What is communications surveillance?

Communications surveillance is the recording and review of employees' business communications, such as email, chat, voice calls and messaging apps, to detect misconduct like insider dealing, collusion, mis-selling or harassment. Firms in trading and advice businesses run it because regulators expect them to. Once AI agents send messages themselves, their communications fall in scope too.

Why does communications surveillance matter once agents act?

Two reasons. First, agents can review far more of the captured material than a team of reviewers, reading context rather than matching keywords. That changes what surveillance can catch. Second, agents now speak to clients and counterparties on the firm's behalf. A client-facing agent's messages are business communications and need to be captured, retained and reviewable in exactly the way a person's would be.

The second point is often missed. If a person may not promise a client a price or hint at a forthcoming deal, an agent may not either, and you need the record to prove it did not.

How does it work in practice?

The traditional approach uses lexicons: lists of risky words and phrases that raise an alert when matched. Lexicons are noisy. "Let's take this offline" can mean many things, and a reviewer may spend most of the day clearing alerts that mean nothing.

An agent reads the whole conversation instead of the flagged phrase. It can see that "offline" was about a broken video link, or that a series of messages between a salesperson and a trader lines up with an unusual order. Its output is a summary with quotes, the participants, the timeline and why it looks concerning, for a surveillance analyst to judge. Our guide on reviewing communications surveillance alerts with an AI agent covers the steps.

An example from a broker's desk

Take a mid-sized broker. A lexicon alert fires on a chat containing "keep it quiet". The agent reads the surrounding thread: two colleagues planning a leaving present. It drafts a note recommending closure, with the quoted context, for an analyst to confirm.

Later the same day it reads a call transcript in which a salesperson tells a client that "something big is coming on that name next week", a few hours before the firm takes on a confidential mandate in that stock. No lexicon word fired. The agent drafts an escalation with the transcript extract and the timing, and an analyst opens an investigation.

How is communications surveillance different from trade surveillance?

Trade surveillance looks at orders and transactions for patterns that suggest market abuse, such as unusual trading ahead of announcements or orders placed and cancelled to move a price. It works on structured data.

Communications surveillance looks at what people said. It works on unstructured text and audio, and much of it concerns intent, which is harder to judge. The two are strongest together: a suspicious trade becomes far more serious when a message shows someone knew in advance.

What limits apply to monitoring staff?

Monitoring employees involves personal data and privacy expectations. Under GDPR and UK GDPR, firms need a lawful basis for the processing and, for systematic monitoring, should carry out a data protection impact assessment. Staff should know their business communications are recorded and reviewed. Whether an AI agent should monitor staff communications for misconduct beyond regulatory surveillance is a narrower question that needs its own justification.

Where Pryme fits

In Pryme, a surveillance agent works at depth 1 or 2: it reads captured communications and drafts findings for a named analyst, and the work sits with the second line, never the last word. Client-facing agents' own messages are filed in the append-only record, so they can be surveilled like anyone else's.