Head of Product – Cards Business


# Real-Time Payments in MENA 2025: Adoption, Rails, and Use Cases

The Middle East and North Africa region is experiencing a transformation in payment infrastructure. Real-time payment (RTP) systems are no longer a future promise—they're reshaping how consumers and businesses transact today.

## The Infrastructure Landscape

Saudi Arabia's **SARIE** (Saudi Riyal Interbank Express) system processes millions of instant payments monthly, offering 24/7 settlement with sub-second response times. The UAE's **UAEFTS** follows suit, enabling instant fund transfers across the Emirates' banking ecosystem.

Egypt's **InstaPay** has rapidly gained traction since its 2022 launch, processing over 10 million transactions in its first year. These systems share common characteristics:

- **Always-on availability**: No batch windows or weekend delays
- **Immediate settlement**: Funds available within seconds
- **ISO 20022 messaging**: Modern, data-rich transaction formats
- **Interoperability**: Cross-bank transfers without friction

## Adoption Drivers

### Consumer Demand

The shift to real-time is consumer-driven. In markets where cash remains king, instant digital payments offer a compelling alternative. Users want:

- **Instant confirmation**: No waiting to verify if rent payment cleared
- **Evening and weekend transfers**: When traditional systems are offline
- **Rich remittance data**: Full transaction context, not cryptic codes

### Business Use Cases

Enterprises are leveraging RTP for:

- **Just-in-time supplier payments**: Optimize cash flow without sacrificing relationships
- **Gig economy disbursements**: Pay delivery partners immediately after shift completion
- **Request-to-pay flows**: Invoice presentation with embedded payment capability

## Technical Architecture

Most MENA RTP systems follow a hub-and-spoke model. The central switch:

1. **Receives** payment instructions from sending bank
2. **Validates** account details and available balance
3. **Routes** to receiving institution
4. **Confirms** credit to beneficiary account
5. **Settles** via central bank liquidity accounts

Response times average 2-5 seconds end-to-end, including fraud screening.

## Challenges Ahead

Despite progress, obstacles remain:

- **Liquidity management**: Banks must maintain real-time settlement balances
- **Fraud prevention**: Instant is permanent—no time for manual review
- **Legacy integration**: Core banking systems weren't built for 24/7 operation
- **Cross-border complexity**: National systems don't yet interoperate regionally

## What's Next

The next phase involves:

1. **API-first access**: Fintech and corporate direct participation
2. **Request-to-pay standardization**: Unified invoice + payment flows
3. **Regional interoperability**: GCC-wide instant payment network
4. **Embedded finance**: RTP as infrastructure for buy-now-pay-later and wallets

Real-time payments aren't just faster—they enable entirely new financial services. As adoption deepens across MENA, we'll see innovation in credit, commerce, and cross-border flows that weren't possible in the batch-processing era.

The infrastructure is here. Now comes the hard part: building products that make instant payments indispensable.


Exploring the rapid growth of instant payment systems across the Middle East and North Africa, from infrastructure to consumer adoption patterns.

Real-Time Payments in MENA 2025: Adoption, Rails, and Use Cases


# AI-Driven Fraud: Velocity, Behavior Analytics, and Graph Signals

Traditional fraud rules—"decline if transaction exceeds $5,000" or "flag if more than 3 transactions in 10 minutes"—are breaking under the weight of modern attack sophistication. Today's fraud prevention requires machine learning models that adapt in real-time, behavioral fingerprinting, and graph-based network analysis.

## The Evolution of Fraud Detection

### Rule-Based Systems (Generation 1)

Static thresholds and simple logic:
- Amount limits by merchant category
- Geographic restrictions
- Simple velocity checks

**Problem**: Attackers easily reverse-engineer and evade fixed rules.

### Machine Learning Models (Generation 2)

Supervised learning on historical fraud labels:
- Gradient-boosted decision trees
- Neural networks for pattern recognition
- Feature engineering: time-of-day, merchant metadata, card-not-present flags

**Problem**: Models lag reality. Yesterday's fraud patterns don't predict tomorrow's attacks.

### Behavioral + Network Analytics (Generation 3)

This is where we are today. Systems that:
- Build baseline behavioral profiles per cardholder
- Detect anomalies from expected patterns
- Analyze transaction networks to spot coordinated attacks

## Velocity Checks 2.0

Simple velocity—"3 transactions in 10 minutes"—is too blunt. Modern systems track:

- **Multi-dimensional velocity**: Cards, devices, IP addresses, merchants
- **Contextual windows**: Faster spending on Friday night vs Tuesday morning
- **Cross-entity patterns**: Same shipping address used across 50 different cards

Example: A card used at 10 gas stations in 2 hours might be normal for a fleet manager, catastrophic for a consumer card. Behavioral context matters.

## Behavior Analytics

Every cardholder develops patterns. ML models learn:

- **Merchant preferences**: Coffee shops daily, luxury retailers quarterly
- **Geographic zones**: Home, office, commute route, vacation destinations
- **Transaction timing**: Lunch purchases at noon, not 3 AM
- **Spend magnitude**: $200/transaction is normal for one user, alarming for another

When a transaction deviates significantly from learned behavior, the system:
1. **Scores** the anomaly (how far outside normal?)
2. **Weighs risk factors** (card-present vs CNP, high-risk merchant category)
3. **Decides**: approve, decline, or step-up authentication (3DS challenge)

## Graph-Based Fraud Detection

Fraudsters operate in networks. Graph analysis reveals:

### Device Fingerprinting Networks

Track relationships between:
- **Cards** → linked by shared device IDs
- **Devices** → linked by IP addresses or browser fingerprints
- **Merchants** → linked by suspicious refund patterns

When one card in a cluster is confirmed fraud, the entire network gets elevated scrutiny.

### Shipping Address Graphs

Map addresses to:
- Number of unique cards shipping there
- Frequency of returns or chargebacks
- Geographic proximity to known fraud rings

An address receiving shipments from 100+ different cards over 2 weeks is a **reshipping mule location**—a logistics hub for stolen goods.

### Velocity Cascade Detection

Track how compromised credentials move:
1. Card A makes 5 small test transactions (checking if card is active)
2. Same device immediately switches to Card B, C, D with similar patterns
3. Graph analysis spots the cascade before individual cards hit velocity limits

## Real-Time Decisioning

All of this must happen in **under 100 milliseconds**—the SLA for payment authorization. Architecture:

- **Pre-computed features**: Behavioral profiles updated hourly, not at transaction time
- **Model serving infrastructure**: ONNX or TensorFlow Serving for low-latency inference
- **Feature stores**: Redis or similar for real-time lookups (device history, IP reputation)
- **Fallback rules**: If ML model times out, hard rules apply

## The False Positive Challenge

Fraud models balance two costs:

- **False negatives**: Approving fraud (direct financial loss)
- **False positives**: Declining legitimate transactions (customer friction, lost revenue)

A 1% false positive rate sounds acceptable until you realize it means declining 1 in 100 legitimate customers. At scale, that's thousands of angry users daily.

Modern systems use:
- **Confidence scores**: High-confidence fraud declines immediately; borderline cases get step-up auth
- **Adaptive thresholds**: Tighten rules during known attack periods, relax during low-risk hours
- **Feedback loops**: Chargeback data retrains models weekly

## What's Next

The frontier is:

- **Federated learning**: Train models across institutions without sharing customer data
- **Generative AI for attack simulation**: Red-team models that invent new fraud patterns
- **Real-time behavioral biometrics**: Typing cadence, mouse movements as fraud signals
- **Cross-border network analysis**: Spot international carding rings across payment systems

Fraud prevention is an arms race. Static defenses lose. Adaptive, intelligent systems that learn from every transaction are the only path forward.


How modern fraud detection systems use machine learning, behavioral analysis, and network graphs to stop sophisticated attacks in real-time.

Trends

Real-Time Payments in MENA 2025: Adoption, Rails, and Use Cases

AI-Driven Fraud: Velocity, Behavior Analytics, and Graph Signals