Not long ago, the biggest debate in banking was whether chatbots could answer customer queries without frustrating clients. Today, that debate seems almost quaint. A far more consequential transformation is underway behind the closed doors of bank boardrooms. Artificial intelligence is no longer merely assisting employees; it is beginning to make financial decisions, execute transactions, investigate suspicious activities, and optimise capital allocation with little or no human intervention.
This shift marks the arrival of Agentic AI—a new generation of autonomous software capable of planning, reasoning, and acting independently to achieve defined goals. Unlike conventional AI systems that wait for human prompts, agentic systems can decide what information to gather, which software tools to use, and what sequence of actions to execute before completing a task. For the global financial industry, this promises unprecedented efficiency. It also raises one uncomfortable question: when autonomous software makes a costly mistake, who is legally responsible?
The answer remains surprisingly unclear.
Banks have compelling reasons to embrace this technology. According to global industry estimates, financial institutions spend well over US$600 billion annually on technology, with artificial intelligence becoming one of the fastest-growing areas of investment. Rising compliance costs, persistent cyber threats, complex regulations, and customer expectations for real-time services have pushed banks to automate almost every operational process possible. Studies by leading consulting firms estimate that AI could generate hundreds of billions of dollars in additional annual value for the global banking sector through better productivity, fraud prevention, and operational efficiency.
Yet the latest wave of AI differs fundamentally from the automation banks adopted over the past decade.
Traditional generative AI writes reports, drafts emails, summarises documents, or produces computer code after receiving instructions from a human user. The human still decides what action to take.
Agentic AI changes that relationship. Instead of asking software to prepare an analysis, a bank might simply instruct an autonomous system to minimise cross-border settlement costs while complying with liquidity regulations across multiple jurisdictions. The software can independently retrieve market data, evaluate foreign exchange rates, choose payment channels, execute transactions, verify compliance, and continuously adjust its strategy as conditions evolve.
In lending, autonomous AI agents are beginning to replace periodic credit reviews with continuous risk assessment. Instead of waiting months to evaluate a borrower’s financial health, these systems analyse real-time transaction data, cash flows, payroll deposits, tax records, and open banking information to recommend dynamic changes in credit limits almost instantly.
Financial crime detection is undergoing a similar transformation. Anti-Money Laundering (AML) teams traditionally spend countless hours examining suspicious transactions and preparing regulatory reports. Increasingly, AI agents can identify unusual behavioural patterns, investigate customer histories, prepare Suspicious Activity Reports (SARs), and recommend immediate account restrictions before a compliance officer even opens the file.
For banks, the commercial attraction is obvious. Decisions that once required days can now be completed in minutes—or even seconds.
But speed creates its own risks.
Modern banking regulation has always rested on one fundamental assumption: somewhere in every important decision-making chain, there is a human being who can be held accountable.
Agentic AI weakens that assumption.
Suppose an autonomous lending system systematically rejects mortgage applications from applicants living in particular neighbourhoods because historical data incorrectly associated those locations with higher default risks. Customers suffer discrimination, but the bias was neither explicitly programmed nor consciously intended.
Who should regulators hold responsible?
The bank could argue that it relied on a commercially licensed AI platform developed by an external technology company. The software developer may respond that its model functioned exactly as designed and merely followed objectives established by the bank. The AI system itself cannot bear legal liability because software has neither legal personality nor moral agency.
This emerging accountability gap represents perhaps the greatest governance challenge facing modern finance.
The risks extend beyond individual lending decisions.
Imagine autonomous trading agents operating simultaneously across competing financial institutions. Each algorithm seeks to maximise returns by responding instantly to changing market conditions. Without communicating with one another, several systems could independently adopt similar strategies during periods of market stress, triggering rapid asset sales, draining liquidity, and amplifying volatility across financial markets.
No trader intended market manipulation. No executive issued unlawful instructions. Yet the collective outcome could resemble coordinated behaviour.
Financial law has little experience dealing with algorithmic coordination that emerges without human conspiracy.
Unfortunately, regulation is struggling to keep pace.
The European Union’s AI Act has established the world’s first comprehensive framework for regulating high-risk artificial intelligence systems, while international banking standards under the Basel framework continue strengthening operational resilience and risk governance. Both represent significant progress. Yet they were largely conceived for AI systems that support human decisions—not autonomous agents capable of adapting their behaviour continuously after deployment.
Even banks’ own governance structures are beginning to show signs of strain.
For decades, financial institutions have relied upon the “Three Lines of Defense” model: business operations, independent risk management, and internal audit. This framework assumes that operational risks evolve slowly enough for human oversight to remain effective.
Autonomous AI changes the timeline entirely.
An AI agent can execute thousands of financial decisions every second. Compliance teams, however, investigate issues over hours or days. By the time auditors identify problematic behaviour, millions of transactions may already have been completed.
That mismatch demands an entirely new regulatory philosophy.
Instead of focusing solely on model accuracy, supervisors should require “Explainability by Design.” Every significant autonomous financial decision must generate secure, tamper-proof records explaining which data were considered, why a particular action was chosen, and whether alternative options were rejected.
Equally important are mandatory circuit breakers. Just as stock exchanges automatically halt trading during extraordinary volatility, autonomous banking systems should immediately suspend high-risk operations whenever predefined safety thresholds are crossed. Human supervisors must retain the unquestioned authority to override autonomous decisions before systemic damage occurs.
The objective should not be to slow innovation but to ensure that innovation remains governable.
There is little doubt that autonomous AI will transform banking over the next decade. Institutions that successfully deploy these technologies will improve customer service, reduce operational costs, strengthen fraud detection, and allocate capital more efficiently than ever before.
However, history offers an important lesson. Every major financial innovation—from complex derivatives to high-frequency trading—initially promised efficiency before exposing weaknesses in governance. Agentic AI is unlikely to be any different.
Banks may increasingly delegate execution to intelligent software, but they cannot delegate accountability. Responsibility for financial stability will continue to rest with boards of directors, regulators, auditors, and policymakers—not with algorithms.
The boardrooms of tomorrow may indeed become quieter as autonomous systems assume more operational authority. But public trust in banking has never depended on how intelligent financial institutions become. It depends on whether someone remains answerable when intelligence fails.
That is the real risk the financial sector must confront before autonomous software is allowed to manage the world’s money.
(The Author is Assistant Professor, Department of Commerce, Guru Ghasidas Vishwavidyalaya, Bilaspur, C.G.)
