AI in Indian Banking: Through the CISO’s Lens
AI adoption changes who can access information, influence a decision and act on an account. A CISO’s view of the controls Indian banks need to make those boundaries work.
Perspectives · AI in Indian Banking · Trust by Design
A bank proposes an AI assistant that can read customer correspondence, retrieve account information and prepare service requests. The demonstration is convincing. It answers quickly, understands conversational language and reduces the employee’s effort.
Through the CISO’s lens, the next questions concern the banking service: whose information can it retrieve, which instructions can it act on, and what happens when it misunderstands a request?
A summary sent to the wrong person, an unauthorised beneficiary change or an inaccurate explanation of a loan can affect a customer even when every server is available. Security therefore needs to follow the information, decisions and actions across the full journey.
The CISO helps the bank define where AI can operate, what information it may use and which actions require independent approval. Business owners, model validators, privacy and compliance teams bring their own responsibilities to that decision. The aim is a service the bank can explain, control and recover when something goes wrong.
1. Match the controls to the AI use case
“AI risk” can conceal quite different problems. An employee using an assistant to draft a generic note has a different exposure from a model influencing credit decisions or an agent executing a financial instruction.
| Exposure | Banking example | CISO’s starting question |
|---|---|---|
| AI used inside the bank | Employee assistants, code generation and customer-service tools. | What information enters the service, and who can retrieve its outputs? |
| AI influencing decisions or actions | Credit recommendations, fraud prioritisation and agent-assisted servicing. | What authority does it have, and which independent controls constrain the outcome? |
| AI used against the bank | Impersonation, manipulated documents and malicious content reaching connected assistants. | Can existing verification and response processes withstand a convincing deception? |
The same project can involve all three. Approval should reflect the actual capability and customer impact, then be revisited when either changes.
2. Know where AI is already operating
An AI inventory should include capability embedded in purchased products. A CRM assistant, fraud platform, contact-centre service or developer tool may acquire new functions through a vendor release. The bank’s original procurement approval may describe a narrower service.
A practical register records the purpose, business owner, users, data sources, model or service provider, deployment location, connected tools, permissions and customer impact. Link each entry to its approval, change history and fallback arrangement.
Offer employees an approved route for useful experimentation. A policy that provides no workable alternative can leave security learning about usage after sensitive information has already been shared.
3. Protect the information before it reaches the assistant
An internal knowledge assistant can expose information through retrieval even when it does not train a model on bank data. A document index, cached answer, conversation history or support log can become another route to confidential material.
Consider a branch employee asking about a customer’s complaint. The retrieval service must enforce that employee’s authorised scope. A broadly privileged connector should not retrieve every complaint and rely on the model to decide which records to conceal.
Map collection, storage, retrieval, processing, logging and deletion across the bank and its providers. Examine retention, subcontractors, administrative access, backup handling and contract terms against applicable requirements. Where location restrictions apply, assess the complete processing chain.
Private-cloud deployment gives the bank choices about infrastructure control. It still needs access governance, protected secrets, secure connectors and disciplined operations. Deployment location alone is insufficient evidence of security.
4. Keep transaction authority outside the model
OWASP identifies prompt injection as a risk when inputs influence model behaviour in unintended ways, including through external files or websites. Retrieval augmentation does not eliminate that risk. A document being relevant to a customer request does not make instructions inside it trustworthy. [5]
For banking architecture, the practical consequence is to enforce transaction authority outside the model. An assistant may interpret a request, but the servicing API must still authenticate the actor, check permissions and apply the bank’s approval rules.
OWASP also describes excessive agency in terms of unnecessary functionality, permissions and autonomy. Those distinctions are useful during design review: access needed to explain an account should not quietly include authority to modify it. [6]
For a fintech channel, LMS in private cloud and CBS in a data centre, preserve the same control chain across every environment. Record the originating user, delegated permissions, application or loan reference, approval and financial outcome. A common correlation identifier helps trace the journey; it does not establish authorisation by itself.
If an execution request times out, verify its outcome before retrying. The assistant should hand an unresolved transaction to a controlled recovery process rather than invent an explanation or create another instruction.
5. Make oversight work for the customer
AI can help explain services in a customer’s preferred language. Test whether the explanation preserves meaning across languages, transliterated text, local expressions and code-switching. Fluent wording can still misstate a fee, repayment obligation or complaint outcome.
A human reviewer needs time, authority and access to the evidence. A queue that encourages staff to accept every suggestion is a weak basis for oversight. Specify which actions require review and which conditions require escalation.
Security should work with fraud, operations and customer-service teams on impersonation scenarios. A familiar voice, convincing video or plausible document should not bypass independent verification for a sensitive request. Customer assistance must remain accessible when verification is difficult.
Credit suitability, fairness and model performance also need accountable review. The CISO contributes security evidence; the relevant business, risk and validation functions must assess whether a model is suitable for its intended banking use.
6. Set clear boundaries for AI in the SOC
An assistant can help analysts summarise alerts, search approved runbooks and organise an investigation. Introduce these uses with a defined evaluation: does the tool improve decisions on representative cases, including ambiguous ones?
Containment needs a separate authority model. Disabling an identity or blocking a connection can interrupt a critical service. Define which actions are pre-authorised, their limits, the evidence required and who can approve exceptions.
Test the response when the assistant produces a misleading recommendation or becomes unavailable. Analysts need a usable fallback, and the bank needs enough independent evidence to reconstruct the incident.
7. Prove the controls before release and after change
A proposed approval package should explain the use case and exclusions, information flows, access controls, evaluation results, identified weaknesses, monitoring, supplier dependencies and recovery arrangements. Significant exceptions need named owners and review dates.
Test the full service with representative data and controlled adversarial scenarios. Include inappropriate retrieval, malicious document content, unauthorised tool requests, ambiguous instructions and provider interruption. Check the financial outcome and customer experience alongside model responses.
A change to the model, retrieval corpus, prompt, connector or permission can alter the exposure. Set reassessment triggers according to materiality. A completed assessment describes a tested configuration; it should not become permanent approval for every later capability.
8. Show leadership where the controls hold
| Leadership question | Useful evidence |
|---|---|
| Where does AI have material authority? | Current use-case inventory with data access, permitted actions and accountable owners. |
| Do controls work in practice? | Scenario results, unauthorised retrieval attempts, blocked actions and unresolved weaknesses. |
| Can people intervene? | Timed stop, escalation and handover exercises with appropriate staffing. |
| Can a supplier change expose the bank? | Change notifications, dependency reviews and demonstrated fallback arrangements. |
| What happens to customers when AI fails? | Incorrect outcomes, correction time, complaints and successful assisted recovery. |
These are proposed management measures, not a universal regulatory return. Their value comes from showing where assurance is strong, where it remains incomplete and what decision leadership needs to make.
The CISO’s contribution is to make adoption conditions explicit and testable. AI can help a bank serve customers better when the institution understands its information access, limits its authority and can explain and recover its outcomes. The bank retains responsibility for delivering that trustworthy service.
The Indian regulatory context
The RBI published the FREE-AI Committee report on 13 August 2025. It sets out seven foundational principles and 26 recommendations across six pillars. It is an important reference for responsible adoption; its recommendations should not all be presented as already operative regulatory obligations. [1]
On 24 June 2026, RBI invited comments on draft model-risk guidance covering a wider range of models, including AI/ML and third-party models. The cited publication is a consultation draft. Banks should verify any subsequent final instrument and its applicability before using it as a compliance checklist. [2]
Existing security governance remains relevant. The Commercial Banks cybersecurity and technology directions cited here specify an independent CISO reporting arrangement and board-level risk reporting. Their definition excludes Small Finance Banks, Payments Banks and Local Area Banks; those institutions need to identify their applicable instruments. One category’s directions cannot be generalised to every Indian bank. [3]
CERT-In’s AI Security Auditing Guidelines, dated 4 November 2026, address lifecycle assurance across governance, security, privacy, resilience and auditability. They recognise risks beyond conventional application testing, including manipulation of models or data, leakage, drift and inadequate oversight. Their relevance is broader than a one-time technical scan. [4]
Continue with Learning
The assistant can read. Should it act? Follow an uploaded document into a beneficiary-change request, test the access boundary and decide what a vendor update should trigger.
Sources and reading notes
Sources reviewed on 21 November 2026. The publications have different scopes and legal effects, as described above. Operating examples and recommendations are the author’s synthesis.
- RBI — Publication of the FREE-AI Committee report, 13 August 2025.
- RBI — Consultation on draft Guidance on Regulatory Principles for Model Risk Management, 24 June 2026.
- RBI — Commercial Banks: Cybersecurity, Technology: Risk, Resilience and Assurance Framework Directions, 2026; applicability and CISO provisions.
- CERT-In — AI Security Auditing Guidelines, 4 November 2026.
- OWASP GenAI Security Project — LLM01:2025 Prompt Injection.
- OWASP GenAI Security Project — LLM06:2025 Excessive Agency.
Put this perspective into practice.
Explore the concepts, make a decision and test what changes when the situation changes.
Topics: Banking Architecture, Identity and Access, Operational Resilience, Perspective, Trust by Design