For compliance teams in financial services, the biggest threat to effective surveillance may not be missed misconduct but the sheer volume of meaningless alerts. Systems that produce excessive false positives breed alert fatigue, draining reviewer time and weakening oversight.
Theta Lake is positioning artificial intelligence as the answer, with a new feature designed to strip away the noise and leave staff free to scrutinise genuine risks.
The company’s Alert Confirmation Analysis (ACA) acts as an automated triage layer, running a secondary review of alerts generated by lexicons. Rohit Jain, distinguished engineer at Theta Lake, who has more than two decades of experience across successive generations of machine learning, has outlined how the tool tackles the problem.
The scale of the issue is stark. While 94% of financial services firms use or plan to use AI-based detection, lexicons remain the foundation of compliance monitoring. Yet in live deployments, up to 99% of lexicon hits turn out to be false positives. Because lexicons focus on isolated keywords and short text fragments, they lack the context needed to judge intent.
Specialised machine learning models can reduce false positives dramatically, but since they search for rare signals within sentences or paragraphs, they still raise a substantial number of false flags to avoid missing real cases. This low base rate problem is where ACA steps in.
The tool ingests classification results together with the full record context and classifier descriptions, then labels each hit as a true or false positive, supported by reasoning and a confidence score. It also produces an overall risk score for the entire record, with an explicit rationale and highlighted text excerpts.
Context is decisive. A phrase such as “you’ll get 10 to 20% commission on it” is flagged within a trading conversation but disregarded during sales training. Likewise, “moving to WhatsApp” triggers a high confidence alert when business is shifting away from approved channels, rather than when colleagues are arranging a social event.
ACA sits between sentence-level classifiers and Theta Lake’s Risk Indicator Analytics, which examines patterns across multiple records over time. Record-level risk scores allow teams to prioritise high-risk items, streamline queues and confidently auto-close or sample lower risk records.
Theta Lake stresses that bigger models are not always better. The stochastic nature of large language models can introduce inconsistency, so ACA relies on finely tuned prompt templates and structured outputs to deliver accurate, repeatable decisions.
In extensive enterprise testing, the feature reduced false positives by 95% to 99% while retaining 98% to 100% of true positives, allowing compliance teams to concentrate on the records that truly demand human attention.
Investors
The following investor(s) were tagged in this article.
