Banking is moving from artificial intelligence that predicts and recommends to artificial intelligence that can plan, decide and act. The technological leap is significant. The governance challenge may be larger.
For most of the digital era, banking software has been obedient. A payment engine moved money because somebody instructed it to. A credit model produced a score for a human or rules engine to consider. Fraud software raised an alert. Treasury systems calculated liquidity positions. Chatbots answered questions. Agentic artificial intelligence changes that relationship: an AI agent can interpret a goal, examine information, decide what should happen next, call external systems, perform a sequence of actions, check the result and adjust its behaviour. Bank of England/FCA
Connect that capability to a bank’s payment infrastructure, customer accounts, credit systems or treasury platforms and the distinction between “software that advises” and “software that operates the bank” begins to disappear. The emerging model is best understood as bounded autonomy: software receives permission to make increasingly consequential decisions inside tightly defined limits, while humans remain responsible for designing those limits and dealing with exceptions.
Banking AI Is Already Widespread. Full Autonomy Is Not.
A joint Bank of England and Financial Conduct Authority survey found that 75% of responding financial firms were already using AI, with another 10% planning adoption within three years. Among international banks responding to the survey, adoption reached 94%. Yet only 2% of reported AI use cases involved fully autonomous decision-making. Some degree of automated decision-making was present in 55% of use cases, but most remained constrained or supervised. Bank of England/FCA, 2024
Europe presents a similar picture of rapid adoption without wholesale automation. In June 2026, the European Central Bank said more than 85% of banks under European banking supervision were using artificial intelligence. ECB, June 2026
That distinction is critical. Banks are not suddenly replacing their operating models with a single intelligent machine. AI is instead spreading through transaction monitoring, fraud detection, coding, compliance, document review, customer support, investment research, payment routing, forecasting and credit assessment. The transition towards agentic banking begins when those systems acquire something earlier generations of banking AI generally lacked: authority to take action.
From Chatbot to Action Engine
One visible example can already be found in retail banking. Malaysia’s Ryt Bank allows customers to initiate banking tasks using ordinary language. A customer can ask the assistant to move money or set up a transfer, while the payment workflow still requires the user to confirm before completion. Ryt Bank
That confirmation step illustrates where much of banking currently sits: AI can understand intent, locate information and prepare an action, while the human retains final authorisation. J.P. Morgan describes the next step in corporate treasury as “policy-bounded execution”: agents acting within predefined thresholds and escalating situations outside them. J.P. Morgan
This is likely to be how autonomy enters banking—incrementally rather than dramatically. A system may first recommend moving liquidity, later move liquidity automatically below a specified amount, and eventually rebalance thousands of positions continuously as long as every transaction remains inside defined treasury, risk and compliance limits.
Payments and Treasury Could Become the First Real Test
A 2025 BIS working paper tested whether a generative-AI agent could perform intraday liquidity-management tasks in a simulated wholesale payment environment. The experiments required the agent to maintain precautionary liquidity buffers, prioritise urgent payments and balance the cost of holding liquidity against settlement delays. The researchers found that the agent could reproduce key cash-management behaviours and concluded that routine liquidity-management tasks could potentially be automated, while stressing the need for safeguards and human oversight. BIS Working Paper 1310
Treasury is unusually attractive for agentic AI because many objectives can be expressed as explicit constraints: maintain a liquidity buffer, avoid a counterparty limit, fund priority payments before a deadline, minimise idle cash and escalate when projected liquidity falls below a threshold. A sufficiently capable agent can observe balances, queues and forecasts continuously and make decisions faster than a human team reviewing those signals manually.
A Financial Decision Is Different From a Recommendation
Risk changes dramatically once an AI system acquires execution authority. If an AI assistant produces a poor summary, an employee may catch the error before acting. If an AI agent incorrectly determines that liquidity should be transferred, a payment may already have occurred. If an automated fraud system freezes a legitimate account or a lending system rejects an applicant, the customer experiences the model’s judgment directly.
The distance between model output and real-world consequence therefore collapses. Traditional governance often assumes software informs a decision somewhere in a process. Agentic systems can compress analysis, decision and execution into one continuous loop. The question becomes not only whether the model is accurate, but what the model is authorised to do when it believes it is accurate.
The Bank Remains Responsible
Delegating a decision to AI does not transfer legal responsibility to the AI. In US lending, creditors using complex algorithms must still provide applicants with specific principal reasons when adverse action is taken; technological complexity does not remove that obligation. CFPB Circular 2022-03
Europe is also moving toward an explicit AI-specific framework. The EU AI Act classifies creditworthiness evaluation of natural persons among high-risk uses. Under the current implementation timetable, high-risk requirements are scheduled to apply from 2 December 2027, while Article 50 transparency obligations began applying on 2 August 2026. European Commission — AI Act
The governance direction is clear: banks may delegate decisions, but they cannot delegate accountability.
The Hardest Problem May Be Correlated Machines
A single badly designed agent can create losses. Thousands of institutions relying on similar agents could create something more serious. The Financial Stability Board has identified third-party dependencies, market correlations, cyber risk, and weaknesses involving model risk, data quality and governance as AI-related vulnerabilities with potential systemic implications. Financial Stability Board
The Bank of England/FCA survey also found that a third of AI use cases were third-party implementations and that the top three providers accounted for 73% of reported cloud providers and 44% of reported model providers. Bank of England/FCA, 2024
This raises the possibility of synchronised behaviour: treasury agents hoarding liquidity at the same moment, trading agents reducing exposure to the same assets, or credit agents tightening lending simultaneously. The systemic risk may not be that machines behave irrationally, but that many machines behave rationally in the same way at the same time.
AI Agents Are Also Cyber Targets With Bank Accounts Attached
A financial agent may possess credentials and access to payment APIs, customer information, treasury systems or internal databases. The FSB identifies cyber risk as a channel through which AI could amplify financial vulnerabilities, while the ECB has warned that newer AI capabilities lower the expertise and time required to discover and exploit vulnerabilities. FSB; ECB
The safer architecture is therefore not simply a smarter model. It is an infrastructure that assumes the model may occasionally misunderstand instructions, encounter hostile data or make incorrect judgments—and prevents those errors from becoming catastrophic actions.
The Autonomous Bank Needs More Rules, Not Fewer
Paradoxically, a bank operated increasingly by AI agents may require more explicit internal rules than a human-centred bank. Payment agents need transaction authorities. Credit agents need defined underwriting boundaries. Treasury agents need hard liquidity and counterparty limits. High-impact decisions need additional verification, and every action needs an auditable record of the data, policy, model, tools and human interventions involved.
The safest architecture is unlikely to be one general-purpose artificial executive with access to everything. It is more likely to resemble a collection of specialised agents, each with narrow permissions and independent control mechanisms.
Regulation Is Still Catching Up With Agentic AI
The difficulty became visible in the United States in April 2026, when the OCC, Federal Reserve and FDIC updated model-risk-management guidance while explicitly stating that generative AI and agentic AI were outside its scope because the technologies were novel and rapidly evolving. The agencies also said further work would consider banks’ use of AI. OCC Bulletin 2026-13
The US Treasury has meanwhile released a Financial Services AI Risk Management Framework, while BIS Innovation Hub Project Noor is exploring tools to help supervisors assess model explainability and robustness. U.S. Treasury; BIS Project Noor
Regulators eventually need the capacity not merely to read banks’ AI policies, but to interrogate how sophisticated models behave in practice.
Humans Do Not Disappear. Their Job Moves Up the Control Stack
Agents could absorb large volumes of information gathering, reconciliation, routine alerts, documentation and approvals. Humans would increasingly define objectives, establish limits, investigate unusual cases and decide when automation should stop. But human oversight can fail through automation bias or through “exception overload”—too many escalations for supervisors to assess meaningfully. J.P. Morgan
Human oversight therefore cannot simply mean placing an approval button at the end of an automated workflow. It must be designed so that intervention is both practical and consequential.
The Economics Could Still Be Transformative
Banks contain countless workflows created around technological limitations that no longer necessarily exist. Agentic software offers the possibility of persistent digital workers that retrieve information through APIs, interpret policy, query databases, prepare documentation and initiate permitted actions. The Bank of England/FCA survey suggests firms expect major future benefits in operational efficiency, productivity and cost reduction. Bank of England/FCA, 2024
Those savings are not free. Agentic systems require high-quality data, secure infrastructure, continuous evaluation, audit mechanisms, model monitoring, cyber defence and specialist staff. They may reduce the cost of executing decisions while increasing the importance of controlling the systems that make them.
Conclusion: The First Autonomous Bank Will Probably Not Look Autonomous
The most realistic autonomous bank is not an institution without bankers. Customers may still speak to employees. Boards will still meet. Risk officers will still set limits. Regulators will still examine the institution. But underneath those visible layers, millions of operational decisions could increasingly be made by software agents.
Banking has spent decades automating transactions. Agentic AI opens the possibility of automating judgment. The winning model is likely to be highly automated, heavily permissioned and relentlessly auditable. The defining question will not be how intelligent the software becomes, but how carefully humans decide what that intelligence is allowed to do.
References
1. Bank of England & FCA — Artificial intelligence in UK financial services: 2024
2. European Central Bank — Strengthening operational resilience for the age of AI
3. BIS — AI agents for cash management in payment systems
5. J.P. Morgan — Agentic AI in Corporate Cash & Treasury Management
6. Financial Stability Board — The Financial Stability Implications of Artificial Intelligence
7. Consumer Financial Protection Bureau — Circular 2022-03 on adverse action and complex algorithms
8. European Commission — Navigating the AI Act
9. OCC — Model Risk Management: Revised Guidance, 2026
10. U.S. Treasury — Financial Services AI Risk Management Framework