9SEPTEMBER 2025CFO TECH OUTLOOKthe actual positive alerts ­ this may be due to calibrated AI factors such as market sentiment, trade history of the dealing desk, and how surveillance personnel have historically closed alerts (to name a few). Written E-Communications--with the turn of the century, particularly from 2010 onwards, e-communications have become commonplace in everyday life. With hundreds of billions of messages being sent and received each day (estimate just for e-mail alone), it is easy to imagine the oversight difficulty in financial services even with a task force of surveillance reviewers. Although lexicon-based monitoring has been a valuable tool in filtering out the noise of e-communications, there have been deficiencies in how it generates alerts for review. This is where AI gains its main stamp of approval ­ the ability to contextualise messages and compare with lexicon-based triggers has meant firms that integrate AI with e-communications monitoring can review better alerts compared to firms purely using lexicon-based tests. Voice Communications--Similar to Written E-Communications, the use of AI in this monitoring space has gained dramatic momentum since 2024. Like written e-communications, there is an incomprehensible number of hours of voice recordings within financial services based on regulatory obligations. The oversight surrounding this type of monitoring typically revolves around sample monitoring, targeted monitoring, or an automated system to oversee relevant persons. Historically, the theory of transcribing a call using software and being able to review a transcript was seen as a potential solution to voice communications monitoring. This then moved to transcribed recordings being integrated into existing written e-communications systems. However, the main limitation was always the accuracy of the transcriptions in voice-to-text. When dealing desks used short-abbreviated financial terms (standard practice) and/or spoke in different languages or accents, transcription capabilities would often fall short regarding accuracy. With the exponential improvement of AI in the last 18 months and multiple other large language models (LLMs) being utilised, voice transcription services with AI integration are beginning to be able to rapidly learn and contextualise relevant conversations in a repeatable manner accurately. Personal Dealings (Conflicts of Interests)--the challenge with this workstream in the context of personal account dealings, gifts and entertainment, and outside business activities, from my own experience, has always been the accuracy of information being inputted by the employee; unfortunately AI in its current form is unlikely to assist this aspect in the near term. However, assuming this input is accurate, to begin with, the second issue that surveillance functions face is typically the identification of alerts alongside other recorded activities. Historically, where there have been control room-related incidents, the difficulty has been the timely identification of other related workstreams, i.e. trading and/or communication events. With AI integration into these functions, the historical challenge for surveillance personnel to cross-identify other related incidents becomes less problematic and more systematic. The caveat is that the AI integration is on a viable platform that ingests more than one workstream of data. ConclusionIn summary, AI will play a crucial role in surveillance, but the future state of optimisation requires the human element to refine secondary review processes and investigations. While AI continues to develop and evolve, the core principles of surveillance governance and framework will remain rooted in traditional practices. AI will play a crucial role in surveillance, but the future state of optimisation requires the human element to refine secondary review processes and investigations
< Page 8 | Page 10 >