“Tweak the contract” triggered the surveillance control. But the surrounding exchange concerned a kitchen renovation. Alex de Lucena, Director of Product and Governance Strategy at Shield shared this example in a recent discussion with RegTech Insight following the publication of the FCA Board-sponsored Mills Review in July 2026.
In late 2025, the FCA Board asked Executive Director, Consumers & Competition, Sheldon Mills to examine how advances in artificial intelligence could transform retail financial services by 2030 and beyond.
The review examines four connected areas:
- How AI technology might evolve, particularly towards more powerful, autonomous and agentic systems.
- How those developments could affect firms and markets, including competition, market structure and UK competitiveness.
- How AI could affect consumers, including outcomes, risks, behaviour and demand for financial services.
- How financial regulators might need to evolve to ensure retail financial markets continued to work well.
The resulting Mills Review, although focused on retail financial services, offers several insights relevant to capital markets and wholesale treasury.
Agents as Process-Actors
Mills describes five roles for the human along a spectrum of increasing AI autonomy: operator, collaborator, consultant, approver and observer. Read as a process model, the framework can also help firms identify who acts at each stage: a human, an agent or both. At observer level, the AI system acts continuously within preset limits while the human monitors outcomes. As the Review puts it: “Control shifts towards setting permissions and boundaries.”Different positions on the spectrum may coexist within a single surveillance workflow. One agent might assemble the context surrounding an alert, another assess the suspected risk and a third challenges that assessment. Closing an alert gives an agent more authority than summarising one, even when both tasks use the same model. Controls should therefore cover each process-actor’s task, authority and decisions, as well as the underlying technology.
Mills notes that “Some harms only become discernible when viewed across the market.” Its proposed Agentic Supervisory Model could include tools that analyse transaction and order-book data across markets to identify potential market abuse more quickly.
Near-continuous supervision would depend on comparable data, agreed semantics and rules governing what an FCA agent could request or infer. It could also generate noise at regulatory scale. A signal connected across firms is not automatically evidence of misconduct, and a supervisory agent’s conclusion would require the same provenance and challenge expected of firms.
Shield’s de Lucena reads Mills as a “permission structure” for firms that have tested AI but remain wary of putting it into production. He sees the five roles framework helping firms describe increasing autonomy and where existing accountability arrangements may come under strain as agentic solutions move into production. “At scale, we have to be much more deliberate in how we build agents; we have to govern them,” he says.A narrowly drawn agent has a stated job, known inputs and outputs that can be examined. In the kitchen example, the agent records why the surrounding conversation changes the meaning of the flagged phrase.
Shield has an agent that can make that distinction and close the case. “It closes it as if there was a reviewer and it provides a reason for why it closed it,” says de Lucena. The alert stays in the system. A reviewer can reopen it, spot-check the decision or draw a random sample. See: Compliance risk is connected. Your AI should be too.
Agents can also give firms a way to test for false negatives. Traditional surveillance assurance has focused largely on the alerts a system produced: whether analysts reviewed them promptly, whether cases were handled consistently and how many proved irrelevant. But what about the messages that never entered the queue?
After a message triggers an alert for one type of conduct risk, Shield’s agents examine the wider context and check the same message for other risks the client has chosen to monitor. “We have agents specifically looking for anything we might have missed,” said de Lucena.
The arrangement provides a basis for a recorded false-negative test that firms can sample, compare with human review and track after a model or policy change. It cannot establish complete coverage, but it offers more evidence than a broad claim about alert quality.
Subscribe to our newsletter
