A banking chatbot can give a customer the wrong answer. An autonomous banking agent could do something with that answer.
The distinction is becoming harder for financial institutions to ignore as AI moves beyond conversational interfaces towards systems capable of initiating transactions, interpreting instructions and carrying out routine tasks. The promise is a banking experience with fewer steps and less friction. The complication is that every additional degree of autonomy introduces another question about authority, accountability and the limits of automation.
For banks, this is not simply a technology problem, it strikes at the foundations of an institution whose business depends on controlling the movement of money, establishing the legitimacy of instructions and maintaining a clear chain of responsibility.
The challenge is particularly relevant to agentic AI, which moves beyond generating responses or recommendations to initiating actions. In an exclusive interaction with Express Computer, Sanjay Mudaliar, Executive Director, Bank of Baroda, identifies this distinction as central to the bank’s approach.
“Agentic AI represents an important evolution because it can initiate actions rather than simply generate responses or recommendations,” he says. The observation captures both the attraction and the difficulty of the technology. An AI system that answers questions can improve access to information. One that acts on a customer’s behalf must also be trusted with the consequences of its decisions.
Convenience is becoming a question of control
Consider the evolution of a banking application. The first generation digitised transactions that customers once completed at a branch, subsequent iterations made those transactions accessible through mobile interfaces, conversational AI seeks to simplify the process further by allowing customers to interact with banking services through natural language.
The next proposed step is more consequential. Mudaliar envisages AI-enabled agents that could verify and pay routine utility bills on a customer’s behalf.
The appeal is obvious. Customers would spend less time navigating menus and executing repetitive tasks. But the apparent simplicity of the transaction conceals a more complicated set of responsibilities. An agent would need to interpret an instruction correctly, operate within authorised permissions and execute the intended action without exceeding the customer’s mandate.
What happens when an instruction is ambiguous, the information available to the system is incomplete or the proposed action differs from what the customer intended? The interaction does not establish a bank would resolve these scenarios. They are, however, the questions that any move towards delegated financial actions must confront.
The issue is not whether such systems can be made useful. It is whether convenience can be delivered without making the exercise of financial authority less transparent.
Mudaliar says the bank’s approach is based on risk-calibrated autonomy. The greater the potential impact on customers or business outcomes, the stronger the oversight and control required. In credit appraisal, for example, AI can consolidate information and generate insights, but final decisions remain with authorised bank officials.
The principle draws a boundary between analytical capability and decision-making authority. As AI becomes more capable, maintaining that boundary may prove more difficult than establishing it in policy.
The distance between a working model and a dependable system
The problem begins well before an AI agent executes a payment. It is already evident in the transition from a promising experiment to an operational system.
Mudaliar acknowledges that Bank of Baroda has recalibrated aspects of some generative AI initiatives, including how responses are generated, how sensitive information is handled and how solutions integrate with core banking systems.
He describes a review process spanning business, technology, security, compliance and data governance. Such scrutiny can lead to changes in system design, data access, monitoring arrangements and the degree of human oversight.
The significance lies in the admission that technical capability is only one part of the deployment equation. A system may work under controlled conditions and still require substantial adjustment before it can operate within the complexity of a bank’s existing infrastructure and obligations.
“A successful proof of concept demonstrates that technology works. The big step is ensuring that the solution is scalable, secure and compliant for enterprise-wide deployment,” Mudaliar says.
The distinction is especially important with generative AI, whose outputs must be assessed not only for usefulness but also for accuracy, consistency and their suitability for the context in which they are used. Integration with core banking systems adds another layer of complexity because the consequences of an error can extend beyond an unsatisfactory response.
Can millions of interactions settle the question of value?
Bank of Baroda’s customer-facing AI systems, including ADI, ADITI and the Virtual Front Office, have handled more than 8.3 million interactions, according to Mudaliar. Its generative AI-enabled employee knowledge platform, GyanSahay.AI, handles more than 4,800 employee interactions daily.
The bank also deploys AI in areas such as early warning systems, collections and risk management. Mudaliar identifies conversational AI as the area delivering the most tangible impact so far, citing improvements in service delivery and efficiency.
These figures indicate the reach of the bank’s AI deployments. They do not, on their own, establish how effectively those systems resolve customer problems, reduce operating costs or improve the quality of decisions.
That is a critical distinction in a market increasingly inclined to use adoption figures as evidence of transformation. An interaction is a measure of activity. Its value depends on what happens as a result.
The same distinction applies to personalisation. Bank of Baroda’s bob World 2.0 combines conversational banking with personal financial management and tailored digital experiences. Mudaliar sees conversational banking as particularly consequential, including features such as Tap-to-Talk for payments.
But making an interface more intuitive is not the same as helping a customer make a better financial decision. The latter requires a more demanding assessment of relevance, accuracy and outcomes.
The next phase of enterprise AI will have to be judged against these distinctions. Without evidence of measurable improvements, the volume of automated activity remains an incomplete account of the technology’s business value.
Why accountability cannot be automated away
For banks, the question of who remains responsible becomes more pressing as AI systems are given greater discretion. Human oversight, audit trails and authorised data access are therefore not merely safeguards surrounding agentic AI. They are central to determining whether the technology can be deployed responsibly.
Mudaliar’s insistence that authorised officials retain final responsibility for credit appraisal reflects this concern. His account also suggests that governance must evolve alongside the technology, rather than be treated as a one-time approval before deployment.
The unresolved challenge is operational. Controls must be effective not just in principle but under the conditions in which systems encounter incomplete information, conflicting instructions and situations that fall outside their intended use cases.
For Bank of Baroda, the ambition is to expand AI across customer service, credit, risk and fraud management over the next two to three years while maintaining human oversight. The wider significance of that ambition lies in the choices the bank and its peers will have to make about where automation ends and human authority begins.
The banking industry’s AI transition will not be defined simply by how convincingly machines can imitate a conversation or how many routine tasks they can complete. It will be defined by whether institutions can demonstrate that automated actions remain authorised, traceable and accountable.