What senior compliance leaders and the FCA really think about AI in communications surveillance

Ask a room of senior compliance leaders what’s stopping them from putting AI to work in surveillance, and you would expect to hear about the technology. But what is really holding firms back is not capability, it’s confidence: the worry that they cannot yet prove to a supervisor that they remain in control of what they have deployed.

That was the clear message from an invite-only breakfast VoxSmart hosted at The Connaught in June, where around forty compliance and surveillance leaders sat down with Dominic Holland, the FCA's Director of Market Oversight, to talk candidly about the future of communications surveillance in financial services.

The morning was held under Chatham House Rule, so the report that came out of it is anonymised throughout. But the picture it paints is clear enough. What follows is a taster of what was said in the room. The full report is available to download on the link below.

Rebalancing risk: a more predictable regulator

The morning opened with a keynote and fireside chat from Dominic Holland, FCA's Director of Market Oversight, which set the tone. The regulator's evolving approach is not about lowering standards. It is about being more predictable, purposeful and proportionate, and about treating growth and market integrity as partners rather than competing opposites. Trusted markets are what attract investment and participation in the first place.

What stood out was how far the FCA is holding itself to the same standard it sets for firms. It ingests billions of transaction reports a year and hundreds of millions of order book messages a day, and it is running its own portfolio of proof-of-concept AI projects to keep pace, always with a human in the loop. In one example shared in the room, AI-assisted triage has cut the review time for suspicious transaction reports from around three days to around three hours. The message to firms was direct: engage with us, use the AI Live Testing cohorts and TechSprints, and treat the regulator as a sounding board rather than something to work around.

Confidence, not capability, is the real brake on AI adoption

One of the most interesting insights of the morning came from the table discussions. What is slowing AI adoption in surveillance is rarely a lack of appetite, and increasingly it is not a lack of capability either. It is a lack of confidence.

Several leaders described a genuine reluctance to deploy while regulatory expectations are still settling, and a fear of breaching rules that are still taking shape. The result is a paradox: caution about the regulator is holding back precisely the responsible innovation the regulator says it wants. Add in two familiar constraints - data quality, and inherited systems that teams did not choose and do not fully trust - and you have a fairly accurate view of why good technology sits unused.

Black boxes, blind spots and noise

Three operational themes ran through every table.
‍The black box: firms running AI-driven surveillance that neither they nor their vendor can fully explain, at exactly the moment supervisors are hardening their expectations on explainability.
The blind spots: a communications perimeter that keeps widening, where the riskiest conversations are the ones deliberately taken off-channel.
And the noise: false positives as the daily drain, and a growing impatience with any approach that answers every new risk by bolting on more detection logic.

The full report goes deeper into each of these, alongside what senior leaders believe firms should be watching over the next twelve to eighteen months.

Download the full report

The complete briefing sets out the five key takeaways from the London roundtable, the FCA keynote themes in more depth, and the questions likely to define the surveillance agenda into 2027. It draws a direct line to what we heard at our Madrid roundtable earlier in the year, where compliance leaders across a different market and a different regulator landed on strikingly similar conclusions.

‍

‍