Compliance teams in financial services are struggling with a basic problem: their surveillance tools produce far more noise than signal.
Lexicon-based monitoring is still the foundation of most supervision programmes, yet in real-world deployments as many as 99% of its hits turn out to be false positives. The result is alert fatigue, which slows reviews and weakens the surveillance it is meant to support.
RegTech firm Theta Lake says its Alert Confirmation Analysis (ACA) feature tackles this directly. ACA acts as an automated triage layer that runs a second analysis on alerts already raised by lexicons.
Rohit Jain, distinguished engineer at Theta Lake, has more than 20 years of experience across several generations of machine learning, and he has set out how the tool cuts noise for compliance reviewers.
The pressure to adopt AI is clear. According to Theta Lake research, 94% of financial services firms already use, or plan to use, AI-based detection. Lexicons scan isolated keywords and short text fragments and miss the wider context. Specialised machine learning models improve on this a great deal. However, because they search for rare signals and are tuned to avoid missing genuine risks, they can still generate a large number of false flags.
ACA addresses this by reviewing each detection alongside the full record and the relevant classifier descriptions. It labels every hit as a true or false positive, with reasoning and a confidence score. It also gives the whole record an overall risk score, with a rationale and highlighted supporting text.
Context changes the outcome. A phrase such as “you’ll get 10 to 20% commission on it” would be flagged within a trading conversation but dismissed during sales training. Likewise, “moving to WhatsApp” triggers a high-confidence alert when business is moving off approved channels, but not when colleagues are organising a social event.
Theta Lake positions ACA between sentence-level classifiers and its Risk Indicator Analytics, which tracks patterns across many records over time. With record-level risk scores, teams can prioritise high-risk items, streamline queues and confidently auto-close or sample lower-risk records.
The company also takes a measured view of large language models. It argues that their stochastic nature can introduce inconsistency, and that bigger models are not always better for basic classification. ACA therefore uses a structured pipeline. Each record is enriched with detection-specific prompts and then processed through finely tuned templates that produce consistent, structured outputs.
In enterprise testing, Theta Lake reports that ACA cut false positives by 95% to 99% while keeping 98% to 100% of true positives. That frees compliance staff to focus on the records that genuinely need human judgement.
Read the full Theta Lake post here.
Copyright © 2026 RegTech Analyst
Copyright © 2026 RegTech Analyst
