Agentic AI in Indian banking runs inside a dense set of published rules: RBI’s FREE-AI framework for responsible AI, the 2025 Authentication Directions for digital payments, the e-mandate framework for recurring debits, the Account Aggregator framework for consented data, and the Digital Personal Data Protection Act. A bank deploying agents should map each rule to a concrete control: consent before data, a person on every commitment, two-factor confirmation of payments, and a record that can be replayed.
This post summarises the published texts in plain words, and shows how Tuvel’s design responds. It is general information, for your risk and compliance teams to review against the texts themselves.
RBI’s FREE-AI framework
In August 2025, an RBI committee published its Framework for Responsible and Ethical Enablement of AI (FREE-AI). It sets out seven guiding principles, the Sutras, and recommendations for AI in financial services, including board-approved AI policies, governance structures and an AI sandbox.
For agents, the practical reading is: know which AI systems you run, grade how much each may do, keep a person accountable, and be able to show what happened. Tuvel records every turn, tool call and gate decision. Autonomy is graded per agent, and for the customer advisor the customer and the bank each hold a key. On the compliance desk, the AI Governance and Inventory agent keeps the bank’s register of AI systems for its own governance.
The Authentication Directions, 2025
RBI’s Authentication Directions (RBI/2025-26/79), in force from 1 April 2026, require digital payments to be authenticated by at least two distinct factors unless a listed exemption applies.
That rule shapes how a customer-facing agent may move money. In Tuvel Advisor, PINs and one-time codes stay out of AI models. A spoken or typed yes never confirms money; a tap bound to the exact amount and payee does. At levels 1 to 3, every money action is confirmed with that tap. Acting without a tap at levels 4 and 5 is On the roadmap, built around the framework for agent-initiated payments on our responsible AI page.
The e-mandate framework
RBI’s e-mandate framework (RBI/DPSS/2026-27/396) requires an additional factor to register a recurring mandate, a pre-debit notice before each debit, and a way for the customer to opt out or withdraw. An advisor that sets up autopay has to work inside those steps. In Tuvel, autopay at level 5 is On the roadmap, within the customer’s limit and the bank’s ceiling, set up under the mandate rules.
Account Aggregator and the consent layer
Under RBI’s Master Direction for NBFC Account Aggregators, financial information is shared only with the customer’s explicit consent, through a consent artefact. That is the right model for any agent that reads a customer’s accounts.
In Tuvel, accounts are linked through a consent journey modelled on Account Aggregator: what is shared, for what purpose, approve or revoke. Tuvel reads linked data only after consent. The AA connection itself is set up for each bank Per deployment.
Consent is also the principle for staff agents. Each agent sees only the customer details and tools it was granted, and a connector can be revoked in one click.
The DPDP Act and Rules
The Digital Personal Data Protection Act, 2023, and the DPDP Rules, 2025, set out notice and consent, limits on purpose, security safeguards, and rights to access, correct and erase personal data.
Tuvel’s design is built to support them. Personal identifiers are masked before text reaches a model. Uploaded documents are read on the server and the original is not stored. Tuvel deploys as a private copy per bank, with its own database, inside the bank’s VPC in an Indian region of Google Cloud, AWS or Azure, and an open-weight model hosted in India keeps every AI model call inside India.
Prepaid instruments and agent payments
RBI’s Master Directions on Prepaid Payment Instruments cap a full-KYC instrument’s balance at ₹2,00,000 and general funds transfers at ₹10,000 a month. Tuvel Agent Card, a prepaid card only the advisor pays from, sits inside those limits by design, and is issued on the bank’s own full-KYC prepaid programme. See Tuvel Agent Card.
Payments by AI agents are moving in India. NPCI, Razorpay and OpenAI piloted UPI payments inside ChatGPT in October 2025, with users pre-authorising the agent within limits they set (Razorpay). The direction is clear: preset limits, chosen by the customer, enforced at the moment of payment.
What this means for staff agents
Most early agentic work in Indian banks is on staff desks, where the rules above still apply. Agents work with GST returns, CKYC, EDPMS and IDPMS, RBI circulars and Indian credit practice. On Tuvel’s compliance desk, the Circular Tracker and Change Impact Mapper maps new circulars to policies and products, and the Regulatory Return Preparer prepares returns. Accepting an interpretation of a circular and submitting a return stay with people.
Languages
For customers outside the big metros, an advisor has to speak their language. Tuvel speaks English and ten Indian languages through Sarvam, by voice or text.
A short map for your compliance team
| Rule | The control to look for |
|---|---|
| FREE-AI | An AI inventory, graded autonomy, a person accountable, a replayable record |
| Authentication Directions | Credentials out of the model; payments confirmed by a bound tap |
| E-mandates | Additional factor at registration; pre-debit notice; easy withdrawal |
| Account Aggregator | Consent with a purpose, before any data is read; revocable |
| DPDP Act and Rules | Masking, purpose limits, documents not kept, and a clear answer on where data is processed |
The full line-by-line map is on our responsible AI page, and the India view is on Tuvel for India.