India processes over 15 billion UPI transactions per month. NPCI's Unified Agent Protocol introduces a new participant category into that infrastructure: verified AI agents with delegated transaction authority. The governance implications for PSPs, member banks, and AI companies building UAP-compatible agents are substantial — and the formal regulatory documentation is still being developed.
UAP Core Parameters
₹15,000 / Month
Autonomous transaction limit per AI agent per user per month. Transactions within this cap execute without per-transaction human confirmation. Above the cap: user confirmation required. This is the externally defined autonomy envelope; banks must define additional internal scope constraints within it.
Agent + PSP / Member Bank
AI agents register through a member bank or PSP that holds UAP registration with NPCI. The PSP/bank is the accountable party. Agent gets a verified identity in NPCI's UAP registry — analogous to merchant registration in UPI but for AI agents. Agent identity is tied to the registering PSP.
Explicit Pre-Authorization
Users must explicitly consent to UAP agent transacting on their behalf. Consent parameters: transaction categories allowed, per-transaction limits within monthly cap, revocation mechanism. DPDP Act consent architecture applies to user data the agent processes. Consent is separate from the UAP registration.
DPIP Integration
UAP agent transactions feed into RBI's Digital Payment Intelligence Platform (DPIP, August 2026). Anomalous agent transaction patterns (velocity, unusual payees, timing) generate DPIP alerts. PSPs responsible for monitoring DPIP alerts for their registered agents and responding to RBI fraud intelligence.
Governance Obligations Across Four Regulatory Frameworks
| Governance Activity | NPCI UAP | RBI PA Directions (Sep 2025) | RBI MRMF (Jun 2026) | DPDP Act 2023 |
|---|---|---|---|---|
| Agent registration and identity | NPCI Agent registered via PSP in UAP registry; PSP accountable | PA Directions PA merchant onboarding due diligence extends to AI agents registered as PA merchants | — | — |
| Authorization scope controls | NPCI ₹15,000/month external cap; agent must not exceed | PA Directions Transaction monitoring for fraud and unusual patterns | MRMF Authorization boundary documentation for agentic AI; internal scope constraints beyond NPCI cap | — |
| Transaction audit trail | NPCI NPCI dispute resolution requires transaction log per agent action | PA Directions PA dispute resolution: complete transaction record for each disputed payment | MRMF AI agent action audit trail for model governance; each agent decision attributable to model version | — |
| Error transaction liability | NPCI PSP accountable in first instance; shared responsibility parameters being formalized | PA Directions PA dispute resolution SLA; chargeback obligations | — | — |
| Fraud monitoring | NPCI Agent transaction patterns in NPCI fraud systems; DPIP feeds | PA Directions PA fraud monitoring; FIU-IND STR for suspicious patterns | MRMF AI model performance monitoring; agent behavior drift detection | — |
| User consent management | NPCI User pre-authorization with defined parameters | — | — | DPDP Separate consent for user data processed by AI agent; purpose limitation; right to revoke |
| Kill switch / agent suspension | NPCI Agent deregistration in UAP registry (parameters being formalized) | — | MRMF Tested kill switch capability for governed AI models — applies to UAP agent as AI model | — |
| Third-party agent governance | NPCI PSP accountable for agent registered through it regardless of AI company building the agent | PA Directions PA due diligence on technology providers | MRMF Third-party model governance: documentation, validation, audit rights from AI company | — |
What to Build Before UAP Goes Beyond Pilot Deployment
UAP was announced at Global Fintech Fest September 2026. Formal NPCI circulars setting out the full operational and governance framework are forthcoming. PSPs and banks that build governance infrastructure now will not be in retrofit mode when the formal requirements are issued.
- Design authorization scope controls beyond the ₹15,000 cap. The NPCI cap is the external ceiling; it is not an authorization policy. Before registering any UAP agent, define internal authorization scope: which transaction categories can the agent execute (bill payments only? peer payments? merchant payments? new payee setup?)? Can the agent transact with payees the user has never paid before, or only established payees? Can it initiate recurring payment setups? Each internal scope constraint reduces the error liability surface and the potential fraud surface beyond the regulatory requirement. Authorization scope must be documented for MRMF's agentic AI authorization boundary requirement and for the audit trail in NPCI dispute resolution.
- Build a UAP-specific transaction audit trail. Standard UPI transaction logs attribute transactions to a user and a device. UAP requires attributing each transaction to the AI agent instance that executed it — including the agent's model version, the input it received, the decision logic it applied, and the consent parameters active at the time. This audit trail has two purposes: NPCI dispute resolution (proving the agent executed within consent parameters when a user disputes a transaction) and MRMF model governance (attributing agent behavior to a specific model version for performance monitoring). One audit trail design serves both purposes if it captures both dimensions from the start.
- Establish DPDP Act consent architecture for UAP agents. UAP user consent (for agent to transact) and DPDP Act consent (for agent to process user financial data) are separate obligations. UAP consent is a payment authorization; DPDP consent governs data processing. An AI agent that reads a user's transaction history to decide which bill to pay, or reads payment preferences to route a transaction, is processing personal financial data under DPDP Act. Purpose limitation applies: consent for an agent to make payments is not consent for that same agent to analyze spending patterns for a different purpose. Build DPDP consent architecture that separates these purposes before any agent processes user financial data.
- Register UAP agents in the MRMF model inventory for bank entities. For bank entities (scheduled commercial banks, payment banks, small finance banks with UPI member participation), UAP agents are AI models deployed by an RBI-regulated entity making consequential financial decisions. They must enter the MRMF model inventory with risk tier classification. A UAP agent executing ₹15,000/month transactions with direct customer financial impact and limited per-transaction oversight is likely MRMF-High risk. High-risk classification requires independent validation before production deployment — including testing of authorization scope enforcement, consent parameter compliance, and error handling for ambiguous user instructions.
- Define and test a UAP agent kill switch mapped to DPIP triggers. MRMF requires a tested kill switch for governed AI models. For a UAP agent, the kill switch mechanism is agent deregistration or suspension in NPCI's UAP registry. Define: what DPIP alert threshold or NPCI fraud signal triggers an agent suspension? Who in the PSP/bank has authority to execute the suspension? What is the target time from trigger to agent deactivation? Test this process before the agent goes to production scale. The PSP that cannot demonstrate a tested agent suspension capability faces NPCI examination risk when the formal UAP governance circular is issued.