Financial institutions throughout Asia are increasingly integrating artificial intelligence into critical banking functions, including lending, payments, customer servicing, Anti-Money Laundering (AML), Know Your Customer (KYC), and transaction monitoring. While initial discussions centered on productivity gains—such as reducing manual effort and accelerating decisions—the industry now faces a more pressing concern: ensuring banks can explain automated decisions. This requires demonstrating not only why a system reached a particular conclusion but also which policies shaped that decision, how those policies were applied across different platforms, and whether their execution remained consistent with regulatory and operational expectations.
The challenge extends beyond traditional model explainability, which focuses solely on interpreting AI-driven predictions. Operational explainability demands a broader perspective: tracing how a specific policy or compliance requirement was applied, what rules were derived from it, which systems enforced those rules, and what final outcomes resulted from the process. A bank might achieve technical explainability within its AI models yet still struggle to account for the full decision-making chain when discrepancies arise between systems.
Explainability Challenges in AI-Driven Banking
Discrepancies often stem from unpropagated policy amendments, varying thresholds across systems, incorrectly mapped data fields, or controls that only partially address transaction workflows. The complexity grows as banking operations rely on fragmented ecosystems, core banking, Customer Relationship Management (CRM), KYC, AML, credit, and workflow systems, each operating with distinct data structures, rule sets, and release schedules. While individual components may pass internal validation, the cumulative process can yield inconsistent results.
For instance, a policy mandating enhanced due diligence when a customer’s risk profile changes materially might be implemented unevenly across multiple platforms. One system could enforce a different risk threshold, another might evaluate a separate customer attribute, and a third may overlook the latest policy revision entirely. Superficially, the procedure may appear compliant, but its original intent could be undermined by gaps between interconnected systems.
Establishing bidirectional traceability is essential to resolve these issues. Policy, compliance, and technology teams require a unified view linking policy language to its interpretation, the resulting rules, and their system-level implementation. Conversely, operations, audit, and control teams must reverse-engineer transaction outcomes, mapping them back to the applied rules and, ultimately, to the governing policies that authorized them. This dual visibility ensures accountability at every stage of the decision-making process.
Read Also: Binance resumes withdrawals after $6bn hack
Regulatory Scrutiny and Explainability Demands
Asia’s financial regulators are accelerating demands for greater transparency in AI-driven banking. The Hong Kong Monetary Authority (HKMA) has prioritized explainability, governance, and validation frameworks for AI applications, expanding its GenAI Sandbox in March 2026 to facilitate controlled testing environments. Similarly, Singapore’s 2026 Model AI Governance Framework for Agentic AI emphasizes human accountability, bounded autonomy, and continuous technical safeguards throughout the AI lifecycle.
The risks of failing to maintain operational traceability are already manifest. In July 2025, the HKMA imposed penalties totaling HK$8.5 million, HK$4 million, and HK$3.7 million against an Indian-origin bank’s Hong Kong branch and two Chinese banking entities. The violations stemmed from omissions in uploading specific transaction types into a shared monitoring system post-migration, subsequent system modifications, and the introduction of new services that bypassed existing controls.
Solving Operational Explainability Gaps
Addressing operational explainability gaps does not necessarily require large transformation programs. A non-invasive AI-powered assurance layer, operating alongside core systems, can provide visibility without altering existing infrastructure. Banks could replay historical transactions against current or future policy-derived controls, identify exceptions, and examine whether controls are missing, outdated, or incorrectly implemented. This proactive approach allows policy adherence to be tested before issues surface through audits, complaints, or regulatory examinations.
The cost of inaction can be disproportionate. Small interpretation or implementation differences, repeated across numerous transactions, can lead to regulatory penalties, remediation costs, and operational losses. The deeper issue is the gap between policy intent and execution. As banks scale AI into consequential decisions, the ability to explain, test, and demonstrate alignment with policy and regulation becomes critical.
