Trading surveillance has traditionally relied on rules. A large order, deviation from a benchmark, or known manipulation pattern triggers an alert, and a compliance employee investigates. The problem is that modern markets produce enormous volumes of activity while subtle misconduct may not fit a predefined condition.
Goldman Sachs and Deutsche Bank are exploring agentic AI as an additional monitoring layer. Deutsche Bank is working with Google Cloud on agents that review order and execution data and flag complex anomalies in near real time. Goldman is also examining agents that scan for conduct indicators.
The systems are not described as making disciplinary decisions. Their role is to connect signals and organise cases that deserve human judgment.
Context Can Reduce Noise and Reveal Relationships
Static controls often generate false positives because they evaluate an event against a threshold without understanding the wider sequence. Compliance teams then spend time clearing ordinary behaviour while a more complex pattern remains dispersed across systems.
An agent can compare orders, executions, timing, price movements, market conditions, communications metadata, and a trader’s history. It may decide which data to inspect next and assess whether a combination is unusual even when no single event crosses a rule.
This could improve both precision and coverage. Fewer low-value alerts let investigators concentrate on difficult cases, while relationships outside the existing rulebook become visible.
Evaluation should measure false positives, missed known cases, time to escalation, investigator workload, and the quality of evidence assembled. A more sophisticated alert is not automatically a more accurate one.
Historical replay can test whether the system detects previously confirmed patterns without learning shortcuts that identify only the known outcome.
Human Investigators Retain the Consequential Decision
Market conduct depends on intent, context, and regulation. An unusual trade may be legitimate, while apparently ordinary activity may form part of manipulation across time or venues.
Agents should surface evidence and competing explanations, not present suspicion as guilt. Compliance officers remain responsible for investigation, escalation, and any report or disciplinary action.
The interface should link every claim to an order, execution, message, or market event. Investigators need to understand which signals mattered and how the case differs from normal behaviour.
Bias can enter through historical enforcement patterns or uneven data coverage. Monitoring should compare alert rates and accuracy across desks, asset classes, regions, and strategies without assuming difference proves discrimination or misconduct.
Employees also need a process to correct the system. Dismissed alerts and discovered cases should improve evaluation, but feedback must not train the model simply to reproduce one team’s habits.
Surveillance AI Must Survive Regulatory Scrutiny
Banks are required to maintain effective systems and controls, regardless of whether a regulation specifically mentions agents. Introducing a non-deterministic component does not transfer accountability to a cloud or model provider.
Model versions, data access, prompts, policies, outputs, investigator actions, and final outcomes need durable audit trails. Changes should trigger regression testing so a new model does not quietly alter coverage.
Security is critical because surveillance systems contain sensitive trading and employee data. Agent permissions should be read-only wherever possible, and tool access must not allow a monitoring system to affect the activity it observes.
Regulators may ask how the bank knows the model works, how false alerts are handled, and whether investigators can challenge its reasoning. Answers need performance evidence rather than general claims about AI capability.
Agentic surveillance is attractive because market behaviour is contextual and fast. Rules remain valuable for known conditions, while agents can explore the uncertain space between them.
The strongest design will combine both. Deterministic controls preserve dependable coverage, agents assemble complex patterns, and experienced compliance staff decide what the evidence means. AI does not remove the need for judgment; it can make the right cases reach that judgment sooner.
