AML transaction monitoring systems are widely deployed across financial institutions, yet they often remain difficult to reconstruct, explain, and evaluate in production. This paper argues that a key limitation is not only insufficient detection logic, but insufficient system observability. It defines AML Observability as the capability to reconstruct detection-relevant decisions across the full processing lifecycle and reframes recurring AML control breakdowns as diagnosability gaps rather than isolated detection misses. The paper makes five linked contributions. First, it proposes a five-layer AML observability architecture and the Transformation Spiral as a sequencing model linking data governance, system observability, AI observability, and AI enablement. Second, it interprets public enforcement cases through an OBASHI-informed lens and derives a compact anti-pattern library, including Upstream Blindspot, Silent Transformation Drop, Threshold Suppression, and Broken Feedback Loop. Third, it formalizes a minimal event-centric proof of concept in Python built around an AMLTrace/AMLTraceEvent model, explicit query semantics for whyflagged, whyₙotflagged, and whatchanged, and privacy-aware trace retention. Fourth, it extends the same trace schema toward AI governance by embedding model-specific artifacts such as feature attribution, confidence distributions, and drift indicators into the compliance trace. Fifth, it adds a production-oriented architecture sketch and a larger synthetic comparison between a monitoring-only baseline and an observable system. The larger evaluation processes 1, 000 synthetic transactions, including 300 injected faults distributed across six fault families spanning the five-layer stack. Across the fault cases, the observable system improves mean diagnosis completeness from 0. 25 to 1. 00, failure-layer attribution accuracy from 0. 17 to 1. 00, and reduces explanation steps from 4. 0 to 1. 0 relative to the baseline. A sizing model based on the current PoC event structure further indicates an average raw telemetry footprint of roughly 3. 38 GB per one million transactions, falling to about 2. 05 GB under the selective retention policy. These results remain synthetic and do not constitute production validation, but they provide stronger empirical and engineering support for the claim that observability improves diagnosability in AML systems. The paper situates these results explicitly within the regulatory anchoring provided by BCBS 239 on risk-data aggregation and the European Central Bank's 2024 guide on effective Risk Data Aggregation and Risk Reporting, both of which identify attribute-level data lineage as a top-priority supervisory expectation. The AMLTrace model presented here operationalizes that expectation for the AML pipeline. The paper also positions itself against the rapidly growing literature on AI auditability: where that literature asks whether AI outputs can be audited, this paper argues that the architectural precondition for auditability is system-level observability across the full processing lifecycle. Observability is therefore neither a substitute for governance nor an alternative to AI, but the architectural condition under which governance becomes evidence-based and AI becomes more governable.
Jürgen Schiller García (Sun,) studied this question.