose the account, the investment, the amount to transfer and when to move the money. Software made those decisions easier to communicate and process, but it did not fundamentally change where the decision originated.
In 2026, I increasingly began to think that this was no longer the most useful way to understand what was happening in banking.
The important change was not simply that banks were adopting artificial intelligence. Technology was moving closer to the point at which financial decisions themselves are formed.
AI could answer questions. Then recommend. Then prepare actions. Increasingly, the industry began discussing systems capable of executing tasks, coordinating workflows and acting with greater autonomy.
Different developments. One structural question.
I initially treated these as separate developments. Increasingly, I don't. They are all Control Point problems.
The Control Point
I have come to think of the underlying question as the Control Point: the place between a financial intention and a financial outcome where the decisive choice is made.
Who controls what happens next?
For a long time, the answer was relatively easy: the customer did. A bank, comparison site, adviser or advertisement might influence the choice, but ultimately the customer selected the product and issued the instruction.
The arrival of increasingly capable AI creates a different possibility.
The difference is where the decision has already been made.
The customer has already chosen the action, amount and destination.
Primary problem: executionThe outcome still depends on product, institution, amount, timing, risk and constraints.
Primary problem: decision + authorityOnce software begins interpreting the objective rather than simply executing a predefined instruction, the nature of digital banking changes.
Digital banking digitised the instruction.
Agentic banking begins to interpret the intent.
AI-native banking redesigns the institution around the consequences.KEVIN MOSERI
Intent is not authorization
The ability to understand what a customer wants does not automatically create the right to act.
An AI system might identify a better savings product, calculate how much money could be moved without compromising expected expenses, and even prepare the transaction. None of that necessarily means it has authority to complete it.
Three lines I keep coming back to.
A system may understand, recommend or technically execute while policy, product design or regulation still keeps consequential authority with the customer.
Even execution is not the end of the question. A customer approving one prepared action has not necessarily delegated continuing decision-making authority to the system.
So once AI enters consequential financial workflows, it is no longer enough to ask how capable is the AI? We also have to ask: what is the AI actually allowed to do?
That reaches beyond technology into permissions, governance, identity, authentication, regulation, risk management and ultimately the architecture of the banking relationship itself.
Why AI maturity is no longer enough
Much of the discussion around AI in financial institutions focuses on maturity: talent, infrastructure, models, investment, deployment and use cases. These are useful questions, but they describe capability.
A bank can have extraordinarily sophisticated AI while keeping every consequential customer action behind explicit human approval. Another institution might have less glamorous technology but allow a tightly controlled agent to perform specific actions inside a bounded mandate.
Those are fundamentally different systems. The difference cannot be captured simply by asking which organisation has “more AI.”
The more useful question is where intelligence sits in relation to authority. And once you look at banking through that lens, the Control Point appears far beyond AI.
Distribution is also a Control Point
Financial products have always depended on distribution. What changes if an intelligent intermediary becomes the place where the customer expresses the objective?
Product discovery can move before the bank ever sees the customer.
The customer may still retain final approval. But the product-discovery process has moved. The agent can influence which options are seen, which institutions are considered and which trade-offs are emphasised.
A bank can build excellent internal AI and still lose the external Control Point if another system becomes the place where customers express their financial objectives.
This is why machine readability becomes strategically important. If an agent is doing the choosing, products, eligibility, fees, rates, restrictions and terms need to be intelligible not only to people but to machines.
If the agent chooses the product, machine readability becomes distribution capability.KEVIN MOSERI
That is not simply an SEO issue. The institution that provides the product may no longer control product discovery, and the institution that controls the account may no longer control the customer interface.
The same problem appears elsewhere
Once I started looking for the Control Point rather than for a particular technology, several apparently separate areas of banking began to line up.
Different surfaces of the same control problem.
Who owns the decision environment when money can move quickly?
Who coordinates payments, cards, accounting and treasury workflows?
Who can see enough of the customer to interpret the objective?
Who can connect, initiate, authenticate, settle and intermediate?
Deposits made the relationship visible
Deposit competition looks like a pricing problem: one bank raises a rate, another launches an acquisition campaign, customers move money and banks respond.
But the deposit account is also an interface. It gives the institution a place to observe behaviour, communicate, introduce other products and become part of the customer's financial routine. Beneath the rate competition is a question about who owns the relationship and the opportunity to influence what happens next.
Business banking made the operating layer visible
A traditional business account is a financial product. Modern business banking increasingly sits inside a larger operating system: payments connect to accounting, cards to employees, accounts to invoicing, bank feeds to bookkeeping and APIs to company software.
The institution no longer merely holds the money. It can occupy the place where financial activity is coordinated. Add AI and the question becomes: does the system show, recommend, prepare, execute—or act repeatedly inside a mandate?
Data and infrastructure define what is possible
A useful financial agent needs context: income, expenses, balances, transactions, obligations and preferences. Open Finance can make more of that context accessible beyond the institution holding the account.
Infrastructure then determines what an intelligent system can actually do with that context: who can connect, initiate, authenticate, settle and intermediate.
An agent without access is an adviser. With data but no execution capability, it is a recommendation system. With execution access but transaction-by-transaction approval, it is something else again. A standing mandate would represent another category entirely.
The intelligence may be similar. The Control Point is not.
The question eventually became measurable
By the middle of 2026, I realised the same question kept appearing in work that had initially seemed unrelated: business banking, deposits, acquisition, privacy and financial data, AI, infrastructure and distribution.
The common issue was not simply digitisation. It was the movement of decision power through the financial system.
The Control Point became the lens I use to understand that shift.
Kevin identifies the structural shift. Torinel® measures it.
Ideas, interpretation and the question of where decision power moves.
Research, methodology and measurement in German banking.
Torinel® is where parts of that lens are now being turned into observable research. It is a German financial-intelligence publication and research institution examining how technology, regulation and competition are changing banking and financial decision-making.
In 2026, that work led to the Torinel AI Banking Index™: an evidence-based attempt to distinguish between what banking AI can do, what it is permitted to decide or execute, and what control architecture surrounds that authority.
Torinel® AI Banking 2026 — foundational research →
Torinel AI Banking Index™ →
The future of banking will not be determined solely by which institutions build the most capable models. It will also depend on where those models sit in the chain between intention and outcome.
What I think matters next
I do not think every financial decision will suddenly be delegated to an AI agent, nor that banks are about to disappear behind autonomous interfaces. Financial systems change more slowly than technology narratives suggest. Regulation, trust, liability and security matter.
But even partial movement of the Control Point could be significant.
An agent does not need complete autonomy to reshape distribution; it only needs to become influential enough in product discovery. It does not need total authority to change banking workflows; it only needs to begin preparing consequential actions. And it does not need to replace the bank to weaken the bank's control of the customer relationship; it only needs to become the environment in which the customer decides what to do next.
That is why I think the most interesting question in financial technology is becoming less about what AI can do.
The more important question is:
Who controls what happens next?KEVIN MOSERI
That question connects technology to distribution, distribution to infrastructure, infrastructure to permissions, permissions to authority, and authority back to the design of the financial institution itself.
For most of digital banking, software waited for the customer to decide. The next era may begin when software starts participating in the decision.
And once that happens, control becomes the architecture.