What did the RBI Authentication Directions 2025 change about AI in digital payment transactions?

RBI's Authentication Mechanisms for Digital Payment Transactions Directions, issued September 25, 2025 and effective April 1, 2026, made two significant changes relevant to AI. First, they moved from prescriptive SMS-OTP requirements to a principles-based two-factor authentication framework — payment system participants may now choose their authentication factors, as long as two distinct factors are used. This opens the door to behavioral biometrics, device fingerprinting, and other AI-native authentication methods that were previously constrained by OTP mandates. Second, the Directions explicitly permit AI/ML-based risk perception checks as an additional layer beyond the mandatory two-factor minimum. Issuers may deploy AI risk scoring calibrated to the fraud risk of each transaction — meaning a high-risk transaction (unusual beneficiary, atypical amount, new device) can be subjected to additional authentication friction while low-risk transactions (familiar payee, typical amount, known device) can proceed with lighter friction. This is India's first explicit regulatory permission for AI-driven adaptive authentication in payments.

Which entities are covered by the RBI Authentication Directions 2025?

The Directions apply to Payment System Providers and Payment System Participants, which includes: all banks operating payment accounts and issuing cards or enabling digital payment initiation; non-bank payment system operators (prepaid instrument issuers, payment aggregators, payment gateways); NBFC-P2P platforms; and any entity that processes digital payment transactions initiated through cards, mobile banking, internet banking, or other digital channels. The scope is intentionally broad: any entity that sits in the transaction chain for a digital payment and has a role in authentication is covered. This includes entities that had previously relied on their payment network's (Visa, Mastercard, RuPay) authentication standards without a separate RBI-compliant AI governance framework — those frameworks may not satisfy MRMF requirements for the AI risk scoring component.

Does RBI MRMF apply to AI fraud scoring models used in authentication?

Yes. An AI model that scores payment transaction fraud risk and whose output influences whether additional authentication friction is applied to a transaction is a material model under MRMF. It influences decisions that affect customers — incorrectly high-risk scoring causes legitimate transactions to be blocked or subjected to excessive friction; incorrectly low-risk scoring allows fraudulent transactions to proceed without adequate authentication. Both error directions cause material harm. Under MRMF, authentication AI models must: be included in the model inventory, have model documentation (model card, training data description, validation results), be subject to independent validation, produce explainable output (why was this transaction scored high-risk?), and be monitored in production (false positive rate, false negative rate, drift from training distribution). A bank that deploys AI risk scoring for authentication without this MRMF framework faces both an authentication compliance gap and a model governance gap.

What is the cross-border CNP transaction AI requirement in the RBI Authentication Directions?

The Directions require card issuers to validate Additional Factor of Authentication (AFA) for cross-border Card Not Present (CNP) transactions when the overseas merchant or acquirer requests it. For AI fraud scoring models, this creates a specific extension requirement: the model must produce fraud risk scores for cross-border transaction patterns, not just domestic Indian transactions. Cross-border CNP transactions have different risk profiles — unfamiliar merchant country codes, unusual currency conversions, cross-border shipping patterns, foreign IP address patterns — and a model trained primarily on domestic UPI or card transaction data will not have sufficient training coverage for these patterns. Banks deploying AI risk scoring must assess whether their models' validation includes cross-border CNP scenarios and document that assessment. MRMF's INDIC-BIAS testing framework should be extended to assess whether cross-border scoring creates disproportionate friction for legitimate international transactions by Indian cardholders.

AI Governance · RBI Authentication · Digital Payments · Fraud AI · India BFSI

RBI Authentication Directions 2025: AI Risk-Based Authentication, Fraud Scoring, and the Compliance Framework India Payment Banks Must Build

RBI's Authentication Mechanisms Directions (September 2025, effective April 2026) explicitly permit AI/ML risk scoring as an additional authentication factor for digital payments — the first binding Indian regulatory permission for adaptive AI-driven authentication. Every bank and payment system operator deploying AI fraud scoring now has MRMF validation obligations attached to a compliance deadline.

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 digital payments infrastructure processes over 15 billion UPI transactions per month. Card-based payments add hundreds of millions more. Every one of these transactions is — or should be — subject to authentication. RBI's approach to authentication has historically been prescriptive: SMS-OTP as the mandated second factor for most transaction types, with limited flexibility for alternatives.

RBI's Authentication Mechanisms for Digital Payment Transactions Directions, issued September 25, 2025 and effective April 1, 2026, changed this architecture fundamentally. The move to a principles-based framework — two distinct factors, freely chosen by the payment system participant — combined with explicit permission for AI/ML risk scoring as an additional factor, creates both an opportunity and a governance obligation.

The opportunity: adaptive authentication that applies the right level of friction to the right transaction — high friction for high-risk transactions, low friction for low-risk ones — enabling a better customer experience alongside better fraud prevention. The obligation: every AI model that contributes to this authentication decision is now a material model under MRMF, with the full documentation, validation, and monitoring apparatus that entails.

What Changed: The Shift from Prescriptive to Principles-Based

The pre-2025 authentication framework was built around SMS-OTP. For most digital payment transactions, two-factor authentication meant a static credential (password, PIN) plus a dynamic one-time password delivered by SMS. This was simple, consistent, and easy to audit — but it was also inflexible. Behavioral biometrics, device-native authentication, and risk-based adaptive friction were technically feasible but existed in a regulatory gray area.

The 2025 Directions resolve this ambiguity. They require:

Two distinct factors of authentication for all digital payment transactions — the two-factor floor remains
Factor independence: the two factors must be distinct in type (knowledge + inherence, knowledge + possession, possession + inherence) — two PINs do not satisfy the requirement
Participant choice of factors: within the two-factor floor, payment system participants choose their authentication methods — the prescriptive SMS-OTP requirement is replaced by a principles-based obligation
Explicit AI/ML permission: issuers may deploy AI/ML-based risk perception checks as an additional layer calibrated to the fraud risk of the underlying transaction

The last provision is the most significant for AI governance. "Additional layer" means the AI risk score sits above the two-factor minimum — it is not one of the two required factors, but a supplemental check that can escalate authentication requirements for high-risk transactions or reduce friction for low-risk ones within a trusted device and customer profile. This is India's regulatory permission for adaptive authentication.

AI Fraud Scoring as an Authentication Component: The Governance Implications

MRMF applies to authentication AI

An AI model that scores payment transaction fraud risk and influences authentication decisions is a material model under MRMF. The impact is bilateral: false positives (incorrectly high-risk scores) cause legitimate transactions to face excessive authentication friction or be blocked entirely — customer harm. False negatives (incorrectly low-risk scores) allow fraudulent transactions to proceed with insufficient authentication — financial harm to the customer and the institution.

MRMF's requirements apply in full:

Model inventory entry: purpose, scope, deployment context, risk tier
Model documentation: training data, feature selection, validation methodology, performance metrics
Independent validation: conducted by a team or party independent of model development
Explainability: the model must produce an explanation for each score — which transaction features drove the high-risk assessment
Ongoing performance monitoring: false positive rate, false negative rate, precision, recall — monitored continuously in production; thresholds defined for escalation

The explainability requirement for authentication AI

MRMF Section 4's explainability requirement — designed primarily for credit AI — applies to authentication AI with an important operational dimension. When an AI fraud score causes a transaction to be blocked or subjected to step-up authentication, and a customer disputes the treatment, the bank must be able to explain why the transaction was scored high-risk.

This is operationally more demanding for authentication AI than for credit AI, because the explainability must be generated and stored at transaction processing speed — there is no post-hoc time window for retrospective generation. The model must produce explanations at inference time: which features (device fingerprint mismatch, unusual payee, atypical time of day, geographic anomaly) drove the risk score, and how each contributed.

The explanation also feeds the dispute resolution process: a customer who disputes an incorrectly blocked transaction needs a meaningful answer about why their transaction was flagged. "Our AI flagged it" is not a compliant response under FREE-AI's consumer protection requirements.

Behavioral Biometrics and Device Fingerprinting: Governance Considerations

The 2025 Directions' principles-based framework enables behavioral biometrics (typing rhythm, screen pressure, swipe velocity) and device fingerprinting (device ID, OS version, app version, network characteristics) as authentication factors. These are increasingly deployed by Indian banks' mobile apps — the 2025 Directions provide the regulatory authorization that previously required case-by-case RBI guidance.

The governance implications:

DPDP Act intersection. Behavioral biometrics and device fingerprinting involve collection of personal data beyond what is strictly necessary for the transaction. Under DPDP Act 2023, this data collection requires a lawful basis — most banks will use contractual necessity (authentication is required to execute the payment) or legitimate interest (fraud prevention). The data minimization principle requires collecting only the behavioral data necessary for the authentication function: a bank cannot collect comprehensive behavioral profiles "in case they're useful later."

Consent and transparency. FREE-AI's consumer protection requirements (Pillar 5) require customers to be informed when AI plays a material role in decisions affecting them. An AI model using behavioral biometrics to assess fraud risk is a material AI system — customers should know that behavioral patterns are being used. The disclosure need not specify the model's technical architecture, but the practice should be disclosed in the bank's privacy notice and terms of service in plain language.

Bias testing for behavioral data. Behavioral biometrics models trained on a narrow population may score differently for users who interact differently with mobile interfaces — elderly users with different typing patterns, users with physical disabilities, users from regions where phone usage conventions differ. INDIC-BIAS testing framework should be applied to authentication AI: are false positive rates elevated for specific customer demographics?

Cross-Border CNP Transactions: The International Extension Requirement

The Directions require card issuers to provide Additional Factor of Authentication for cross-border Card Not Present transactions when the overseas merchant or acquirer requests it. For AI fraud scoring models, this creates a specific validation requirement: the model must be validated on cross-border CNP transaction patterns, not just domestic transactions.

The practical challenge: most Indian banks' AI fraud models are trained primarily on domestic transaction data. UPI transactions, domestic card transactions, IMPS/RTGS transfers — this is the data that is plentiful. Cross-border CNP transactions are a smaller fraction of transaction volume and represent different fraud patterns: shipping address manipulation, foreign IP address anomalies, cross-border merchant category patterns, currency conversion fraud.

A model that produces high-quality risk scores for domestic transactions may be poorly calibrated for cross-border CNP scenarios. Banks deploying AI risk scoring must:

Assess whether their model's training data includes sufficient cross-border CNP transactions for reliable scoring
Validate the model specifically on cross-border CNP test cases
Document whether the model's performance meets acceptable thresholds for cross-border CNP scenarios
If the model is not validated for cross-border CNP, implement a fallback authentication approach (step-up to human review or additional factor) for cross-border CNP transactions pending model improvement

AI Authentication Governance Checklist: RBI Directions 2025

☐ Authentication factor inventory: all factors in use documented; two-factor independence verified per Directions
☐ AI fraud scoring model in MRMF model inventory: risk tier assigned; model card complete
☐ Training data documentation: data sources, training period, transaction types covered (domestic vs. cross-border CNP coverage assessed)
☐ Independent validation completed: validation methodology, test dataset, performance metrics (precision, recall, F1, AUC at operating threshold)
☐ Inference-time explainability: model produces feature-level explanation per transaction score; explanation stored in transaction log
☐ Production monitoring: false positive rate and false negative rate tracked daily; alert thresholds defined; escalation process documented
☐ Cross-border CNP validation: specific validation on cross-border CNP test cases; performance threshold met or fallback authentication documented
☐ Behavioral biometrics DPDP alignment: lawful basis documented; data minimization assessment; customer disclosure in privacy notice
☐ INDIC-BIAS testing: false positive rate by demographic segment; elevated rates for elderly, differently-abled, regional usage patterns assessed
☐ Dispute resolution: process for explaining blocked transactions to customers; routes to human review; explanation based on model output, not generic "AI flag"
☐ Customer disclosure: AI fraud scoring disclosed in terms/privacy notice; plain language; FREE-AI Pillar 5 compliant
☐ Third-party authentication AI: vendor contract AI-specific provisions (MRMF validation access, update notification, right to audit)
☐ RBI Directions effective date: April 1, 2026 compliance confirmed; legacy OTP-only systems assessed for Directions alignment

The 2025 Directions represent a significant regulatory maturation in India's digital payments governance: a shift from prescriptive rules to principles-based flexibility, with AI explicitly recognized as a tool for better authentication outcomes. The governance obligation that accompanies this permission — MRMF validation for authentication AI, DPDP compliance for behavioral data, INDIC-BIAS testing for fair authentication — is the price of that flexibility. Banks that implement the permission without the governance apparatus face both the compliance exposure and the operational risk of unvalidated AI in a high-stakes real-time decision context.