AI in Banking: Trust, Security and Governance Will Define the Future, Says Tulshekar Gangireddy | Exclusive

As artificial intelligence reshapes global banking, the challenge is no longer just adopting advanced technologies but ensuring they deliver measurable value without compromising security, accountability and customer trust. From agentic AI and real-time fraud prevention to data governance and responsible innovation, financial institutions are entering a new era of transformation.

In an exclusive interaction with Srajan Agarwal of News4Bharat, Tulshekar Gangireddy,  Executive Director & Head of Data Strategy – Trust & Safety & Safety Line of Business, shares his perspectives on scaling enterprise AI, securing autonomous financial systems, navigating India's evolving regulatory landscape and the technological priorities that could redefine banking by 2030. Edited Excerpts:

1. What separates organisations successfully scaling AI from those struggling to translate investments into measurable business value?

In my experience, the biggest differentiator is not the sophistication of the AI model; it is the organisation’s ability to connect intelligence with business outcomes, trusted data, and operational execution.

Over the last 25 years, I have seen several waves of technology transformation, from service-oriented architectures and microservices to cloud platforms, big data and now generative and agentic AI. Each wave has created enormous opportunities, but the organisations that consistently succeed are those that move beyond experimentation and fundamentally rethink how technology delivers value.

Many enterprises still approach AI as a collection of isolated use cases or proof of concepts. They demonstrate what a model can do, but they do not adequately address the quality and accessibility of data, integration with business workflows, operational resilience, governance, or the accountability required when AI influences consequential decisions.

AI creates value when it becomes part of the decision-making fabric of the enterprise, rather than remaining an isolated technology capability.

Five essential pillars of successful AI adoption

• Trusted data foundations: AI is only as reliable as the data, context and knowledge on which it operates.
• Business-led prioritisation: Use cases must be evaluated against measurable outcomes such as fraud losses prevented, improved customer experience, operational efficiency, reduced risk and faster decisions.
• Production-grade architecture: AI must be integrated with enterprise platforms, APIs, real-time data streams, monitoring and established operational processes.
• Governance by design: Security, privacy, explainability, auditability and human accountability must be built into every step of the workflow from the beginning.
• An engineering and operating model that scales: Reusable platforms, common standards and federated business ownership are essential to avoid creating a new generation of disconnected AI solutions.

My work in financial services has reinforced the importance of this approach. I have led data and platform transformation initiatives supporting environments that process more than 500 million transactions daily, while driving improvements in fraud detection and operational efficiency. The lesson is that scale, reliability and measurable business impact must be designed together.

In my article From Data Platforms to Decision Platforms: Architecting AI-Ready Enterprises, this evolution is central to the argument: the next stage of enterprise transformation is not simply about making more data available. It is about enabling the right decision, with the right context, at the right time.

Ultimately, organisations that succeed with AI will be those that treat it as an enterprise transformation agenda, combining data, architecture, people, governance and business accountability. The measure of AI maturity is not how many models an organisation deploys, but how consistently those models produce trusted, measurable and sustainable outcomes.

Also Read | Who Will Finance India's Next Growth Decade? Dr. Amit Goenka on Private Credit & AI

2. As agentic AI gains momentum, how should banks balance automation and customer convenience with security, accountability, and digital trust?

I believe we are approaching a fundamental shift in banking: moving from AI that recommends an action to AI that can independently initiate and execute it. That transition creates enormous opportunities, but it also changes the nature of financial risk.

Recent reports that major US banks are restricting certain third-party AI agents from accessing customer accounts illustrate the challenge. The concern is not simply whether an AI agent can authenticate successfully; it is whether the agent is authorised to perform a particular action, whether that action reflects the customer's actual intent, and who is accountable if something goes wrong.

I have explored this subject in my articles TA14: The Missing Trust Layer for Autonomous AI and TA14 Implementation: Building a Decision Certificate Authority for Autonomous AI, as well as in my writing on AI-led transformation and autonomous finance.

My central argument is that autonomy must be earned through verifiable trust, not granted simply because an agent has access to credentials.

Traditional security models largely focus on identity, authentication and access permissions. Agentic AI requires us to go further. A legitimate agent operating under a legitimate identity can still make an incorrect decision, misinterpret an instruction, or be manipulated by malicious content. Authentication alone does not establish that every subsequent action is appropriate.

I believe banks need a layered approach.

• First, establish a verifiable identity and mandate for every agent. The bank should know which agent is acting, on whose behalf it is acting, what permissions it has, and when those permissions expire.
• Second, enforce policy at the point of execution. An agent proposing a payment or changing a financial instruction should be checked through preset/dynamic rules against transaction limits, beneficiary restrictions, customer intent, risk signals and applicable policies before the action is executed.
• Third, introduce risk-based autonomy. Read-only access and low-risk tasks can be automated more freely. Higher-risk transactions should require stronger authentication, explicit customer confirmation or human approval, depending on the circumstances. Unusual or policy-violating actions should be blocked.
• Fourth, make every consequential action independently verifiable and auditable. My proposed Decision Certificate Authority concept explores how a decision can be bound to a specific action, its parameters, the applicable policy and the authorisation under which it was approved. If the action changes after approval, the authorisation should no longer be valid. Cryptographic integrity can support this process, although it does not by itself establish that the underlying decision was correct.
• Finally, banks must retain the ability to revoke an agent's authority, suspend its activity and recover safely from failures.

I see this as the next evolution of zero-trust security: moving beyond never trust, always verify towards verify the identity, validate the intent, enforce the policy and revalidate before execution.

The objective should not be to prevent autonomous banking. It should be to create an environment where customers can delegate tasks confidently without surrendering control over their money. Trust will become a competitive differentiator for banks that make autonomy safe, transparent and accountable.

3. Why do enterprises still struggle with fragmented systems and actionable intelligence, and what architectural changes are essential for AI-driven organisations?

The fundamental problem is that many organisations have modernised their infrastructure without modernising how information is governed, shared and converted into decisions. This is often associated with a ‘lift-and-shift’ approach, in which existing systems are migrated without fundamentally redesigning their underlying processes.

Over my career, I have worked across large-scale data platforms, streaming architectures, distributed systems, data services and real-time decisioning. I have seen platforms handling more than 10 TB of data daily and supporting millions of transactions. Yet even with this scale of infrastructure, organisations can struggle to answer a basic question consistently: What do we know about this customer, transaction, merchant or risk event right now, and what should we do about it?

The problem is rarely a complete absence of data. It is the fragmentation of context.

Different business units maintain different definitions of the same entity. Data is duplicated across systems, governance is inconsistent, and access to information depends on lengthy integration projects. Batch-oriented pipelines can introduce delays, while operational systems may have information that analytical platforms cannot access quickly enough. AI adds another challenge: models need not only historical records, but also reliable context, relationships, lineage and timely signals.

In my article The Data Leadership Imperative: Governing Centrally in a Federated World for Regulated Enterprises, I emphasise the need to reconcile enterprise-wide governance with distributed ownership and execution.

I believe future-ready architecture must evolve in four ways.

• First, move from data silos to a trusted data foundation. Organisations need common definitions, ownership, metadata, lineage, quality controls and discoverability. Data products should have explicit business owners, service expectations and measurable quality standards.
• Second, combine centralised governance with federated execution. Business domains should retain the agility to build and operate their data products, while enterprise-wide policies enforce consistent security, privacy, access, quality and regulatory controls.
• Third, connect batch intelligence with real-time decisioning. Streaming platforms, event-driven architectures, APIs and analytical services should work together so that operational decisions can incorporate the latest relevant information rather than relying exclusively on periodic reports.
• Fourth, evolve from data platforms to decision platforms. A decision platform brings together trusted data, business rules, analytical models, AI services and execution mechanisms. It does not merely tell us what happened; it helps determine what should happen next, within defined policy boundaries. For example, in fraud prevention, the value is not simply storing transaction history. It is combining transaction patterns, beneficiary relationships, device signals, behavioural context and current risk indicators quickly enough to inform a decision.

The same architectural principle applies to AI agents. Agents need governed access to trusted knowledge and enterprise services, rather than unrestricted access to fragmented databases and application interfaces.

My view is that the next generation of enterprise architecture will be measured less by how much data it stores and more by how reliably it enables decisions. That requires treating data, intelligence, governance and execution as interconnected capabilities.

Also Read | Financial Inclusion in India: J. Balakrishnan on Responsible Microfinance, Technology and Women Empowerment | Exclusive

4. Is AI adoption advancing faster than institutional safeguards, and what should banks prioritise for responsible deployment?

Yes, in some areas, the pace of AI capability development is outstripping the maturity of the controls needed to deploy it safely. But I would distinguish between the pace of experimentation and the readiness for autonomous execution.

Banks have decades of experience managing model risk, operational risk, cybersecurity, privacy and regulatory compliance. The challenge is that generative and agentic AI introduce new combinations of risks: non-deterministic outputs, changing behaviour, complex dependencies, indirect prompt injection, tool misuse and the potential for one automated decision to trigger a chain of consequential actions.

The Federal Reserve's 2026 interagency model-risk guidance explicitly notes that generative and agentic AI are not within the scope of that guidance, while stating that established governance practices should guide the management of tools and processes outside its scope.

This highlights an important distinction: existing risk-management principles remain relevant, but institutions must adapt their implementation to new AI capabilities.

In my view, banks should prioritise five areas.

1. AI inventory and risk classification. Institutions need visibility into where AI is used, what decisions it influences, what data it accesses and the potential consequences of failure.
2. Security and governance throughout the lifecycle. Risk assessment must cover data preparation, model development, deployment, agent tools, third-party dependencies and ongoing operation.
3. Meaningful human oversight. Human involvement should be proportionate to the risk and designed so that reviewers have the information, authority and time needed to intervene. A nominal approval step is not a meaningful safeguard.
4. Continuous monitoring and adversarial testing. Models and agents must be tested for drift, unexpected behaviour, manipulation, data leakage, privilege escalation and failures under abnormal conditions.
5. Operational resilience and accountability. Banks need clear ownership, auditable decisions, incident-response procedures, kill switches, rollback mechanisms and tested recovery plans.

I would add a sixth priority: govern the action, not just the model.

A model may pass validation and still be unsafe when connected to an inadequately controlled payment service, customer database or external tool. The system must validate the specific action an AI component proposes, against the authority and policies applicable at that moment.

This is where my work on TA14 and admissible execution becomes relevant. The concept is to establish an enforceable boundary between an AI system's ability to reason and its authority to act. Intelligence can propose; policy determines whether the proposed action is admissible; and the execution layer enforces that decision.

Responsible AI should not be treated as a compliance checkpoint that slows innovation. It should be an engineering discipline that makes innovation scalable.

For banks, the goal is not to eliminate all risk. It is to understand risk, constrain it within acceptable boundaries, detect deviations quickly and ensure that the institution can intervene when necessary.

5. How can Indian banks translate the RBI's FREE-AI framework into practical AI governance while continuing to innovate?

I see the RBI's FREE-AI framework as an important opportunity for Indian financial institutions to make responsible AI a source of competitive advantage rather than treating it purely as a regulatory obligation.
The framework, published in August 2025, sets out seven foundational principles and 26 recommendations across six strategic pillars: infrastructure, policy, capacity, governance, protection and assurance. Its central ambition is to enable innovation while managing the risks associated with AI adoption in finance.

For banks, the real challenge is translating these principles into everyday engineering and business practices.

I would approach implementation through five practical measures.

1. Establish risk-based AI governance: Every AI use case should have a documented business purpose, accountable owner, risk classification, approved data sources and defined performance criteria. A customer-service assistant, a fraud detection model and an autonomous payment agent should not be subject to identical controls because their potential consequences differ significantly.

2. Embed governance into the technology architecture: Governance should be enforceable through the platform, not dependent entirely on policies and manual reviews. Access controls, data lineage, model monitoring, audit logs, explainability mechanisms and policy enforcement should be integrated into the development and deployment lifecycle.

For agentic systems, banks should additionally validate each consequential action against its authorised purpose and applicable risk limits.

3. Enable controlled experimentation: Banks should establish secure AI sandboxes and reusable platforms where teams can experiment with new models, synthetic or appropriately protected data, and emerging architectures. A clear pathway from experimentation to production should define the evidence required at each stage.

This enables innovation to proceed without bypassing the controls required for regulated environments.

4. Make assurance continuous: Model validation before deployment is not sufficient. Banks should monitor model drift, bias, performance degradation, data quality, cyber threats and unexpected outcomes throughout the operational lifecycle. They should also test whether human intervention, escalation and recovery mechanisms work under realistic failure conditions.

5. Invest in people and shared capabilities: Responsible AI requires collaboration among engineering, data science, cybersecurity, compliance, legal, risk management and business teams. AI literacy must extend beyond technical specialists to the executives and business owners accountable for AI-enabled decisions.

My experience establishing Data and AI Centres of Excellence has reinforced the value of combining common enterprise standards with distributed execution. Central teams can provide reusable platforms, governance patterns and assurance mechanisms, while business domains retain ownership of their use cases and outcomes.

I would also encourage banks to measure governance effectiveness through operational indicators: time to approve a use case, percentage of models with complete lineage, policy violations detected, incidents contained, model performance over time and measurable business value delivered.

The objective is to make the safe path the easiest path for engineering teams.

India has an opportunity to demonstrate that responsible AI and rapid innovation are complementary, not competing, priorities. With strong digital infrastructure, sophisticated payment ecosystems and a growing engineering talent base, Indian banks can develop governance practices that support both domestic financial inclusion and globally competitive financial technology.

6. How should financial institutions rethink data architecture and governance in response to India's DPDP Rules, 2025, while preserving AI-driven innovation?

I believe data protection should be treated as an architectural principle, not a compliance layer added after a platform has been built.

India notified the Digital Personal Data Protection Rules, 2025, in November 2025, establishing a phased implementation framework. Certain provisions took effect earlier, while many substantive requirements have an 18-month transition period. Financial institutions must therefore map their obligations to the relevant commencement dates.

For financial institutions, this has direct implications for how data is collected, stored, shared, analysed and used by AI systems.

In my thought leadership on The Data Leadership Imperative and Securing Data in the AI Era: Why Visibility, Ownership, and Observability Matter, I emphasise that organisations cannot govern what they cannot discover, classify and monitor.

I would focus on five architectural changes.

• Build a discoverable data estate. Maintain inventories of personal data, data ownership, lineage, sensitivity classifications, processing purposes and downstream consumers. This must extend to data replicated into analytical platforms, feature stores, AI applications and third-party services.
• Make purpose and access controls enforceable. Access should be governed by the intended use, the user's or service's authority, and applicable policies. Data minimisation should be implemented at ingestion, feature creation, retrieval and model-serving stages.
• Strengthen data lifecycle management. Retention, deletion, correction and other applicable data-handling obligations need to be reflected in operational processes and downstream dependencies, including derived datasets and AI pipelines where relevant.
• Enable privacy-conscious AI development. Use data minimisation, masking, tokenisation, synthetic data where appropriate, and controlled environments for experimentation. Privacy-enhancing techniques should be selected according to the sensitivity of the data and the use case.
• Make governance observable. Maintain evidence of access, processing, policy decisions, data lineage and relevant changes. Integrate security monitoring and incident response so that the organisation can investigate and respond to failures.

There is an important distinction between centralised governance and centralised data storage. A bank does not necessarily need to consolidate every dataset into one physical repository to achieve effective governance. A federated architecture can allow domains to retain operational ownership while common policies govern how data is discovered, accessed and used.

For AI, this architecture becomes even more important. A model may not directly access a customer database, but its features, embeddings, retrieved documents or derived outputs may still expose sensitive information. Governance must therefore cover the complete information lifecycle rather than only the original data source.

I would also avoid treating privacy and innovation as opposing objectives. Clear data ownership and purpose controls can actually accelerate innovation because teams know which datasets they may use, under what conditions, and what approvals are required.

The strategic shift is from asking, Are we compliant with data protection requirements? to asking, Can we demonstrate that every use of personal data is appropriately governed throughout its lifecycle?

That is the foundation for building AI-driven financial institutions that can innovate confidently while preserving customer trust.

Also Read | “Build Institutions, Not Just Unicorns”: Dr Nirmal Bansal on India’s AI Future | Exclusive Interview

7. How close are we to preventing financial threats before they occur, particularly AI-assisted phishing and deepfake-enabled fraud?

I believe we are moving from reactive fraud detection towards predictive and increasingly preventive financial security. However, we should be realistic: no system can guarantee the prevention of every fraud attempt, particularly as attackers use AI to adapt their methods.

My experience working with AI/ML platforms processing more than 500 million transactions daily has reinforced a critical principle: fraud prevention is not just a classification problem; it is a real-time intelligence and decision problem.

Traditional fraud systems often evaluate individual transactions against known patterns, rules or model-generated risk scores. These approaches remain valuable, but sophisticated attacks increasingly exploit relationships across identities, devices, accounts, beneficiaries, communication channels and transaction sequences.

AI-assisted phishing and deepfake-enabled impersonation make this problem harder because attackers can produce convincing communications, imitate voices and exploit legitimate customer behaviour. Recent industry research has highlighted the growing prevalence of such scams in India.

My article Beyond MFA: Building Context-Aware Fraud Defense in a Real-Time, AI-Driven Global Economy explores the need to move beyond relying on a single authentication event towards evaluating the broader context of a transaction.

I see four capabilities as particularly important.

• First, contextual and continuous identity verification. Authentication should not be treated as a one-time event. Systems should continuously assess relevant signals, including device integrity, behavioural patterns, session characteristics, account history and transaction context. A legitimate login does not automatically make every subsequent action legitimate.
• Second, entity resolution and relationship intelligence. Banks should be able to identify meaningful relationships among accounts, devices, beneficiaries, merchants and other entities, even when attackers attempt to obscure those relationships. Graph analytics and temporal graph neural networks can help identify suspicious connections and evolving patterns that individual transaction models might miss.
• Third, real-time risk evaluation and adaptive intervention. Systems should combine transaction features, behavioural signals, network relationships and emerging threat intelligence to make decisions while a transaction is being initiated. Depending on the risk, the response might be to approve, step up authentication, delay execution, request independent confirmation or block the transaction.
• Fourth, coordinated AI-assisted defence. Multiple specialised capabilities can work together to analyse identity risk, transaction behaviour, communication signals and network relationships. However, coordination must be supported by clear permissions, independent policy enforcement and monitoring.

Adding more AI agents does not automatically make a system safer.

Consider a scenario in which a customer receives a convincing deepfake voice call instructing them to transfer funds to a new beneficiary. A conventional system may see a successfully authenticated customer initiating a seemingly valid payment. A context-aware system could additionally evaluate the new beneficiary, device and session signals, unusual transaction patterns, recent account changes and other relevant risk indicators. The combined evidence could trigger independent confirmation or block the transaction before funds are released.

This is the distinction between detecting fraud after a loss and interrupting the chain of events that makes the loss possible.

My broader view is that prevention requires integrating intelligence with enforceable security controls. Even a highly accurate model cannot prevent a transaction if the execution architecture ignores its risk assessment or permits the decision to be overridden without appropriate authorisation.

This connects directly to my work on Knowledge Is the New Infrastructure: Rethinking Enterprise Security in the Age of AI. Protecting financial institutions increasingly requires understanding not only where data resides, but also how identities, knowledge, decisions and actions interact.

I expect substantial progress towards anticipatory fraud defence by 2030. But the strongest systems will combine AI with high-quality data, graph intelligence, secure identity infrastructure, real-time decisioning and well-designed recovery mechanisms. The goal is not simply to predict more accurately; it is to intervene at the earliest appropriate point (shift left philosophy) while minimising friction for legitimate customers.

8. What major technological shifts do you foresee over the next five years, and what should financial institutions invest in today to remain competitive and trusted by 2030?

Over the next five years, I expect banking technology to evolve along three interconnected dimensions: from data to intelligence, from intelligence to autonomous action, and from perimeter-based security to continuous trust enforcement.

My 25-year journey across enterprise architecture, distributed systems, cloud, data platforms and AI has taught me that major technology shifts rarely replace everything that came before them. Instead, they change how existing capabilities are connected and how value is created.

In my articles From SOA to Microservices to AI Agents — My 25-Year Journey and EJBs and AI Agents — An Architectural Déjà Vu from a 25+ Year Journey, I explore the architectural evolution towards increasingly autonomous software. The important lesson is that new abstractions do not eliminate the need for sound engineering fundamentals. Reliability, explicit contracts, isolation, observability and controlled execution become even more important as systems gain autonomy.

I see five major shifts shaping banking through 2030.

1. From analytical platforms to real-time decision intelligence: Banks will increasingly combine streaming data, graph intelligence, AI models and business policies to make decisions in milliseconds or seconds, depending on the use case.

This will affect fraud prevention, credit assessment, payments, customer service, liquidity management and operational resilience. The competitive advantage will come from the quality and timeliness of the context available at the point of decision.

2. From AI assistants to governed autonomous agents: AI agents will increasingly perform multi-step tasks across banking workflows, from information gathering and reconciliation to operational investigations and selected customer-authorised financial activities.

However, autonomy will need to be graduated according to risk. Read-only tasks and reversible operations can generally support greater automation than irreversible financial transfers or high-impact credit decisions.

The critical infrastructure will include agent identity, scoped permissions, policy enforcement, transaction-level authorisation, auditability and the ability to revoke access immediately.

This is where my TA14/admissible execution work becomes especially relevant. By separating an agent's ability to propose an action from its authority to execute that action, banks can expand autonomy without granting unrestricted control.

3. From fragmented data estates to governed enterprise knowledge: Banks will increasingly invest in trusted knowledge layers that connect structured data, documents, business rules, entity relationships and operational context.

This will support retrieval-augmented generation, specialised financial models, AI-assisted investigations and more context-aware decision systems. But knowledge infrastructure must also address provenance, permissions, freshness, conflicting information and the risk of sensitive data being exposed through AI outputs.

I expect the evolution towards decision platforms to be just as important as the adoption of foundation models themselves.

4. From periodic security controls to continuous, adaptive defence: Identity, cybersecurity, fraud prevention and operational risk will become more tightly integrated. AI-assisted attacks will require institutions to evaluate not only whether a user or device is trusted, but also whether the requested action remains consistent with the authorised purpose and current risk context.

This will strengthen the role of zero-trust architecture, continuous authentication, graph-based threat detection, behavioural analytics, automated response and policy-enforced execution.

5. From isolated AI deployments to industrialised, governed AI platforms: The next competitive divide will be between organisations that deploy AI as individual projects and those that operate it as a reliable enterprise capability.

Reusable model-serving infrastructure, evaluation frameworks, MLOps, agent observability, security testing, cost management and standardised governance will become essential. Banks will also need to balance the use of large general-purpose models with smaller, specialised models that can deliver better economics, predictable latency or greater control for specific tasks.

What should banks invest in today?

I would prioritise six foundational capabilities.