The RBI expects banks and other regulated entities to run AI agents within the rules they already follow: IT governance and cyber security, outsourcing limits that keep key decisions in-house, digital lending conduct, customer grievance redress and fraud risk management, with the DPDP Act 2023 governing personal data. The regulated entity remains accountable for every action an agent takes.
What does the RBI require, in plain words?
The Reserve Bank of India regulates banks, non-banking financial companies, payment system operators and other entities. It has shown supervisory interest in responsible use of AI, but agents are governed today by existing directions and guidance. In substance, five sets matter:
- IT governance, risk and cyber security. Board-approved IT strategy, risk management, access control, change management, audit trails, business continuity and cyber incident reporting to the RBI.
- Outsourcing. A regulated entity remains responsible for outsourced activities and must manage service providers through due diligence, contracts, monitoring and exit plans. Core management and decision-making functions, such as sanctioning loans and deciding KYC compliance, stay with the regulated entity.
- Digital lending. Lending through digital channels must follow conduct rules on disclosure, data collection with consent, oversight of lending service providers and grievance handling.
- Customer service and grievances. Fair treatment, timely complaint resolution, internal escalation and access to the RBI's ombudsman scheme.
- Fraud risk management. Systems to prevent, detect and report fraud.
Alongside, the Digital Personal Data Protection Act 2023 governs personal data processing. For a broader map of Indian rules, see which Indian rules apply to AI agents in banking.
Who does it apply to?
Entities regulated by the RBI: scheduled commercial banks, small finance and payments banks, cooperative banks, NBFCs and payment system operators, with expectations scaled to size and type. Insurers and securities firms answer to IRDAI and SEBI respectively.
What does this mean when the system is an agent that acts?
Decisions stay with the bank. An agent that recommends loan sanction can support the decision. An agent that sanctions loans on its own, supplied and run by a vendor, raises the question of whether a core decision function has been handed outside. The safer design keeps the agent drafting recommendations and a named officer of the regulated entity deciding, or, where the bank itself runs an agent that acts, keeps authority, limits and evidence firmly with the bank.
The vendor is a service provider. An agent platform hosting customer data and connected to core systems is an outsourcing arrangement. Due diligence, contract terms, audit rights, data location and exit plans apply. The model provider behind the platform is part of the chain to understand.
Digital lending conduct reaches the agent's words. An agent that explains loan terms must give accurate disclosures. An agent that collects data must collect only what is needed, with consent. An agent that contacts borrowers in arrears must follow conduct rules on recovery.
Grievances must reach a person. Customers complaining about an agent's action need a route to a person with authority, and the bank's complaint data should show where agents were involved.
Audit trails must cover the agent. IT governance expects records of who did what in each system. For agents, that means every action, the data used and any approver.
Take an illustrative case. An Indian NBFC deploys an agent in its personal loan app to answer questions, gather documents and prepare credit files. The agent cannot sanction loans; a credit officer does. It collects only the documents the product needs, with consent captured in the app. Every disclosure it gives is drawn from the approved key fact statement. Complaints about the agent route to the grievance team, and the NBFC's internal audit reviews a sample of agent conversations each quarter.
How do you meet RBI expectations for agents?
| Requirement | Control | Evidence it leaves |
|---|---|---|
| Board-approved IT governance | Agent policy approved by the board or its committee | Policy and minutes |
| Access and change control | Least-privilege agent credentials; changes to model, brief and tools controlled | Access reviews; change log |
| Audit trails | Every agent action logged with data used and approver | Append-only action log |
| Outsourcing limits | Core decisions such as loan sanction kept with named officers | Decision records with officer names |
| Service provider management | Due diligence, contracts, audit rights and exit plans for agent vendors | Outsourcing file |
| Digital lending conduct | Disclosures drawn from approved documents; need-based data with consent | Disclosure sources; consent records |
| Grievance redress | Route from agent to a person; complaints tagged for agent involvement | Complaint records |
| Fraud risk | Agent actions screened; untrusted documents cannot authorise payments | Fraud alerts; refused actions |
| Data protection (DPDP Act) | Notice, consent and purpose limits for agent processing | Notices; consent logs |
| Cyber incidents | Agent misuse covered by incident response and reporting | Incident records |
What will an RBI inspection ask?
- Which agents operate, and what can each do without an officer's approval?
- Does any agent take decisions that should stay with the regulated entity?
- Who supplies the agent and its model, and what do your contracts secure?
- How are digital lending disclosures by agents controlled?
- How do customers complain about an agent, and what do those complaints show?
- Can you produce the record of an agent's actions on a given account?
For how banks build the validation evidence behind these answers, see how banks validate AI agents.
What do RBI expectations not cover?
They do not create an AI-specific approval process, and they do not replace the DPDP Act on personal data or SEBI and IRDAI rules for securities and insurance business. Comparing regimes helps when you operate in several markets: Nigeria's central bank applies a similar existing-rules approach, described in CBN and AI agents in Nigerian banks. For the full map, see our guide to AI agent regulation mapped to controls.
This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.
Questions people ask
Can an Indian bank let an AI agent sanction loans?
Credit sanction is a core function that stays with the regulated entity. A bank-run agent supporting or performing parts of that work needs bank-owned authority, limits and evidence, and many banks keep a named officer as the decision-maker. Confirm your design with counsel and your RBI-facing compliance team.
Is an agent platform an outsourcing arrangement under RBI rules?
Where a third party performs activities the bank would otherwise do, or hosts and processes its data, it is likely to be. Treat it under your outsourcing policy, including due diligence, contracts and exit plans.
Does the DPDP Act require consent for agent processing?
The Act relies mainly on consent, with certain legitimate uses permitted without it. Agents must process personal data for the purposes notified and consented to, and stop when consent is withdrawn.