The phrase “real-time surveillance” has long functioned more as a marketing label for market abuse detection than an accurate description of how firms operate. While modern technology can generate an alert the second a trade executes, the capacity to immediately review and act on that alert is far rarer.
At A-Team Group’s recent AI in Capital Markets Summit London, Jason Shiu, Surveillance Compliance Manager at Vanguard, joined Justin Nathan, Global Head of Surveillance at DimeTrades, to address why true real-time review remains elusive, how artificial intelligence is altering trade surveillance, and what regulators ultimately expect.The Capacity Gap and Rules-Based Realities
Whether a firm can achieve real-time monitoring depends on its organisational shape and resourcing rather than the capability of its detection engine. Addressing the practicalities for larger market participants, Shiu noted that “this idea, this talk of real-time surveillance – in terms of practicalities as to whether it is a reality, I don’t think we’re there yet.”
A large asset manager may not be staffed for continuous review, whilst a single compliance officer at a smaller firm is already forced to prioritise anti-money laundering work (for example) over instant trade alerts. Nathan highlighted the core contradiction in generating alerts instantaneously if no one is available to review them. “Many firms would be glad to reach T+10, let alone T+0,” he suggested.
Whilst continuous monitoring is a commercial necessity for exchanges and prediction market venues seeking to protect market integrity, banks and hedge funds have rarely interrogated whether instant alert generation is truly necessary. This dynamic has highlighted the limitations of traditional, rule-based surveillance systems. Whilst parameters and fixed thresholds remain popular because they are legible and traceable for audit purposes, they struggle with flexibility. For instance, anticipating that US tariff announcements would trigger a wave of fixed-income alerts, Shiu’s team had prepared for an April overflow that failed to materialise on schedule, only for a surge to arrive the following month driven by non-US events, noted Shiu.Yet, rule-based systems still serve a vital role for binary behaviours. When monitoring actions like front running, where the underlying compliance principle remains constant, AI models work best when overlaid on top of existing rules to better discern true positives based on timing, rather than replacing the rules entirely.
Efficiency Gains and the Shift Beyond Trade Data
For larger firms at the time of speaking, the primary value of AI currently sits in enterprise efficiency rather than instant false-positive reduction. Investigations have seen dramatic improvements, with models ingesting and summarising vast amounts of market data in moments – work that previously consumed hours of analyst time. Management information reporting has similarly accelerated as senior leaders seek to cut the human capital costs of manual report preparation.
Interestingly, e-communications surveillance, which historically lagged behind trade surveillance by relying on rigid keyword lexicons that generated endless false hits, embraced large language models first and rapidly outpaced trade surveillance tools. However, broad adoption of generative AI across both domains remains constrained by information security and privacy risks surrounding proprietary data leaving the firm. Beyond alert triaging, Nathan argued that “when it comes to the application of AI for surveillance, we shouldn’t just think about the surveillance, we should think about the surveillance operating model as a whole” to unlock wider efficiency gains.
Pragmatic Regulatory Standards and Explainability
As these tools proliferate, the debate surrounding explainability versus effectiveness has intensified. Taking a contrarian stance on the debate, Nathan stated that he is “leaning in and doubling down onto effectiveness versus explainability,” resting his case on the gap between what surveillance is asked to prove and the standard applied anywhere else. In his view, expecting compliance teams to explain the mathematical inner workings of complex models holds them to a stricter standard than the frontier labs building the technology (a limit OpenAI CEO Sam Altman has acknowledged). What matters is whether a tool works and if its practical workflow can be demonstrated.
Shiu agreed that senior managers should be able to explain the input-to-output process to a regulator in twenty to thirty minutes without needing to act as a data scientist. To satisfy regulatory inquiries, Shiu stressed that “having that documentation, be it self-created or actually from the vendor, to explain how whatever tool you’re using is operating” remains key as an “AI artefact library.” Ultimately, regulatory standards across jurisdictions are expected to settle on outcomes and principles as familiarity with AI grows.
The Future Operating Model: AI Agents in Triage
Looking ahead, the surveillance operating model is poised for a structural shift. First-line triage reviewers will increasingly migrate towards judgement-heavy secondary and tertiary reviews as AI agents take over mechanical first-pass filtering. Nathan compared this evolution to supermarket self-checkouts, where human cashiers do not disappear but instead transition to a supervisory role managing exceptions. However, both experts pushed back on allowing autonomous agents to adjust fraud detection thresholds on their own.
Whilst real-time threshold tuning is technically feasible, leaving AI agents to recalibrate surveillance parameters unsupervised carries significant operational risks that most firms are not yet ready to take.
While neither expert expects real-time review to become the norm overnight, the window for handling alerts is rapidly narrowing. Triage teams are becoming leaner, documentation libraries are being established, and firms are placing their bets on operational effectiveness while regulatory frameworks gradually take shape.
Subscribe to our newsletter
