For banks and non-banking financial companies (NBFCs), the next phase of AI adoption will not be determined by how many processes can be automated, but by how much decision-making authority institutions are willing to delegate to machines. In a sector where errors can affect customer data, credit outcomes and financial exposure, the distinction between automating tasks and automating decisions is becoming critical.
Krupesh Bhat, Founder and CEO of Melento, believes financial institutions will approach agentic AI incrementally. Rather than attempting to automate entire lending or onboarding journeys at once, organisations are likely to begin with bounded operational tasks, assess their reliability and expand their use as governance frameworks mature.
This reflects a broader reality for enterprise technology leaders: the path from AI experimentation to production requires more than model performance. It demands clarity on accountability, oversight and the consequences of an incorrect action.
Beyond conventional workflow automation
Traditional automation follows predefined sequences of tasks. Agentic AI introduces a different model: a system receives an objective and determines the steps needed to achieve it. This can make it useful in processes involving multiple documents, information sources and decision checkpoints.
In financial services, potential applications include extracting information from customer documents, verifying details, compiling due diligence findings and identifying discrepancies for review. These activities often consume significant employee time before an underwriter, compliance officer or operations team can make a decision.
The opportunity is to reduce this preparatory workload without assuming that the final decision should also be automated.
For Bhat, this distinction is central to how enterprises are approaching adoption. AI can gather and organise information, highlight exceptions and give employees a more complete view of a case. Whether it should act independently depends on the process, the institution’s risk appetite and the consequences of getting it wrong.
This also changes how technology leaders should assess automation opportunities. The relevant question is not simply whether a task can be performed by an agent but whether its output can be validated and its actions constrained within an acceptable risk boundary.
Why adoption will be gradual in BFSI
Banks and NBFCs operate under regulatory obligations and established internal controls. Consequently, introducing autonomous systems into sensitive workflows is unlikely to be an all-or-nothing decision.
Bhat expects institutions to take a staged approach, beginning with processes that have relatively limited regulatory implications and expanding into more consequential areas as confidence develops. “Vendor onboarding and document verification, for instance, may offer a more manageable starting point than granting an agent the authority to approve exceptions to credit policy,” he says.
Such sequencing allows technology, risk and compliance teams to understand how a system behaves under real operating conditions before extending its remit.
For CIOs and CTOs, this makes the design of autonomy boundaries a governance decision as much as a technology decision. Organisations need to establish which actions an agent may execute independently, which require approval and which must remain exclusively under human authority.
The approach also has implications for measuring success. “A pilot that performs well under controlled conditions may not demonstrate the reliability required for deployment across thousands of cases. Production readiness depends on how a system handles unusual inputs, incomplete information and exceptions — precisely the situations that can complicate financial workflows,” he elaborates.
The unresolved questions of security and reliability
AI agents introduce risks beyond those associated with conventional rules-based automation. They may generate incorrect information, misinterpret instructions or behave unpredictably when operating with insufficient context or poorly defined boundaries.
Human oversight remains one way to contain these risks, but its effectiveness depends on where it is applied. If employees must recheck every action, the efficiency gains may be limited. If oversight is too weak, errors can propagate through connected workflows.
“The challenge for enterprise architects is to identify the right control points: validate information at critical stages, restrict what agents can access or execute, flag deviations and maintain records that allow decisions and actions to be reviewed,” says Bhat, adding that security assessments are equally important when AI capabilities are introduced through third-party platforms. Financial institutions must evaluate how sensitive information is handled, how access is controlled and whether the technology fits within their existing security and infrastructure requirements.
These considerations make governance a prerequisite for scaling agentic AI, rather than a layer to be added after deployment.
Redesigning operations, not just automating tasks
The most immediate opportunities may lie in operational processes that involve repetitive information handling, manual verification and coordination across teams. Customer service is another area where organisations are seeking faster responses and more consistent experiences.
However, automating individual steps does not necessarily resolve the underlying inefficiencies of a process. Enterprises also need to reconsider how information moves between systems, how exceptions are escalated and where human judgement adds the greatest value.
For technology leaders, this means evaluating AI against operational outcomes: reduced manual effort, faster processing, consistent handling of exceptions and the ability to explain how a system arrived at its output.
The longer-term question is how far these systems can move from assisting employees to executing complete processes. In regulated industries, that transition is likely to be uneven, shaped by the sensitivity of each workflow and the institution’s ability to demonstrate control.
Agentic AI may eventually change how enterprise operations are organised. But in BFSI, progress will depend less on pursuing autonomy for its own sake and more on establishing where autonomy is justified, where human judgement remains essential and how the two can work together without weakening accountability.