New White Paper: AI and the Operational Reality of e-Comms Surveillance

The rapid expansion of electronic communications channels has fundamentally reshaped how financial firms interact with clients, counterparties and markets. Email and Bloomberg messaging have been joined by collaboration tools, instant messaging platforms, voice calls and video meetings, all of which now sit squarely within the scope of regulatory scrutiny. For compliance teams, this proliferation has created a familiar but unresolved problem: surveillance systems are processing more data than ever yet struggling to surface genuinely meaningful risk.

At scale, the numbers are striking. Respondents reported scanning and archiving millions of electronic messages every day. Tier 1 institutions cited daily volumes “between 10 and 20 million” messages, while smaller firms still referenced figures in the hundreds of thousands or low millions. A Tier 1 global financial institution reported scanning around 5.5 million messages per day yet generating alerts on just 0.1% of that total.

Part of the issue lies in the mechanics of modern communications platforms. Instant messaging tools dramatically inflate message counts because, as one respondent put it, “every time you hit the enter button, that’s a message.” Short conversational exchanges can generate hundreds of records, swelling ingestion volumes without necessarily increasing risk.

The result is a surveillance process where scale has outpaced precision. While less than 1% of messages typically generate alerts, those alerts rarely represent actual compliance violations.

False Positives

Across firm types and geographies, false positives emerged as the single most significant weakness in electronic communications monitoring. Interviewees consistently reported false positive rates exceeding 90%, with many placing the figure in the high 90s and some as high as 99%.

The cost is not only inefficiency. High volumes of low-quality alerts divert attention from genuinely risky activity and contribute to burnout and turnover within surveillance teams. Several respondents acknowledged that their highest-risk areas were not receiving the focus they deserved because resources were consumed clearing noise.

Legacy Technology

Despite widespread dissatisfaction with alert quality, many firms continue to rely on surveillance architectures that are five to ten years old or more. These systems are frequently described as “old,” “technically inflexible” and “too manual,” with heavy reliance on static lexicon-based detection.

The limitations are structural. Lexicon-driven systems tend to trigger alerts on isolated words or phrases, without understanding the broader conversation or its relationship to trading activity. As a result, alerts are often generated on standalone messages even when surrounding context clearly indicates benign intent, such as informal discussion unrelated to business.

Several respondents noted that their existing platforms were not sophisticated enough to support modern AI techniques or to integrate communications data with trade surveillance. In some cases, this inflexibility has driven firms to build outside the system – or to replace it entirely – rather than attempt incremental upgrades.

Expanding Coverage

The report shows that most firms have achieved broad coverage across traditional and newer electronic channels. Email, Bloomberg messaging, SMS, Slack, Microsoft Teams and Symphony are widely captured, alongside mobile and turret voice communications transcribed into text.

However, significant gaps remain. Full video content from platforms such as Zoom or Microsoft Teams is generally not monitored, with firms citing the physical impossibility of reviewing the sheer volume of meetings, alongside technical and ethical concerns. Instead, most capture only the audio component and rely on transcription to feed voice data into existing e-comms engines.

Encrypted messaging apps such as Signal, Telegram and personal WhatsApp remain difficult to monitor reliably. In response, firms have turned to governance rather than technology, enforcing strict approved-channels policies. As one respondent stated, “If we can’t monitor it or we can’t capture it or we can’t record it, they’re not permitted to get there. Period.”

Where AI helps – and Doesn’t

Against this backdrop, artificial intelligence is widely viewed as essential, but not a panacea. The primary driver for AI adoption is clear: reducing false positives. Firms reported using or piloting AI for risk-based prioritisation, contextual understanding, anomaly detection and, critically, voice transcription.

The report includes examples of tangible impact. One firm reduced alert volumes from approximately 900,000 to 16,000 through large language model tuning, while another achieved transcription accuracy of more than 92% using AI to convert voice into text for surveillance purposes.

Yet enthusiasm is tempered by scepticism. Several respondents questioned the practical value of vendor-supplied AI, suggesting that marketing claims often outpace real-world results. Explainability remains a major concern, particularly where vendors are reluctant to disclose how models operate. As one respondent put it, “the jury’s still out” on whether current AI tools genuinely solve the surveillance problem.

Beyond single vendor solutions, the report points to a broader change in how firms are approaching electronic communications surveillance. Rather than viewing AI or new tooling as a standalone fix, respondents describe a reassessment of how surveillance is designed, evolved and governed over time. The emphasis is increasingly on flexibility, control and incremental improvement, particularly in response to the limitations of legacy platforms that cannot easily absorb new techniques as they emerge.

A Hybrid Approach

This shift in approach is reflected in growing interest in hybrid or internally controlled solutions and in more pragmatic integration of surveillance capabilities. Firms are not seeking to replace entire stacks overnight, but to reduce dependency on rigid, long-term vendor deployments and to introduce modular components – such as improved transcription, contextual filtering or policy management – that can be adapted as requirements change. As one respondent noted, being locked into inflexible systems makes it difficult to benefit from the “regular advances in AI,” reinforcing the need for architectures that can evolve without constant wholesale replacement.

When asked what they would change with a “magic wand,” respondents returned to two themes: eliminating noise and improving context. The most radical wish was to “delete all lexicons,” reflecting deep frustration with keyword-based detection. Others emphasised the need to unify structured and unstructured data so analysts can see intent and behaviour in one place, rather than chasing isolated alerts.

Efficiency also featured prominently. Several respondents expressed a desire for systems so effective that only a small core team would be required to manage surveillance flows, freeing experienced staff to focus on genuinely complex risk.

An Industry Benchmark

Taken together, the findings paint a consistent picture across buy-side and sell-side firms alike. Surveillance teams are processing unprecedented volumes of communications, yet struggling with outdated technology, excessive noise and rising regulatory expectations. AI offers real promise, particularly for false positive reduction and voice integration, but only when paired with flexible architectures, strong governance and realistic expectations.

For compliance practitioners seeking to benchmark their own programmes, the full report provides detailed charts, firm-type comparisons and verbatim insights from peers facing the same challenges.

Downloading the report HERE for a data-driven view of where electronic communications monitoring stands today – and where it may realistically go next.