By Shashank Pattekar, International Banker
Artificial intelligence (AI) is already widely embedded across banking, with predictive models currently used to classify transactions, detect anomalies, generate summaries and support decision-making. More recently, a new phase of automated support has emerged in the form of agentic AI, which is radically changing how work is organised inside banks and how important decisions are made over time.
AI agents are programmed to execute tasks and workflows to achieve certain objectives. While humans define the initial goals, agentic systems then plan, execute and iterate across multi-step tasks with limited human intervention to achieve those goals. “They take inputs, reason, decide on tasks to perform, interact with other agents and tools, review their outcomes and determine the next steps required,” Accenture explained. “The agents ‘know’ a domain context in which they have been trained within an organization (e.g. on specific training data and documents). They have long-term memory and can learn from past interactions to optimise decision-making.”
A major upgrade from AI and generative AI, agentic AI architectures aim to coordinate entire workflows. And given that banking is typically process-heavy, with many of its core functions, including lending, compliance, payments and portfolio management, being sequences of interdependent actions, several processes could significantly benefit from agentic systems.
In practical terms, this means that an agent can monitor a stream of data in real-time, decide when certain actions are required, execute those actions across multiple systems and learn from the results, thereby refining and improving its future choices of action. “Every decision, outcome, and customer interaction helps the system refine its models and strategies, making it smarter and more effective,” according to Arturo Rivasainz, chief product officer at digital-banking firm Meniga.
The system can link analysis to execution in a continuous loop, meaning that banking processes that previously took hours or even days to complete—with intermittent human intervention required—can now be completed in minutes. This can be hugely beneficial for improving payments, fraud detection and credit decisions.
Of the broad areas of the industry most likely to be transformed by this technology, retail banking could see a dramatic change in customer interaction resulting from the use of agents. A customer interacting with a bank’s mobile platform, for example, might set a broad objective of increasing savings or reducing debt. An agentic system can then translate that objective into a sequence of actions: reallocating funds, adjusting spending limits and identifying opportunities to refinance liabilities.
Rather than simply suggesting the best actions for us to take, therefore, agentic AI is primarily concerned with the successful execution of those actions.
Rather than simply suggesting the best actions for us to take, therefore, agentic AI is primarily concerned with the successful execution of those actions. A system that understands income patterns, spending behaviours and financial goals can move funds, adjust savings allocations or renegotiate products without requiring repeated user input. This changes the customer’s role—interaction becomes less about initiating transactions and more about setting preferences and boundaries.
“The key is personalisation,” Bryan Zhang and Kieran Garvey of the Cambridge Centre for Alternative Finance (CCAF) explained. “Agentic AI will transform customer interactions by drawing on data permission capabilities enabled by open banking and embedded finance to create personalised AI agents that manage finances, optimise decisions, and align with an individual’s specific habits, goals, and risk appetite. This isn’t just about convenience—it’s about enabling everyday consumers to make smarter and more intuitive financial decisions.”
Compliance and risk management are also expected to be significantly impacted by these agents. Regulatory compliance has traditionally been resource-intensive, relying on periodic reviews and manual processes. Agentic systems enable continuous monitoring, with transactions analysed and anomalies escalated for further investigation in real-time. The agentic system can also gather relevant data, assess risk and generate reports for compliance teams, greatly reducing the lag between detection and response. And instead of compiling reports retrospectively, banks can maintain continuously updated datasets to improve accuracy and reduce operational burdens.
In capital markets, meanwhile, the focus is on execution and optimisation. Trading desks already use algorithmic systems, but these are typically constrained to specific strategies. Agentic systems can operate across strategies, adjusting positioning, managing risk and responding to market conditions in a more integrated manner.
As for lending and credit, the implications are more structural. Loan origination has historically involved multiple stages, including application, verification, risk assessment, approval and monitoring. Agentic systems can coordinate these stages to reduce latency and improve consistency. They can also continue to operate after origination, adjusting credit limits or repayment structures in response to changing conditions.
Take the example of a mid-sized business applying for a working-capital facility. The process usually begins with the submission of financial data, transaction histories and supporting documents. An agentic system ingests this information and initiates verification by cross-referencing it with external sources and internal records. Risk assessment is also conducted in tandem, with the system evaluating cash-flow patterns, sector exposure and macroeconomic conditions to generate a dynamic risk profile of the business.
Based on this assessment, the agent proposes a credit structure that includes limits, pricing and covenants, all of which are aligned with both the borrower’s profile and the bank’s risk appetite. The proposal is then executed, with documentation generated, compliance checks completed and funds disbursed, with the system logging decisions at each stage and maintaining audit trails to meet regulatory requirements.
The process does not end at origination. The agent continues to monitor the borrower’s financial position, potentially adjusting credit limits or triggering interventions should risk conditions change. If cash flow deteriorates, for example, the system can initiate restructuring discussions or recommend additional safeguards.
What distinguishes the agentic model from previous AI milestones, therefore, is continuity. The same system that originates the loan manages it throughout its lifecycle, creating a feedback loop that improves decision-making over time.
The efficiency gains and cost savings of deploying agentic AI cannot be ignored, either. “As banks grapple with falling revenues, some are exploring opportunities to use AI for productivity gains. AI’s potential effect on banking will depend on factors such as banks’ ability to become fully agentic and the extent to which banking customers adopt AI to manage their finances,” according to a December 2025 report from McKinsey & Company, which conducted scenario analysis on how AI could reshape banking functions and consumer behaviour.
“In the most likely scenario…consumers will use AI agents as a new banking channel for some financial matters but will continue to value interactions with banks,” the consulting firm added. “This moderate adoption could enable banks to reshape functions and achieve cost reductions of 15 to 20 percent.”
Nonetheless, the deployment of agentic AI introduces distinct challenges for banks and regulators. As with other AI-based architectures, the problem of algorithmic bias is particularly acute in the banking industry, in which accuracy is essential and errors can be costly.
“Agentic AI banking systems make decisions that can affect everything from loan approvals to fraud investigations. If trained on biased or incomplete data, these systems may reinforce harmful patterns,” according to Salesforce, which provides AI-agent solutions for banking. To minimise this risk, therefore, Salesforce recommended that banks should “use diverse and representative datasets; invest in explainable AI models that can justify decisions, and implement oversight mechanisms, including regular audits and fairness testing”.
Control of the deployed agents becomes another central issue. Systems that can act autonomously must operate within clearly defined boundaries. Ensuring that those boundaries are respected at all times requires banks to have robust governance frameworks established. Errors in autonomous systems can also propagate quickly, particularly when actions are executed at scale. Managing this risk, therefore, requires both sufficient technical safeguards and organisational oversight to be firmly in place.
Ultimately, the shift from assistance to execution under agentic AI enables systems to take action autonomously. In a sector defined by complex, multi-step processes, this capability has far-reaching implications, not least from a regulatory perspective. That said, it is clear that banking is moving towards a model in which decision-making and execution are increasingly integrated, with agentic systems highly likely to be at the heart of that process. As such, those institutions that can quickly adapt to agentic AI and effectively manage the risks that come with it will be best-positioned to benefit from this cutting-edge technology.