At one wholesale bank, a coding defect prevented data from reaching a financial-crime control. The control’s continued operation offered no assurance that it was monitoring the complete population. The bank’s data monitoring and completeness checks also failed to expose the omission. Unrelated activity eventually brought the gap to light.
This example sits at the centre of the Financial Conduct Authority’s (FCA’s) latest wholesale banking supervisory findings. The missing data generated no alert because it never entered the system’s control flow and routine oversight failed to identify cases outside the control’s field of vision.
The FCA found comparable blind spots in communications surveillance, algorithmic trading, data-loss prevention and artificial intelligence. Communications surveillance received incomplete data, testing omitted a critical trading scenario, human review missed inaccurate AI-generated code, and an employee adapted their behaviour to bypass data-loss controls.Each control could produce normal-looking output while not detecting data, scenarios or behaviour outside its scope. Compliance needs evidence of the control’s perimeter: what should have entered, what arrived and whether system changes altered coverage.
An Established Pattern
Less than three weeks earlier, the FCA had identified the same assurance problem in Market Watch 86, following a review of seven UK retail brokers. Every firm had automated surveillance and documented procedures. The FCA still found material weaknesses in control design, implementation and operation.
Some surveillance staff could not explain their systems’ alert logic or calibration. Others relied on standard settings supplied by vendors. Flat price-movement thresholds ignored differences in volatility. Minimum trade values excluded a concerning proportion of client activity from monitoring.
Testing failed to expose the missing coverage. Several firms compared alert volumes with trading volumes or replayed historical data. Those checks did not identify cases where alerts failed to trigger. Senior managers lacked the information and technical understanding needed to challenge the results.
Market Watch 86 focused on retail brokers, but the FCA said its findings could apply more widely. The latest wholesale-bank publication reinforces that wider relevance.
Market Watch 86 and the wholesale findings expose different routes to incomplete coverage. In the first, calibration excluded activity that reached the system. In the second, data or scenarios never reached the control or its tests. Both produced normal-looking output from a partial view.
Testing for Gaps
Quality assurance over completed alerts cannot identify records that never entered surveillance. Firms must reconcile the expected monitoring population against the records received and test scenarios that configuration may exclude. Market Watch 86 points to quality-assurance findings, Suspicious Transaction and Order Reports and calibration reviews as useful evidence. The wholesale findings add a prior test: did the control receive the full population?Daniel Ihrig, eDiscovery and Surveillance Architect at Macquarie Group, has argued that “completeness and reconciliation controls are no longer considered a day two activity.” They belong in implementation, change and production monitoring.
The FCA also identified a weakness in how one bank tested a change to a trading algorithm. Simulation testing did not cover near-simultaneous fills in separate locations. It therefore missed an error in the order-management component responsible for communicating fill notifications. The algorithm received an incorrect executed quantity, overfilled the order and accumulated a significant unintended intraday position. Firms must test changes across the complete order lifecycle, including interactions with venues and downstream systems.
AI Assurance
AI controls showed the same gap between approval and production. Inaccurate AI-generated code passed human review and caused a loss of service. Internal audits found incomplete AI inventories, undocumented production exceptions and conditional approvals recorded as full approvals. They also found insufficient kill-switch training and testing, and no documented list of authorised operators.
Human review needs its own performance test. Firms should define what reviewers must challenge, give them access to the underlying evidence and test whether they detect known faults. Controlled cases can include inaccurate recommendations or defective code. Repeat the test after material changes to the model, workflow or reviewer guidance.
Banks need evidence of complete coverage. Change testing must cover downstream interactions and omitted scenarios. AI inventories should reconcile approvals with production deployments and named intervention rights.
Management information must expose missing coverage before an incident or supervisory review does.
Subscribe to our newsletter
