Can an AI agent do market abuse surveillance?

Partly. The surveillance system still generates the alerts; an AI agent can take each one and build the picture around it: the orders, the market moves, the news, the trader's messages and their history. Deciding that conduct is suspicious, reporting it to the regulator and acting against a trader stay with a surveillance analyst or compliance officer.

Where does the agent fit in a surveillance programme?

Market abuse surveillance has two halves. Detection runs calibrated scenarios over orders and trades, looking for patterns such as trading ahead of announcements, layering, spoofing or marking the close. Review takes each alert and decides whether there is a reasonable suspicion of abuse.

Detection should stay with the surveillance system, where scenarios are documented, tested and tuned under governance. Review is where agents help. Most alerts close as explained by normal activity, but each one still needs someone to pull five or six data sources together before they can say so. That assembly is what consumes an analyst's day.

What evidence does the agent assemble for each alert?

• The order and trade record for the account across the alert window, including amendments and cancellations, not just fills.
• Market context: price and volume for the instrument, sector moves and index moves over the same period.
• Public information: announcements, results dates and news that could explain the trade.
• Inside information records: whether the instrument was on the watch list or restricted list, and whether the trader's desk was wall-crossed.
• Communications around the trade, drawn from communications surveillance records under the agent's role scope.
• History: the trader's previous alerts in this scenario and how they were closed.

It then drafts an alert narrative: what happened, what explains it, what does not, and a recommended disposition with the evidence for it. The working method for order-based alerts is in how to review trade surveillance alerts with an AI agent.

Which decisions stay with a person?

The analyst or compliance officer decides:

• Whether an alert is closed, escalated to an investigation or kept open for more data.
• Whether there are reasonable grounds to suspect abuse, which triggers the duty to report suspicious orders and transactions to the regulator.
• Whether to speak to the trader or their manager.
• Any restriction on the trader, the account or the instrument.
• Changes to scenarios, thresholds or the watch list.

The reason is consequence. A report to the regulator can lead to an investigation of a named person. Closing an alert wrongly can leave abuse unreported. Both carry personal accountability for the people who make them, so the agent recommends and they decide.

The agent must also never contact the trader or anyone on the desk about an alert. A message from a compliance tool asking about a trade is exactly the kind of signal that warns someone they are being looked at.

What does a review look like?

Take an asset manager. The system raises an insider dealing alert: a portfolio manager bought shares in a mid-sized retailer two days before it announced a takeover offer, and the price rose sharply on the news.

The agent pulls the order history and finds the purchase was part of a gradual build over three weeks, in line with a rebalancing note filed a month earlier. The retailer was not on the watch list. The desk had no wall-crossing record. The manager's messages around the trade discuss sector valuations but not the company. Similar trades by the same manager over the past year closed without concern.

The agent recommends closure and lists what it checked and what it could not see, including the manager's personal phone, which is outside any record the firm holds. The analyst reads the draft, adds a note on the rebalancing timeline and closes the alert. Had the build started after a wall-crossing, the same evidence would have pointed the other way, and the decision to escalate would still have been the analyst's.

What controls make this defensible?

• Scenarios stay governed. The agent does not generate its own alerts or quietly widen thresholds.
• Every check is listed, including null results. A reviewer can see the agent looked for a wall-crossing record and found none.
• Closure samples. A person re-reviews a sample of alerts the agent recommended closing, every cycle.
• No contact with the business. The agent's role scope excludes messaging traders or desks.
• Draft and decision on one record. The disposition record shows what the agent recommended, what the analyst decided and why. If your messaging alerts are reviewed the same way, see how to review communications surveillance alerts with an AI agent.

The rule for compliance agents generally, that they prepare and people decide, is in AI compliance agents: second line, never the last word.