Reducing Authorised Push Payment (APP) fraud by 92% using high-performance predictive ML.

The Challenge

A London-based fintech processing £500M+ in monthly transactions was facing a surge in Authorised Push Payment (APP) fraud — the fastest-growing type of financial crime in the UK. Unlike traditional card fraud, APP fraud relies on sophisticated social engineering: criminals impersonate banks, HMRC, or delivery services to trick customers into voluntarily authorizing payments to fraudulent accounts.

The existing rule-based fraud system was failing badly:

  • False positive rate of 12% — legitimate customers were being blocked, creating support overhead
  • Detection rate of only 34% — most APP fraud slipped through because transfers were "authorized" by the account holder
  • Average response time of 8 seconds — too slow for real-time payment decisioning
  • £2.8M annual losses from fraud reimbursements under the new PSR regulations

The fundamental problem: rule-based systems can only detect patterns they've been explicitly programmed to recognize. APP fraud constantly evolves its tactics.

The Solution

Azura AI implemented a multi-layered predictive defense system that analyzes transactions across three parallel dimensions simultaneously, delivering a fraud confidence score in under 50 milliseconds.

Layer 1: Behavioral Biometrics

We built a real-time user behavior model that learns each customer's "digital fingerprint":

  • Typing patterns: Keystroke dynamics, input speed, and correction frequency
  • Session behavior: Time-of-day patterns, typical transaction amounts, usual recipients
  • Device signals: Screen size, browser configuration, network origin

When a customer's interaction pattern deviates significantly from their baseline — for example, unusually fast typing (suggesting script-driven input) or an atypical time of day — the system raises the risk score.

Layer 2: Graph-Based Network Analysis

We constructed a live transaction graph where:

  • Nodes represent accounts (both sender and receiver)
  • Edges represent money flows with timestamps and amounts

Using graph neural networks (GNNs), the system identifies suspicious network patterns such as:

  • Newly created accounts receiving payments from multiple unrelated senders
  • "Mule chains" where money flows through 3-4 accounts before reaching a final destination
  • Account clusters that share device fingerprints or IP ranges
# Simplified graph analysis pipeline
from torch_geometric.nn import GATConv

class FraudGraphModel(torch.nn.Module):
    def __init__(self):
        super().__init__()
        self.conv1 = GATConv(in_channels=64, out_channels=32, heads=4)
        self.conv2 = GATConv(in_channels=128, out_channels=16, heads=2)
        self.classifier = torch.nn.Linear(32, 1)

    def forward(self, x, edge_index):
        x = F.relu(self.conv1(x, edge_index))
        x = F.relu(self.conv2(x, edge_index))
        return torch.sigmoid(self.classifier(x))

Layer 3: LLM Narrative Analysis

The most innovative component: we fine-tuned an LLM to analyze the payment reference text and chat logs (when available from in-app messaging) for known scam patterns:

  • "Urgent" language suggesting pressure tactics
  • References to fake invoices, tax refunds, or delivery fees
  • Payment references that don't match the stated purpose of the transfer

This layer catches fraud that pure numerical analysis misses — because the social engineering happens in natural language.

Ensemble Decision Engine

All three layers feed into a gradient-boosted ensemble model (XGBoost) that produces a final fraud probability score between 0 and 1. The system operates in three bands:

Score Action Volume
0.0 – 0.3 Auto-approve ~88% of transactions
0.3 – 0.7 Soft challenge (additional verification step) ~10%
0.7 – 1.0 Block + human review ~2%

Financial Impact

After 120 days of production deployment:

Metric Before After Change
Successful fraud attempts (monthly) 340 27 -92%
False positive rate 12% 0.08% -99.3%
Detection latency 8,000ms 47ms -99.4%
Customer friction complaints 2,100/month 180/month -91%
Annual fraud reimbursement costs £2.8M £0.4M -£2.4M saved

"Azura AI's system didn't just catch more fraud — it made the experience better for our legitimate customers. Our support tickets dropped by 60% in the first month." — CTO, [Client]

Key Takeaways

  1. Multi-modal analysis wins: No single technique catches APP fraud. Combining behavioral, graph, and language analysis creates defense in depth.
  2. Speed is non-negotiable: At 47ms response time, fraud scoring happens before the payment confirmation screen renders — invisible to the user.
  3. False positives matter more than false negatives: Blocking legitimate customers has a compounding cost on trust and retention. Reducing FP rate from 12% to 0.08% was the highest-value outcome.

Related Insights