Fifty Thousand Wires Nobody Was Watching

On 3 August FinCEN assessed a $125 million civil money penalty against UBS Financial Services Inc., the largest Bank Secrecy Act penalty ever imposed on a broker-dealer. The conduct ran from January 2019 to June 2023. Parallel matters brought by the SEC, the CFTC and FINRA were closed at the same time.

The headline number will do the rounds and the finding underneath it will not, which is a shame, because the finding is the useful part. FinCEN found that the firm failed to monitor more than 50,000 foreign currency wires with a combined value exceeding $10 billion.

Those wires were not scored incorrectly, and they were not alerted on and dismissed. Fifty thousand payments moved through a regulated institution and the system responsible for looking at them did not look at them, because they were never presented to it.

It is a different category of failure from the one the industry spends its time discussing, and it belongs to nobody in particular.

Detection quality and coverage are not the same problem

Almost every conversation about automation in financial crime is a conversation about detection quality. False positive rates. Threshold calibration. Whether a machine learning layer over a rules estate suppresses the wrong alerts. Whether a model is explainable enough to defend to a regulator, and who signs the validation report. These are real questions and I have written about several of them.

They all share a premise. They assume the transaction reached the system.

Coverage is the prior question, and it is plumbing. Which source systems feed the monitoring platform. Which product lines, entity types, currencies, correspondent relationships and booking locations are in the feed and which are not. Whether a migration two years ago quietly dropped a table. Whether a new payment corridor was onboarded commercially in March and connected to monitoring in November, or never. Nobody writes conference papers about this because it is unglamorous, and it is precisely where the ten billion dollars went.

The uncomfortable property of a coverage gap is that it is silent in the metric everyone watches.

Consider what a monitoring function actually looks at week to week. Alert volumes. Alerts per analyst. Conversion rates from alert to case to SAR. Model performance against a tuning sample. Now introduce a gap in which an entire class of payments never enters the pipeline. Alert volumes do not spike, because there is nothing extra to alert on. They do not fall, because those payments were never generating alerts in the first place. Conversion rates are unaffected. The tuning sample is drawn from the data that arrived, so the model tests clean against exactly the population it can see.

Every indicator on the dashboard says the estate is healthy. The estate is missing fifty thousand wires.

A detection failure produces evidence of itself: alerts that should have fired and did not, cases closed that should have escalated, a SAR filed late enough to notice. A coverage failure produces silence, and silence reads identically to safety.

The objection, which is that this is an engineering problem

The strongest response to all of this is that coverage is not a compliance failure at all. It is a data engineering failure, it belongs to the technology function, and dressing it up as a financial crime issue confuses ownership. Monitoring teams cannot be expected to audit ETL pipelines. That is what data lineage tooling and change management are for.

There is something in that, and it is exactly why the problem recurs.

Ownership is the mechanism here. Model validation validates the model, and its scope statement will say so; the validator is asked whether the model performs against the data provided, not whether the data provided is the whole population. Data lineage sits with engineering, where the acceptance criterion is that pipelines run without error rather than that they carry every in-scope payment. Change management approves a release. The MLRO reads a monitoring pack. Internal audit samples the control environment, and a sample drawn from the system’s own records cannot reveal what never entered it.

Each function does its job to its own boundary, and the gap lives in the space between the boundaries. That is not a story about incompetence. It is a story about a control that has no owner because it does not look like a control. It looks like infrastructure.

The regulatory record keeps confirming it. FinCEN’s action is described as recidivist for a reason: UBS Financial Services paid a $14.5 million penalty in December 2018 over failings that included inadequate monitoring of foreign currency wire transfers, and the conduct FinCEN penalised this month began the following month and ran for four and a half years. A firm that had been told, in writing, with a number attached, that its foreign currency wire monitoring was inadequate, did not close it.

Remediation is where this kind of finding goes to die. A model tuning fix is a project with a testable outcome. Connecting an orphaned source system to a monitoring platform crosses three functions, two budgets and a vendor contract, and its success criterion is that nothing visible changes.

What the AI conversation misses about all of this

There is a version of this argument that treats it as a problem the current wave of technology will solve, on the basis that better systems ingest more. I would not rely on it, and I think the direction of travel runs the other way.

Sophistication in the detection layer raises the cost of a coverage gap rather than lowering it. A rules engine with a documented scenario library at least tells you what it is looking for, and a scenario that names foreign currency wires is a question somebody can ask of the feed. A behavioural model trained on historical activity infers normality from what it has seen. If a payment class is absent from the training data, the model does not treat it as missing. It treats the world as a place where that behaviour does not occur, and it will keep doing so with high confidence and good validation statistics.

I have written elsewhere about how the EU AI Act’s Annex III leaves transaction monitoring outside the high-risk regime while capturing credit scoring, and about the FCA’s preference for supervising outcomes rather than technologies. Both arguments were about what sits outside the frame of regulation. The UBS finding puts the same point one level in. Even inside a firm’s own model risk management, where the documentation exists and somebody signs it, the completeness of the input population is assumed rather than tested.

Article 10 of the AI Act, on data governance, gets closer to this than anything in the UK regime, requiring that training and testing data sets be relevant, sufficiently representative and complete. It is the right instinct. It also does not apply, because the systems concerned are not in Annex III.

What to ask on Monday

If you work anywhere near a monitoring estate, the question is not how the model is performing. It is a reconciliation question, and it is answerable.

Take the payment volumes from the source of record for a given month, by product, currency, booking entity and channel. Take the volumes ingested by the monitoring platform for the same month, cut the same way. Compare the two. Any line where the second number is materially lower than the first, or absent entirely, is a coverage gap, and the size of the gap is the size of the exposure.

It is a tedious piece of work, it requires cooperation from people who do not report to compliance, and it is the single control most likely to be missing. Ask when it was last done, who owns it, and whether it runs on a schedule or ran once during an implementation four years ago. If the answer is that the platform receives everything, ask how that is evidenced.

Fifty thousand wires is not a subtle number. Somebody could have counted.