Which Indian rules apply to AI agents in banking?

Two sets, mainly. The Reserve Bank of India supervises banks and sets expectations on IT governance, outsourcing, customer protection and fair lending, which apply to any agent a bank runs. The Digital Personal Data Protection Act 2023 governs how the bank, as a data fiduciary, processes personal data, including through agents. SEBI and IRDAI add rules for securities and insurance.

Who regulates an agent in an Indian bank?

What do the rules require, in substance?

RBI expectations. The RBI expects banks to govern technology risk at board and senior management level, to manage outsourced IT services so that accountability stays with the bank, to treat customers fairly with clear grievance redress, and to follow fair practices in lending, including transparency about terms and reasons. A bank cannot outsource its responsibility to a platform or model provider; the RBI must be able to supervise the activity. The RBI's view of agents is set out in RBI and AI agents in Indian banks.

The DPDP Act. In substance, the bank as data fiduciary must give notice and obtain consent for processing, unless a legitimate use permitted by the Act applies; process only for the specified purpose; keep data accurate where it is used for decisions; protect it with reasonable security safeguards; notify personal data breaches to the regulator and affected people; erase data when the purpose is served; and respond to people's rights to access, correction, erasure and grievance redress. Processing by a data processor, such as an agent platform, must be under a valid contract. Larger fiduciaries designated as significant carry additional duties.

This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.

Where does the line fall, and why?

The reasoning behind both regimes is that the bank stays accountable to the customer and the supervisor. An agent can carry the workload; it cannot carry the accountability.

• Agent may: answer customer queries in several languages from approved content, collect documents for account opening, run checks on KYC documents, monitor transactions and draft alerts, process routine service requests within limits.
• Agent may, within limits: reverse small charges, set up standing instructions the customer requested, approve pre-approved credit offers within policy.
• Stays with a person: credit decisions that refuse or reduce credit, closing or freezing accounts, resolving grievances escalated beyond first response, decisions on suspected fraud, and approval of any new purpose for personal data.

Two features are specific to India. Grievance redress is a strong supervisory focus, so an agent that frustrates complaints or loops customers between channels creates supervisory risk quickly. And the DPDP Act's consent model, with notice in clear language, means each agent purpose that relies on consent must have been covered in the notice the customer actually saw.

An illustrative example

Take a private sector bank that deploys an agent to handle credit card service requests in Hindi, Tamil and English. The agent resolves address changes, card blocks and fee queries. Fee reversals up to a set amount it handles alone; above that, it holds for a service officer.

The bank's data protection review finds the agent was using card spending data to suggest personal loans during service chats. The original notice covered servicing, not cross-selling. The bank removes loan offers from the service agent, routes them through a separate consented marketing process, and adds service chats to its grievance monitoring so complaints raised with the agent are tracked to resolution.

Which controls meet both regimes?

• Board-approved agent governance with named owners.
• Outsourcing assessments of the platform and model provider, with RBI access to records preserved.
• Purpose-bound sources matched to the notices customers received.
• Holds on adverse credit and account decisions.
• Grievance tracking that includes agent-handled complaints.
• Breach procedures that cover agent errors, such as data sent to the wrong customer.

For comparison, see which Kenyan rules apply to AI agents and which Mexican rules apply to AI agents in finance. The full map is in AI agent regulation, mapped to controls.