BFSI moves from AI pilots to production, but rationalisation must accompany modernisation
For India’s banking sector, the AI conversation is moving beyond pilots and proof-of-concepts. Banks are beginning to put AI-led systems into production across operations, technology transformation, testing, customer service and application modernisation. But as these deployments scale, a different set of challenges is emerging around context, governance, token economics, explainability and the complexity of running multiple technology and AI layers.
Kishan Sundar, Senior Vice President and CTO, Maveric Systems, says the most significant change over the past year is that AI initiatives have moved from experimentation to measurable production outcomes.
“When we last spoke, we were either doing POCs, or we were doing a paid POC or a free POC, or we were doing some pilots,” Sundar says. “Whereas now we can quantify and qualify what has gone well, what we should keep in mind when we are taking it into production, as well as the lessons learned.”
The shift is visible across both banking operations and technology transformation.
From AI pilots to production-grade operations
One example is an agentic anti-money laundering (AML) solution deployed with a regional bank in North America. The system uses three levels of agents to analyse AML alerts, with different degrees of autonomy and human involvement.
At L1, the process has been made autonomous through agents. At L2, human operators remain in the loop and are assisted by AI, while at L3, the AI agent acts as a co-pilot to support more complex activity.
The deployment has reduced the need for L1 intervention while increasing the productivity and judgement capabilities of L2 and L3 teams. It is also helping the bank generate more complete regulatory reports, including suspicious activity and suspicious transaction reports.
A similar transition is taking place in customer service. An agent-assist system can identify why a customer is calling and recommend the next best action while helping complete the closure documentation. Across around 10 user journeys, the prediction accuracy for identifying the reason for a call has increased from 75% to 90%. Call closure times have also fallen from around 15 minutes to below five minutes.
For Sundar, these deployments illustrate a broader change: AI is increasingly being measured through operational outcomes rather than the novelty of the underlying technology.
Context becomes the new AI battleground
As banks move towards production, however, the challenge is no longer simply accessing a powerful large language model. It is ensuring that the model understands the organisation’s own context, policies and technology environment.
Sundar identifies three major issues emerging as banks scale AI. The first is token economics. Organisations that initially encouraged developers to make extensive use of AI are now finding that token consumption itself needs to be managed. The focus is shifting towards providing the right context and measuring productivity while ensuring optimal use of tokens.
The second is determinism.
A general-purpose LLM may provide an answer based on information learned from a broad body of industry data. For a bank, however, the answer also needs to align with its own policies, procedures and technology environment.
The objective, therefore, is to ensure that AI systems retrieve relevant organisational context before generating an answer rather than relying primarily on the model’s broader knowledge.
Maveric’s approach uses a context layer and knowledge graph to provide that information to the model. The aim is to make AI outputs more deterministic while also controlling input and output token consumption.
The third challenge is keeping humans in the loop without making human oversight another bottleneck. This requires explainability: users need to understand quickly why an AI system has suggested or taken a particular action.
“While I spend less time on making a decision, I’m 100% sure that there is a wider coverage,” Sundar says, describing the objective of explainability in AI-assisted decision-making.
AI-native development changes the technology lifecycle
The AI transformation is also extending into how banks build and maintain technology.
Maveric has taken its testing expertise into an AI-native platform called Pulse, which is designed to assess the technology landscape and simplify testing. According to Sundar, the platform is helping identify gaps earlier in the development process, including during user acceptance testing and requirements elaboration, allowing testers to operate closer to the pace of developers.
The company is also developing ContextOS, an AI-native software development lifecycle that acts as a knowledge layer for the bank. The system is designed to provide context to different participants across the technology and operations lifecycle.
A product manager can start with a problem statement and, through interactions with agents, develop an epic, user stories, acceptance criteria and business or functional requirements, along with an indication of the budget required. Architect agents can then translate those requirements into high-level and low-level designs that fit the bank’s existing technology stack. Developers can subsequently use their preferred AI coding tools within those guardrails before testing is brought back into the process through Pulse.
This is significant because it shifts AI from being an assistant used by individual employees to becoming part of the workflow connecting requirements, architecture, development and testing.
Modernisation becomes a security strategy
AI is also changing the economics and speed of legacy modernisation.
Banks continue to operate applications built on older technologies, while some are approaching end-of-support milestones for their operating systems and technology stacks. Sundar says AI can now help teams understand undocumented code and dependencies and accelerate application refactoring and migration.
Maveric is working on programmes involving .NET framework upgrades, C++ to Java migration and TIBCO to Spring Boot migration. According to Sundar, programmes that previously took around a year can potentially be compressed to approximately three months with AI-assisted modernisation.
This also has a security dimension.
Rather than continuously adding additional security layers around ageing infrastructure, organisations can use modernisation to remove vulnerabilities at their source. AI can help identify dependencies, map application flows and assess the impact of changes across upstream and downstream systems.
At the same time, AI introduces its own security risks, including data poisoning, adversarial attacks, prompt injection, model inversion and model extraction. The security challenge is therefore becoming two-sided: organisations need to protect legacy systems while also securing the AI infrastructure being introduced into the environment.
The next priority: rationalise before modernising
Perhaps the most important message for technology leaders is that AI should not become another technology layer added to an already complex banking environment.
Banks are increasingly operating across multi-cloud environments and now adding multiple LLMs to the mix, creating another layer of architectural complexity.
Sundar argues that modernisation therefore needs to be accompanied by application rationalisation.
“Application rationalisation and modernisation should be the new mantra,” he asserts.
The objective is to identify overlapping functionality across applications, determine which systems can be consolidated or retired and build a roadmap that reduces the technology estate without disrupting critical business capabilities.
For banks with multiple off-the-shelf applications, this can mean identifying common functionality and deciding which products can be sunsetted while retaining the required capabilities. The result, Sundar points out, is a leaner technology estate with lower operational and infrastructure costs.
That may ultimately become the more important measure of AI-led transformation in BFSI. The winners will not necessarily be the banks that deploy the most AI models or add the most AI tools. They will be the ones that can use AI to make existing technology estates simpler, more responsive and easier to govern, while keeping human judgement, organisational context and accountability firmly in the loop.
