RBI Payment System Provider AI Governance 2026: PSP, PPI, and PA Obligations Under MRMF and Cybersecurity Directions

India's payment ecosystem is one of the world's most AI-intensive financial infrastructures. UPI alone processed over 14 billion transactions in a single month in FY2026 — and each transaction passes through multiple AI models before settlement. Fraud detection engines, merchant risk scoring systems, KYC verification AI, and adaptive transaction limit models operate across Payment System Operators, Payment System Providers, Prepaid Payment Instruments, and Payment Aggregators. The regulatory question — which RBI AI governance frameworks apply to which entity — has not been formally resolved. This analysis maps the obligation landscape entity by entity.

India's Payment AI Stack

Entity Types and Their Regulatory Anchors

India's payment ecosystem spans four distinct entity categories, each regulated under different RBI instruments:

• Payment System Operators (PSOs): NPCI (UPI, RuPay, IMPS, NACH) and CCIL — systemically important infrastructure operators under direct RBI oversight. NPCI's AI governance obligations are among the most significant in the Indian financial system given the scale of rails it operates.
• Payment System Providers (PSPs): Authorized third-party UPI app operators — PhonePe, Google Pay, Paytm, and others — regulated under PSP Authorization. PSPs are non-bank entities; they do not hold settlement accounts directly and operate through sponsor banks.
• Prepaid Payment Instruments (PPIs): Paytm Payments Bank, Airtel Payments Bank, Jio Payments Bank, and non-bank PPI issuers — regulated under the PPI Master Direction 2021. Bank PPIs are also subject to full bank regulation including MRMF; non-bank PPIs have a more limited obligation set.
• Payment Aggregators (PAs): Razorpay, PayU, Cashfree, CCAvenue — regulated under the Payment Aggregator Directions 2020 (updated 2023). PAs sit between merchants and acquiring banks; their merchant onboarding and transaction risk AI is high-stakes and under increasing regulatory scrutiny.

Five AI Use Cases Across the Payment Stack

Which RBI Frameworks Apply — Entity by Entity

PSPs: UPI App Operators

PSPs are authorized non-bank entities. RBI's MRMF (June 2026) was issued to regulated entities — banks and NBFCs. PSPs are not directly in scope of MRMF. This does not mean PSPs are ungoverned:

• PSPs operating under bank infrastructure (most major UPI apps use bank partners for settlement) face MRMF obligations through their bank partner — the sponsor bank's MRMF compliance extends to AI systems the PSP operates that affect the bank's settlement function.
• PSP fraud detection AI affects the banking system — RBI has signaled it will extend AI governance to payment system participants. PSPs should treat MRMF as the anticipatory standard for their AI governance even if not yet directly bound.
• Cybersecurity Directions 2023 explicitly cover PSPs — any AI used in security monitoring, anomaly detection, or incident response is within scope.
• DPDP Act applies unconditionally to PSPs — transaction behavioral data used for fraud detection and limit setting is personal data, and DPDP Act obligations for AI processing this data are direct and immediate.

PPIs: Prepaid Payment Instruments

PPI governance is bifurcated by entity type. Bank PPIs (payments banks) are directly under bank regulation including MRMF — Paytm Payments Bank, Airtel Payments Bank, and Jio Payments Bank are banks and carry the full MRMF obligation set. Non-bank PPI issuers face a different landscape:

• PPI Master Direction 2021 does not explicitly reference MRMF, but its KYC and cybersecurity provisions create overlapping obligations.
• Cybersecurity Directions 2023 cover non-bank PPIs — AI in security monitoring is in scope.
• FREE-AI Fairness Sutra: any PPI using AI for transaction risk scoring that affects customer access faces fairness obligations — the INDIC-BIAS framework applies to adaptive limit models.
• RBI's KYC directions for PPIs require PMLA-compliant verification — AI-based KYC must meet Prevention of Money Laundering Act standards regardless of automation method.

PAs: Payment Aggregators

PA Directions 2020 and 2023 establish comprehensive cybersecurity obligations but do not yet explicitly reference RBI's AI governance frameworks. However, PA AI governance obligations are substantial:

• Merchant risk scoring AI affects bank partners — RBI oversight extends through the chain, and acquiring bank MRMF obligations cover AI models used in the PA-bank interface.
• CERT-In 6-hour reporting: PAs must report cyber incidents including AI system compromises — a fraud model failure exploited by adversarial actors triggers reporting.
• DPDP Act: PAs process massive volumes of customer transaction data — DPDP Act obligations for AI processing this data are direct and immediate, with DPIA requirements for high-risk large-scale processing.

Fraud Detection AI: MRMF Application to Payment Fraud Models

Payment fraud detection AI is one of the highest-stakes AI deployments in India BFSI. Monthly UPI transaction volume exceeds 14 billion (FY2026 projection). At a 0.01% fraud rate, that is 1.4 million fraudulent transactions per month — and AI is the primary defense layer. Model governance for payment fraud AI must address four specific obligations:

Model Documentation

Fraud detection models must document training data sources, feature set, target variable definition (the fraud label), training period, and known limitations. Data sourcing from NPCI's fraud data consortium requires documenting consortium data terms and any data sharing limitations that affect model completeness. A model trained only on consortium-shared fraud patterns will underperform against emerging fraud typologies not yet shared.

Performance Monitoring

MRMF requires ongoing monitoring, and fraud models decay faster than credit models because fraud patterns evolve. Monthly or even weekly performance review is appropriate for payment fraud AI. The performance monitoring cadence should be documented in model governance policy and calibrated to the transaction volume and fraud typology mix of the entity.

False Positive Management

MRMF's control effectiveness framework requires measuring false positive rate. Payment fraud AI false positives — blocking legitimate transactions — cause direct customer harm and generate regulatory attention. RBI has specifically highlighted false positive rates in payment fraud AI as a governance metric, not merely an operational one. Entities must track false positive rates by channel, transaction type, and customer segment and report them in model performance dashboards.

Adversarial Testing

MRMF validation must include adversarial scenarios. For fraud detection AI, this means testing against known fraud patterns including synthetic identity fraud, account takeover, and carousel fraud — which is common in the India UPI ecosystem. CERT-In Adversarial Robustness guidelines apply to AI used in financial security functions.

UPI Transaction Limit AI: Adaptive Limits and NPCI Guardrails

Many UPI app providers use AI to dynamically set transaction limits for individual users — higher limits for established users with clean transaction histories, lower limits for new users or accounts with anomalous behavioral patterns. This is a governance-intensive use case with multiple regulatory anchors:

NPCI Regulatory Floors

NPCI sets maximum transaction limits for UPI. AI-driven adaptive limits must operate within NPCI-set ceilings — the AI cannot override regulatory maximums, only personalize limits below the ceiling. Documentation of the constraint architecture (how the AI is constrained to NPCI limits) is part of model documentation requirements.

Fairness Obligation: INDIC-BIAS Testing

If the AI model systematically assigns lower transaction limits to users based on proxy demographic characteristics — name, location, phone number registration area, device type — this violates FREE-AI Fairness Sutra. The INDIC-BIAS framework applies specifically to India context proxy variables. PSPs and PPIs using adaptive limit AI should test quarterly for proxy discrimination across geographic, linguistic, and demographic dimensions.

New-to-Credit and New-to-Digital Population Risk

India's largest unbanked population is also a significant UPI user base. AI models trained primarily on established user behavior will systematically underperform for new-to-credit users with thin digital histories — creating a bias toward lower limits for the population that most needs payment access. Validation must explicitly include new-to-digital user population testing, not just aggregate performance metrics.

Explainability for Limit Decisions

If a user is denied a higher transaction limit by AI, they have an implicit right to understand why under RBI fair practices expectations. PSPs should build explanation capability for limit decisions — "your limit is set based on your account age and transaction history" is materially more compliant than unexplained limits that appear arbitrary to users who have been denied access.

Merchant Risk Scoring AI: PA-Specific Governance

Payment Aggregators use AI for merchant onboarding risk scoring — whether to approve a merchant, at what transaction limits, and with what escrow requirements. This is financial credit risk AI with significant and growing regulatory attention:

PA Directions 2023 KYC Obligations

Merchant onboarding AI must comply with PMLA KYC requirements. AI-based document verification and KYC cannot replace the regulatory KYC process — AI can automate the process but must meet the same verification standards as manual processes. A merchant rejected by AI-based KYC must have a human review pathway available.

Merchant Fraud Ring Detection

Graph AI used to identify connected merchant fraud rings is sophisticated but carries significant false positive risk — a legitimate merchant group can be incorrectly flagged as a fraud ring based on shared infrastructure, common directors, or geographic clustering. MRMF validation should include false positive rate measurement specifically for merchant fraud ring detection, with attention to industry sector concentration effects.

Escrow Governance

PAs hold merchant funds in escrow before settlement. AI that adjusts escrow release timing based on risk scoring must document the escrow decision logic. Regulatory examination can require a PA to justify escrow decisions on a merchant-by-merchant basis — the AI decision must be reconstructible and explainable, not a black-box output.

RBI Merchant Onboarding Scrutiny

RBI has increased scrutiny of PA merchant onboarding following several high-profile payment fraud cases involving compromised merchants. AI-assisted onboarding decisions will face examination — documentation of the AI decision process and the ability to explain individual decisions is regulatory defense, not optional enhancement.

Five Priority Controls for Payment AI Governance

1. Fraud AI performance dashboard: An automated weekly report on fraud detection model performance — fraud detection rate, false positive rate, false negative rate, broken down by transaction type and channel. This feeds the MRMF performance monitoring obligation and provides a continuous audit trail of model performance over time.
2. INDIC-BIAS testing for transaction limit AI: A quarterly proxy discrimination test for any AI that affects customer transaction access — adaptive limits, risk scoring, KYC decisions. Test specifically for geographic, linguistic, and demographic proxy variables using the INDIC-BIAS framework developed for India BFSI context.
3. Merchant risk AI documentation: A model card for every merchant risk scoring model that includes training data source, full feature set, performance metrics, validation history, and known limitations. This is required for both internal governance and RBI examination — a PA without model cards for its merchant scoring AI is unprepared for examination.
4. Incident classification tree: A decision tree that classifies any AI system anomaly — fraud model accuracy drop, KYC false positive spike, transaction limit algorithm anomaly — by reporting obligation: internal escalation, CERT-In 6-hour reporting, RBI incident reporting, NPCI notification, and future India AISI reporting. Build this before an incident occurs, not during one.
5. DPDP consent architecture for transaction data: Transaction behavioral data used for AI risk scoring is personal data under DPDP Act. A documented consent framework must specify which personal data is used for which AI model — fraud detection, limit setting, and dispute resolution must have separate documented processing bases, distinguishing between legitimate interest (for fraud prevention, where the RBI regulatory mandate supports processing) and explicit consent (for enhancement purposes).

Frequently Asked Questions