What AI governance obligations apply to PPI issuers under the RBI PPI Master Directions 2021?

The RBI Master Directions on Prepaid Payment Instruments, 2021 (RBI/DPSS/2021-22/82, CO.DPSS.POLICY.No.S-479/02.14.006/2021-22, dated August 27, 2021, as amended November 2021, February 2023, and February 2024) create AI governance obligations for PPI issuers — banks issuing PPIs, non-bank PPI issuers, and NBFC wallet operators. Four AI systems are in scope: (1) V-CIP AI for Full-KYC PPI upgrades: MD Clause 9 requires Full-KYC PPIs to use an RBI-approved KYC process; AI-based V-CIP for KYC upgrade is subject to the RBI KYC Master Direction's face liveness and anti-spoof obligations, and the PPI issuer bears full regulatory liability for AI deepfake failures during the upgrade process; (2) Transaction velocity check AI: the MD's transaction limits (Rs 10,000/month for Small PPIs, Rs 2,00,000/month for Full-KYC PPIs) are enforced through AI velocity-check engines; these models are material decision models under MRMF 2026 and require board-approved model inventory, periodic validation, and PSI monitoring; (3) Fraud scoring AI: real-time AI fraud detection for PPI transactions — flagging suspicious payment patterns, unusual recipient chains, and velocity anomalies — is a model used in transaction-blocking decisions and must be in the MRMF inventory with independent validation and performance monitoring; (4) KYC upgrade recommendation AI: AI models that identify Small PPI holders eligible and ready for Full-KYC upgrade are models used in material customer decisions and require MRMF documentation.

How does the 2024 PPI amendment address AI model ownership in co-branding arrangements?

PPI co-branding arrangements — where a fintech or merchant co-brands with a bank or non-bank PPI issuer (PhonePe's co-branded cards with Yes Bank, Amazon Pay with Axis Bank, etc.) — were tightened by the February 2024 amendment to the PPI Master Directions. The amendment requires clear contractual delineation of responsibilities between the PPI issuer and the co-branding partner, specifically including responsibility for fraud detection and prevention systems. For AI governance, this creates a model ownership question: when a co-branded PPI uses an AI fraud scoring model, who is the MRMF model owner — the PPI issuer (the regulated entity with the RBI license) or the co-branding partner (who may have built and operates the AI model)? MRMF assigns model risk governance to the regulated entity's board. The PPI issuer cannot contractually assign its MRMF obligations to the co-branding partner — the model must be in the PPI issuer's MRMF inventory even if the co-branding partner developed and operates it. The 2024 amendment's delineation requirement means the co-branding agreement must: (a) identify which AI models in the co-branded product are governed under the PPI issuer's MRMF and which are the co-branding partner's own models; (b) give the PPI issuer audit rights over the co-branding partner's fraud AI; (c) establish change notification obligations so the PPI issuer can assess MRMF implications before any model update.

What is the Digital Payments Intelligence Platform and how does it create AI explainability obligations for PPI issuers?

The Reserve Bank of India has proposed the Digital Payments Intelligence Platform (DPIP) — a consortium-based real-time fraud intelligence system where transaction fraud signals are shared across payment system participants (banks, PPIs, UPI PSPs). DPIP is aligned with RBI's Payments Vision 2025 and builds on NPCI's existing Federated Fraud Risk Model (FRM) architecture. For PPI issuers integrating DPIP AI fraud signals into their transaction decisioning, an AI explainability obligation arises: when a DPIP-sourced AI fraud signal causes an automated transaction block on a customer's PPI wallet, the customer's dispute resolution request requires an explanation of why the transaction was blocked. The PPI issuer — not DPIP — is the entity with the customer relationship and the regulatory accountability for the block decision. If the block was based on a DPIP signal that the PPI issuer's own systems cannot explain, the issuer cannot satisfy RBI's payment dispute resolution requirements or the DPDP Act's transparency obligations for automated decisions affecting the customer. Governing this requires: (a) the PPI issuer must have access to the DPIP signal's reason code — not just the binary block/allow recommendation; (b) the issuer's dispute resolution process must include a pathway for explaining DPIP-signal-driven blocks to customers; (c) the issuer's MRMF documentation must cover the DPIP-integrated model as a third-party AI component under MRMF Section 7.

How does DPDP Act 2023 limit behavioral profiling AI for PPI limit enhancement?

PPI issuers use AI behavioral profiling to assess whether a Small PPI holder is eligible for Full-KYC upgrade and higher transaction limits — analyzing transaction frequency, recipient diversity, merchant category patterns, and account tenure as proxies for creditworthiness and financial behavior. Some PPI issuers have extended this profiling to include third-party behavioral data: purchase history from co-branded merchant platforms, geolocation data, app usage patterns. DPDP Act 2023's data minimisation principle (Section 6(2)) requires that personal data processing be limited to what is necessary for the stated purpose. For PPI limit enhancement AI, the stated purpose is the KYC upgrade decision — determining whether the customer is eligible for Full-KYC PPI limits. Third-party behavioral data (purchase history from an unrelated merchant, geolocation outside the payment transaction context) is not necessary for this purpose and requires explicit consent for the specific AI use. A behavioral profiling AI for PPI limit enhancement that uses third-party data without DPDP-compliant consent is in violation of Section 6(2). Additionally, PPI limit enhancement decisions that use AI-derived behavioral signals must be explainable to the customer under the DPDP Act's transparency principle — a customer denied PPI limit upgrade based on an AI behavioral score must be informed that an automated decision was made and, on request, the basis for it.

AI Governance · RBI PPI · Wallet AI · Fraud Scoring · MRMF · India BFSI

RBI PPI Master Directions 2021 AI Governance: Wallet Fraud Scoring, Velocity Check AI, Full-KYC Upgrade Obligations for PhonePe, Paytm and NBFC Wallet Issuers

India's 300+ million active PPI wallets operate under AI systems governed by the RBI PPI Master Directions 2021 (RBI/DPSS/2021-22/82). Full-KYC V-CIP AI liability falls entirely on the PPI issuer. Velocity check AI is a material decision model under MRMF. The 2024 amendment requires AI model ownership delineation in co-branding agreements. DPIP integration creates explainability obligations PPI issuers cannot outsource. DPDP data minimisation limits behavioral profiling for limit enhancement. Zero global competitor coverage.

Zytra Research Team · October 2026 · 12 min read

📖 Part of the India BFSI AI Governance Complete Guide 2026 — RBI, SEBI, IRDAI, NPCI & DPDP in one place.

India's Prepaid Payment Instrument (PPI) ecosystem — wallets issued by banks (HDFC PayZapp, SBI Yono Pay), non-bank PPI issuers (Ola Money, MobiKwik), and co-branded wallet products — processes billions of transactions annually. The RBI Master Directions on Prepaid Payment Instruments, 2021 (RBI/DPSS/2021-22/82, dated August 27, 2021, as amended) are the primary regulatory framework for all PPI issuers.

AI has become structural to PPI operations: fraud scoring at transaction time, velocity check enforcement for the MD's transaction limits, V-CIP-based Full-KYC upgrades, and behavioral profiling for limit enhancement decisions. Each of these AI systems has specific governance obligations under the PPI MD, MRMF, DPDP Act, and the RBI's Fraud Risk Management Master Directions (July 2024) that PPI issuers have not fully mapped.

Full-KYC V-CIP AI: The PPI Issuer's Full Liability

The PPI MD distinguishes between Small PPIs (minimum KYC, lower transaction limits) and Full-KYC PPIs (complete KYC, higher limits including Rs 2,00,000/month for full-KYC wallets). The upgrade from Small to Full-KYC requires a compliant KYC process. V-CIP is the standard non-face-to-face KYC channel for this upgrade.

The MD Clause 9 requirement that Full-KYC PPIs use an RBI-approved KYC process means the V-CIP AI for PPI KYC upgrades is subject to the same obligations as bank V-CIP AI under the RBI KYC Master Direction (DBR.AML.BC.No.81/14.01.001/2015-16): face liveness/anti-spoof detection at "high degree of accuracy," MRMF model card, independent validation, 8-year audit trail retention.

Two PPI-specific features make V-CIP governance particularly important for PPI issuers:

Deepfake upgrade fraud: A successful deepfake bypass of the PPI issuer's V-CIP AI does not merely open an account — it upgrades an existing Small PPI wallet to Full-KYC status, multiplying the fraudster's transaction limit from Rs 10,000/month to Rs 2,00,000/month instantly. The financial exposure from a deepfake-enabled upgrade fraud is directly tied to the V-CIP AI's anti-spoof performance.

Vendor liability gap: Many PPI issuers, particularly non-bank PPIs, outsource V-CIP to vendors. The PPI issuer's regulatory liability under the MD for KYC failures does not reduce because the AI model failed rather than a human. MRMF Section 7 applies: the vendor's V-CIP AI must be in the PPI issuer's model inventory, independently validated by the issuer, and governed under the issuer's MRMF framework.

Velocity Check AI: A Material Decision Model Under MRMF

The PPI MD's transaction limits are not static thresholds applied by a rule engine — they are enforced through AI velocity-check models that track transaction patterns in real time, identify attempts to circumvent limits through split transactions or multiple wallets, and flag accounts for review before limit breaches.

A velocity-check model that incorrectly blocks legitimate transactions (false positive) creates a customer experience and Consumer Protection Act exposure for the PPI issuer. A model that fails to catch genuine circumvention attempts (false negative) creates regulatory exposure for limit violations. Both error types are material — the model is making a binary decision with regulatory and customer consequences.

Under MRMF 2026, the velocity-check AI is a model used in material decisions and must be: in the board-approved MRMF model inventory, independently validated, and monitored with PSI/CSI metrics. The PPI issuer's risk committee should receive periodic performance reports on velocity-check false positive and false negative rates by customer segment.

Co-Branding AI Ownership: The 2024 Amendment Gap

The February 2024 amendment to the PPI MD tightened co-branding governance, requiring clear contractual delineation of responsibilities between PPI issuers and co-branding partners for fraud detection systems. This directly addresses a governance gap that had emerged: co-branded PPI products where the fraud AI was operated by the technology-capable co-branding partner (a fintech) while the regulatory accountability rested with the PPI issuer (the bank), but without formal governance connecting the two.

MRMF resolution for co-branded PPI AI:

The PPI issuer's MRMF model inventory must include co-branding partner AI models that are used in fraud detection, velocity check, or KYC for the co-branded product — they are third-party models under MRMF Section 7
The co-branding agreement must give the PPI issuer audit rights over the partner's fraud AI — documentation access for independent validation
Change notification: the partner must notify the PPI issuer before any material model change, giving the issuer time to assess MRMF implications
Performance reporting: the partner must provide the PPI issuer with fraud model performance metrics on the co-branded portfolio — not aggregate platform statistics

DPIP Integration: The Explainability Obligation That Cannot Be Outsourced

The DPIP (Digital Payments Intelligence Platform) consortium shares real-time fraud intelligence across payment system participants. PPI issuers integrating DPIP signals into transaction decisioning are effectively using a network-level AI model as a component in their fraud detection stack.

The governance challenge: DPIP is an industry utility; the PPI issuer is the regulated entity accountable to the customer for the transaction decision. When DPIP's AI signal causes a payment block on a customer's wallet, the customer's dispute is with the PPI issuer. The PPI issuer must be able to explain the block — at minimum to the customer's complaint, and potentially to RBI's payment ombudsman.

An explanation that amounts to "our network fraud intelligence flagged this transaction" is not sufficient for consumer dispute resolution. The PPI issuer must have access to the DPIP signal's reason code and must be able to translate it into a customer-understandable explanation within the dispute resolution SLA.

PPI AI Governance Checklist

☐ Full-KYC V-CIP AI: model card in MRMF inventory; independent validation conducted (not vendor statistics); deepfake adversarial testing; 8-year audit trail; INDIC-BIAS demographic liveness testing
☐ Velocity check AI: in MRMF model inventory as material decision model; board-approved; independently validated; PSI/CSI monitoring; false positive rate by customer segment tracked
☐ Fraud scoring AI: in MRMF inventory; independent validation; PSI monitoring; performance reporting on co-branded portfolio if applicable
☐ Co-branding AI governance: co-branding agreement includes audit rights, change notification, performance reporting, and MRMF inventory designation for partner's fraud AI
☐ DPIP integration: reason code access for DPIP-sourced blocks; explanation pathway for customer disputes; DPIP AI covered as third-party model under MRMF Section 7
☐ Behavioral profiling AI for limit enhancement: data minimisation audit — only data necessary for upgrade decision used; third-party behavioral data has DPDP consent; explainability for adverse upgrade decisions
☐ KYC upgrade recommendation AI: in MRMF inventory; documentation of recommendation criteria; false rejection rate by customer segment monitored
☐ Fraud Risk Management MD 2024 compliance: fraud AI governance aligned with RBI Fraud Risk Management Master Directions; incident reporting process for AI-related fraud events
☐ DPDP data processing agreement: data processing by co-branding partner AI models has DPDP lawful basis; purpose limitation documented for each AI use
☐ RBI examination readiness: MRMF model cards for all PPI AI models, validation reports, performance monitoring dashboards — packaged for examination