How DPA’s Agentic AI Platform Is Changing Trade Surveillance

It started with a conversation with client’s compliance team was spending three to five hours a day listening to call recordings, and still only getting through a sample of a few dozen at most. Today that same team has close to 4,000 calls reviewed every day across six languages, and their listening begins only after the system has told them what to focus on.

That shift, from sampling a slice of conversations to covering all of them, is the most practical thing AI has done for communication surveillance so far.

It matters because almost every client conversation in financial services carries some compliance weight. A recommendation made on a call. A disclosure that never got said. A product reference that was never suitable for that client. A client code left incomplete. None of these are exotic problems, and all of them live inside audio files that somebody has to actually listen to.

Why sampling stopped working

For years, trade surveillance has run on manual review and statistical sampling. A compliance officer pulls a subset of calls, listens, notes what looks wrong, and documents the findings. When call volumes were modest, that was a reasonable way to work.

Volumes are no longer modest. A mid-sized brokerage or advisory business can generate thousands of client conversations in a single day, spread across desks, products, geographies and languages. The arithmetic gets uncomfortable quickly: the more your business talks to clients, the smaller the share of those conversations anyone is actually reviewing.

Regulators have noticed. FINRA's 2026 Regulatory Oversight Report again flags weaknesses in how firms supervise electronic communications, including thin sampling and gaps in reviewing communications conducted in languages other than English. In other words, thin coverage is no longer just an operational inconvenience. It is a supervisory finding waiting to happen.

What AI surveillance actually does

Strip away the vocabulary and a modern surveillance workflow is doing four things in sequence. Speech-to-text turns the recording into a transcript. Natural language processing picks out the terms, phrases and patterns that matter. Real-time translation makes non-English conversations readable to whoever is reviewing them. Then the conversation gets tested against the firm's own business rules, and anything that looks like a problem gets flagged.

The change in the compliance team's day is subtle but significant. The question stops being "what happened in the calls we happened to sample?" and becomes "which conversations need us today?"

Configurable rules matter too. Generic keyword searching produces noise. Rules built around the actual risks a firm cares about, such as unsuitable recommendations, missing client information, or specific escalation triggers, produce something a compliance officer can work with.

This is also roughly where the industry is heading. FINRA has noted that firms are increasingly exploring AI across communications, investment processes and operational functions, while continuing to press on supervision, reliability, transparency and control over how these tools are used. Deploying AI in a compliance function without being able to explain how it works is its own kind of risk.

Where this goes next

Financial services keeps getting more digital, more data-heavy and more multilingual. Periodic sampling was built for a smaller, slower, more monolingual business, and it is being asked to do a job it was never designed for.

Continuous oversight is the more realistic destination: technology surfacing potential risk across every conversation, and experienced people applying judgement where judgement is actually needed. The point was never to review more calls for the sake of volume. It is to catch risk earlier, resolve it faster, and be able to show your work afterwards.

For firms trying to strengthen governance without adding operational complexity in equal measure, this turns one of the most labour-intensive parts of compliance into a source of usable risk intelligence.

If you are running a surveillance function today, I would be curious to know where you have landed on this. Still sampling, or already moving toward continuous review?

Decimal Point Analytics combines financial domain expertise with AI, analytics and automation to help institutions build compliance workflows designed for scale, multilingual complexity and audit readiness.