When AI Starts Moving Money- Who Is Actually in Control
Why AI agents are changing how banks manage authority, permissions and control
Artificial intelligence is already becoming part of everyday banking. It can analyse documents, detect unusual transactions, support customer interactions and help employees navigate increasing volumes of information.
But most of these applications still share one characteristic: AI helps people understand what is happening or decide what to do next.
Agentic AI changes that.
AI agents could go beyond analysing information and recommending actions. They could prepare payments, coordinate liquidity management and potentially initiate financial transactions.
For Business and Corporate Banking, this raises a fundamental question:
What is AI actually authorised to do?
From AI assistance to AI action
A lot of the current value of AI in banking comes from helping humans manage complexity.
In corporate credit, AI can analyse financial statements, contracts, risk information and customer data before a decision is made.
In trade finance, it can extract information from letters of credit, invoices and shipping documents, compare different sources and identify discrepancies.
In KYC and financial crime prevention, it can process documents, cross-reference information and surface findings for specialists.
The common theme is not necessarily replacing human decisions. It is helping humans navigate information and processes that span documents, systems and organisational boundaries.
Agentic AI takes this a step further.
While a chatbot might answer a question about an account balance, an AI assistant could investigate why a company's liquidity has changed. An AI agent, however, could work towards a broader objective, such as ensuring sufficient liquidity for an upcoming payroll.
This shift becomes particularly relevant in treasury management.
Imagine an AI agent with access to bank balances, expected cash flows, payment obligations and available credit facilities.
A treasurer could give the agent a simple objective: “Manage today's liquidity.”
The AI could identify an expected liquidity shortage and recommend postponing selected payments. It might even prepare those payments for human review.
But how far should its authority extend? Could it decide which account should make a payment, move liquidity between accounts, draw on a credit facility or release a payment itself?
This scenario is closer to reality than it might sound. In 2025, researchers tested a generative AI agent performing high-level intraday liquidity management in simulated wholesale payment scenarios. The agent maintained precautionary liquidity buffers, prioritised urgent payments and balanced liquidity use against settlement delays. At the same time, the researchers emphasised the need for safeguards, human oversight and further research. (BIS, 2025)
The question is therefore no longer purely theoretical.
How much autonomy should AI agents get?
Not every AI agent needs the same level of authority.
There is a fundamental difference between an agent that analyses account balances, one that prepares payments for human approval and one that can execute transactions within predefined boundaries.
As autonomy increases, so do the requirements for permissions, oversight and accountability. The question is no longer just what an AI agent can do, but what it is authorised to do.
Intelligence is not authority
Traditional Corporate Banking is built around financial authority.
A person has an identity. That identity is connected to accounts, roles and permissions. Mandates, transaction limits, signature classes, approval workflows and four-eyes principles determine what that person may actually do.
An AI agent complicates this model.
Imagine a CFO defines an objective. The AI determines that a payment is required, selects an appropriate action and prepares the transaction.
Potentially, it also initiates it.
But under whose authority does that happen?
The person who defined the objective? The organisation operating the agent? The individual who configured it? The technical identity assigned to the agent?
AI can understand context, analyse information, compare alternatives and orchestrate processes.
But the fact that an AI system can determine an appropriate financial action does not mean that it should automatically be authorised to perform it.
The distinction is crucial: AI can determine what could or should happen. But banking control infrastructure must determine what is actually permitted to happen.
This suggests a clear separation between two layers:
AI layer
AI understands, analyses, recommends, prepares and orchestrates.
Trusted banking control layer
Trusted banking infrastructure enforces permissions, validates actions, applies approval requirements, controls execution and records what happens.
In other words, intelligence and authority serve different purposes. The AI can interpret intent and determine an appropriate course of action. The banking control layer determines whether that action is authorised.
Give AI capabilities, not unlimited access
That distinction also changes how banks should think about providing AI agents with access to banking services.
Rather than giving an agent broad “online banking access”, banks could expose individual capabilities according to the task the agent needs to perform.
Each capability can be subject to specific permissions, transaction limits and approval requirements. An agent authorised to analyse liquidity does not need permission to move money. And an agent authorised to initiate predefined low-value transactions should not automatically be allowed to execute a €10 million cross-border payment.
That turns access control from a supporting technical function into a central part of the agentic banking architecture.
What this means for banks
As AI moves from analysing information towards taking action, the architecture around the AI becomes increasingly important.
The question is not only what the AI can do, but whether banks can demonstrate that they remain in control of what it does.
This is also increasingly reflected in the supervisory discussion. In February 2026, Pedro Machado of ECB Banking Supervision noted that more than 85% of large European banks under its supervision already use AI in some form. As generative and agentic AI expand, the ECB is emphasising clear accountability for AI-driven decisions, effective senior-management oversight and challenge from risk management, compliance and internal audit. (ECB Banking Supervision, 2026)
For agentic banking specifically, this translates into concrete design questions:
- Which capabilities, accounts and transaction types can an agent access?
- Which permissions, financial limits and approval requirements apply?
- When must a human intervene?
- How are agent actions monitored, recorded and audited?
- How can access be modified or revoked?
These are not entirely new questions.
Corporate Banking has been dealing with identity, mandates, entitlements, transaction limits, approvals and auditability for decades.
AI does not make these control mechanisms obsolete. It makes them more important.
The technology may be new, but many of the fundamental control questions are remarkably familiar: Who is acting? What are they allowed to do? Who authorised them? What limits apply? Does somebody else need to approve the action? And can the bank reconstruct what happened afterwards?
The challenge for banks is therefore not simply to connect AI agents to existing banking services. It is to ensure that those services can distinguish between what an agent is capable of doing and what it is authorised to do.
Because in agentic banking, the ability to act must never be confused with the authority to do so.
Interested in exploring this topic further?
Next Thursday, we'll be taking a closer look in a new episode of our podcast coco:cast, exploring what it means for corporate banking. Subscribe today so you don't miss it!