AI Governance for Payment Systems Banking: PSD3, Real-Time Payments Fraud, and APP Fraud AI Controls

What is settled is the direction of travel: the agreed PSR text makes PSPs liable to reimburse consumers for losses from authorised push payment (APP) fraud and impersonation scams where the PSP failed to implement adequate fraud-prevention measures, and it introduces a mandatory "verification of payee" check on credit transfers. Banks that wait for the final Official Journal date to start building AI fraud controls will be building them under examination pressure instead of on their own schedule.

Payment fraud AI governance is the highest-throughput governance domain in banking, full stop. A major retail bank processes hundreds of millions of transactions a year, each one needs a fraud call in milliseconds, and every wrong call — a fraud let through, or a legitimate payment blocked — has a direct customer impact and, once PSR's liability framework lands, a direct line to the bank's P&L.

The governance tension here is structural: fast AI tends to be less accurate, accurate AI tends to be slower, and explainable AI is often both slower and less accurate than the opaque model that would have scored higher on a leaderboard. Real-time payments remove the one thing that used to protect banks from their own fraud models — time. There's no window to escalate to a human analyst or call the customer before a UPI or Faster Payments transaction settles in seconds. The AI's decision is the decision, full stop, which is exactly why the governance has to happen before deployment, not after an incident.

Whatever the final PSR liability threshold turns out to be, the economics already favour AI investment in APP fraud detection — industry estimates widely put APP fraud at more than half of total European payment fraud volume, and a PSP that can't detect it will eventually absorb losses it currently pushes onto consumers. The governance question isn't whether to build AI fraud detection; it's whether that AI is validated, monitored, and documented well enough to survive both a PSR-style liability test and a PSD3 explainability demand from a customer whose payment got blocked.

Four AI Payments Governance Obligations

Real-Time Fraud Model Governance

Fraud models in real-time payment rails need governance built around their actual operational constraints, not borrowed wholesale from batch credit risk models. Validation has to check latency alongside detection accuracy — a model that hits 95% fraud detection at 800ms isn't viable if the rail needs a decision inside 200ms, no matter how good the accuracy number looks on a slide. Production monitoring needs both detection performance (catch rate, false positive rate) and latency performance (p50/p95/p99) tracked together, because a model can degrade on either axis independently.

Drift monitoring for payment fraud AI has to catch three distinct failure modes: data drift (a new payment channel or type shifts the transaction distribution), performance drift (detection accuracy degrading on its own), and adversarial drift (fraudsters adapting to beat the current model). Adversarial drift is the hardest to catch because it looks like good news at first — a sudden drop in fraud catch rate at a stable false positive rate usually means fraudsters have found the model's decision boundary and are now designing transactions specifically to sit just outside it.

APP Fraud Detection Governance

APP fraud needs a governance framework distinct from unauthorized-transaction fraud detection, because the customer authorized the payment themselves — from the model's point of view, nothing looks wrong. The deception sits outside the transaction entirely: a fraudster has convinced the customer the payment is legitimate. Detecting it requires modelling deception indicators instead of transaction anomalies — a newly registered payee, an unusually large amount for this customer, unusual inbound contact immediately before the payment, a payment made outside the customer's normal hours.

APP fraud AI needs its own validation track measured against deception-detection performance specifically, not folded into general fraud detection metrics. The false-positive cost here is unusually sharp: blocking a legitimate large payment to a new payee — a first home deposit, a new supplier relationship — is genuinely disruptive to a real customer, so thresholds need documented, deliberate calibration between APP fraud catch rate and legitimate large-payment throughput, not a default tuned purely for fraud catch.

SCA Risk-Based Exemption Governance

PSD3/PSR's eventual rules on Strong Customer Authentication exemptions via transaction risk analysis (TRA) will only let a PSP skip SCA when its fraud rate for that transaction value band sits below a regulatory threshold. That makes the AI exemption engine's fraud-rate performance a continuous compliance obligation, not a one-time calibration: if the exempted-transaction fraud rate rises past the threshold, SCA has to come back automatically for that band, not after someone notices the breach in a quarterly report. Regulators can and will ask how a specific exempted transaction was assessed as low-risk, so the engine needs to produce a risk-assessment explanation per transaction that holds up under examination, not just an aggregate fraud-rate dashboard.

Payment Fraud Explainability for Consumer Rights

PSD3's direction of travel requires PSPs to explain why they blocked or flagged a payment as suspicious — and "our systems assessed this as high-risk" isn't an explanation, it's a refusal to give one. The real explanation has to name the actual trigger — unusual payee, unusual amount, unusual timing, device anomaly — translated into plain language a customer can act on. That means feature-level attribution (SHAP or an equivalent method) has to sit in the production decision pipeline, not get bolted on afterward as a compliance exercise. The trick is doing this without handing fraudsters a blueprint: describe the category of concern ("unusual recipient") to the customer, not the specific model feature ("payee registered fewer than 30 days ago with zero prior transaction history") that a fraudster could use to engineer around the control.

India's UPI Scale Changes What "Real-Time" Means

UPI processes hundreds of millions of transactions a day, making it the highest-volume real-time payment system anywhere by transaction count. That scale creates a fraud governance problem with no real international equivalent: sub-100ms latency, across a population spanning sophisticated urban smartphone users and first-time rural users with no transaction history on feature phones. India's banking system is building the operational muscle for real-time payment AI governance faster than anywhere else, mostly because it has had to.

NPCI sets the fraud risk management framework UPI-participating banks operate under — transaction monitoring AI, incident reporting to NPCI within defined timelines, and participation in shared fraud intelligence. A bank whose AI model produces too many false positives doesn't just annoy its own customers; it drags down NPCI's published success-rate metrics for the whole network, which is a different kind of pressure than most banks' internal governance committees are used to applying.

India's APP fraud landscape has its own texture: fake customer-care calls impersonating bank staff, WhatsApp investment scams, SIM-swap fraud against UPI-registered numbers, and QR code substitution at the point of sale. Fraud detection AI tuned on US or European scam scripts will miss social engineering conducted in Hindi, Tamil, or Telugu — multilingual pattern detection for APP fraud is a governance requirement that genuinely has no clean international template to borrow from. RBI has issued specific UPI APP fraud prevention guidance, and participating banks are expected to have implemented the recommended controls, not just acknowledged them.

UPI transaction data is sensitive personal financial data, and training fraud models on it means documenting a legal basis under the Digital Personal Data Protection Act 2023 — most likely legitimate interest for fraud prevention, but that basis has to be written down and assessed for proportionality, not assumed. Processing has to be limited to the stated purpose, and retention limited to what fraud investigation and regulatory obligations actually require — not "as long as it might be useful someday."

What to Do About It

• Start building PSD3/PSR-aligned fraud governance now against the agreed direction of travel, rather than waiting for the Official Journal date — examiners reward banks that can show the work in progress.
• Separate your APP fraud detection validation track from general fraud detection metrics, with its own documented false-positive threshold calibrated against real customer disruption, not just catch rate.
• Build per-transaction explainability for SCA exemption decisions now, in plain language, before a regulator asks for one you don't have.
• Stand up adversarial-drift monitoring that looks specifically for "catch rate falling at stable false-positive rate" as its own alert condition, distinct from generic performance drift.
• For UPI-scale deployments, document your DPDP Act legal basis for fraud-model training data in writing before your next model refresh, not retroactively.

Zytra governs payment fraud AI with real-time model validation, APP fraud detection governance, SCA exemption engine monitoring, fraud explainability aligned to PSD3's direction of travel, and UPI-scale governance — producing the evidence PSR's coming liability framework and RBI/NPCI guidelines require.

Talk to sales