The Institutional Challenge: The Cost of Being Wrong Is Paid Twice
Fraud economics have outgrown the rules-based monitoring stack.
Alert overload is now a conduct problem. Legacy threshold engines flood investigators with low-value alerts while genuinely high-risk cases wait. A false positive no longer costs only an analyst hour, it costs a blocked legitimate customer, a complaint, and eventually a conduct conversation.
Typologies mutate faster than rules can be re-tuned. Authorised push-payment scams, mule networks and synthetic-account rings reconfigure weekly. Static thresholds detect yesterday's fraud at tomorrow's volumes.
Black-box scores fail the governance test. HKMA's 2019 High-level Principles on AI and its 2024 GenAI guidance require governed, transparent, explainable AI-assisted decisions. A scoring model that catches the right transactions but cannot say why will not survive the first regulatory inquiry.
Channel silos blind the bank. FPS, card-not-present, cross-border wires and e-wallet top-ups are monitored channel by channel, while the defining feature of modern fraud is the cross-channel, cross-account pattern.
Reference Architecture: A Transparent Decision Core with an Agentic Perimeter
The design resolves the false choice between rules and AI: probabilistic intelligence proposes, deterministic rules dispose.
Ingestion & Orchestration: Sub-Second, Across Every Channel
Every payment event across 11+ product channels instantiates an executable process on the orchestration engine under a unique business key, idempotent by design, so retries never double-debit or double-alert. The ingest, enrich, analyse, decide, act/escalate lifecycle executes in sub-second time, with SLA tracking on every escalated case and an audit record written by the engine itself. Real-time means the fraudulent transaction is held before settlement, not investigated after it.
Walkthrough, an authorised push-payment attempt at 21:14 on a Sunday:
Ingestion. An FPS transfer of HKD 180,000 to a newly added payee triggers the payment-vetting process.
Enrichment and scoring. The Harmony Engine scores the transaction against behavioural baselines: first-time payee, session on a newly registered device, amount just under a reporting threshold, transfer velocity inconsistent with the customer's history.
Network analytics. Graph traversal over accounts and counterparties shows the payee account received funds from six unrelated customers in 72 hours and forwarded 94% onward within minutes, a mule topology invisible to transaction-level rules. Two prior victims are surfaced for proactive safeguarding.
Deterministic decision. An Executable DMN decision table, high-value transfer × newly seen beneficiary × velocity breach × adverse network signal, returns hold and escalate. Because the rule is FEEL-readable, the duty manager sees exactly which conditions fired.
Action and evidence. The transfer is held pre-settlement; the customer is contacted through an out-of-band channel; the case opens with SLA countdown; InvestLM drafts the analyst summary with the feature drivers and source citations attached. Every artefact is audit-ready by construction.
Decision Engine (DMN): One Rule Set, Eleven Channels
Approximately 100 rules and scenarios, expressed as readable FEEL decision tables, govern all 11+ channels, coverage without per-channel rebuilds. Business analysts read, simulate and version every rule in the visual process modeler and back-test changes against historical transaction populations before promotion. When a new scam typology emerges, a rule update ships in days, governed by a four-eyes gate, the opposite of a code release.
Human-in-the-Loop & Auditability
High-risk flags route to humans under SLA tracking; confidence-scored AI context supports, never substitutes, the investigator's decision. Guardrails enforce continuous bias monitoring and PDPO-by-design data handling. The decisive property is explainability: every flag traces to a named DMN rule version and a grounded rationale, defensible to HKMA examiners and internal audit alike.
Empirical Baselines & Regulatory Alignment
Capability | Published baseline | Source |
|---|---|---|
Real-time, network-aware fraud decisioning | ~HKD 35M losses prevented in year one; ~200 customers safeguarded; 11+ channels on one rule set | Tier-1 HK bank reference implementation (AI Services series) |
False-positive reduction | 60-90% at peer institutions; HSBC's published AI/graph deployment cut alerts 60% while detecting 2-4x more suspicious activity | Peer benchmarks; HSBC × Google Cloud published results |
Decision latency | Batch review (minutes) to sub-second real-time scoring | Reference implementation |
Regulatory alignment by design:
HKMA AML/CFT Regtech programme, the architecture directly operationalises the four priorities of AML/CFT Regtech: Case Studies and Insights Volume 2 (September 2023): real-time fraud monitoring, ML alert handling, NLP media monitoring and Single View of the Customer.
HKMA circular, Use of Artificial Intelligence for Monitoring of Suspicious Activities (September 2024), AI-powered monitoring with governed models, feasibility evidence and implementation planning.
HKPF anti-fraud initiatives and FMLIT-style intelligence sharing, network analytics surface additional victims for proactive safeguarding; explainable evidence chains are shareable with intelligence partners.
HKMA 2019 AI Principles / 2024 GenAI guidance; PDPO (Cap. 486), governance, transparency, human oversight and data minimisation are structural, not documented after the fact.
The Global Scaling Implication
Fraud decisioning is the entry point for the enterprise decisioning fabric. The same DMN governance, process orchestration and agentic perimeter extend to sanctions screening, KYC knowledge graphs, post-trade exceptions and regulatory reporting as re-uses of one pattern. Each extension lowers the marginal cost of the next, and the bank that industrialises explainable, real-time decisioning first converts a compliance cost centre into a compounding operational moat across every risk domain it touches.
FAQ
Q: How does the system hold a fraudulent FPS transfer pre-settlement without delaying legitimate customers?
A: Scoring executes in-line and in sub-second time inside the payment flow. The vast majority of transactions clear deterministically with no friction; only combinations of conditions defined in the DMN tables, for example, high-value transfer × newly seen beneficiary × velocity breach × adverse network signal, route to a hold-and-escalate branch. Because DMN rules are business-readable and back-tested against historical populations before promotion, thresholds are tuned to catch typologies, not customers.
Q: How do we evidence to HKMA why a specific transaction was held or released?
A: Every decision record contains the exact DMN rule version and fired conditions, the Harmony Engine's feature contributions and confidence score, the graph evidence for any network finding, the InvestLM-generated rationale with citations, and the human disposition where a case was escalated. The record is written by the orchestration engine as the process executes, so producing it for an examiner is a query against the audit log, not a reconstruction exercise.
