The Financial Conduct Authority's Mills Review sets out how wealth managers can deploy AI continuously inside agreed mandates, and who is liable when the machine gets it wrong.
The UK Financial Conduct Authority has spent the year asking what AI should be allowed to do in wealth management. The Mills Review, published 6 July 2026, is its answer, and it draws a sharper line than the trade press has so far captured.
The review, commissioned by the FCA in January and led by the regulator's executive director of consumers and competition Sheldon Mills, is the first formal UK regulatory document to treat AI as an agent that acts on a client's behalf inside an agreed mandate, rather than as a tool an adviser picks up.
The review's central forecast is a shift from "human-led, episodic financial activity" to services that are "AI-enabled, continuous and delegated." Trade press has shortened this to "agentic finance." The review itself avoids the label. The operating model it sketches, though, is concrete enough to test against three scenarios the document lays out for investment management.
First, analysts work alongside AI on investment themes rather than generating them alone. The review expects firms to deploy AI to surface and stress-test ideas, with humans retaining research ownership. Second, advisers use AI to compare products from a wider catalogue before signing off on a recommendation, narrowing the search cost that has historically given human advisers an information edge. Third, AI systems carry out portfolio management tasks within agreed mandates while clients monitor outcomes through dashboards and exception alerts.
In all three scenarios, the human stays in the loop. The legal responsibility for what the machine does inside the mandate stays with the adviser. Trade press has tended to frame this as a forecast about technology adoption. It is also a forecast about liability.
According to Finextra's summary, the report contains seven recommendations for the FCA Board and executive. The substance of those recommendations, and the timetable for turning them into supervisory expectations, will determine whether the operating model above is a sketch or a binding map.
Freshfields' analysis notes that the regulator's vision assumes AI systems that are accurate and free of bias, which the market has not yet demonstrated at scale. Baker McKenzie's regulatory briefing and Norton Rose Fulbright's regulation tracker flag the same gap: the legal framework is ready for AI agents, but the assurance regime for those agents is not.
The operating question for wealth managers is no longer whether AI will review portfolios or draft recommendations. That is already happening. The question is who signs off, who monitors the AI's actions inside the mandate, and whose name is on the disclosure when the model produces a result the client did not expect.
The Mills Review answers that question by putting the adviser's name on the disclosure. Firms that read it only as a forecast about AI adoption will miss the part that is binding. The next test will be how the FCA supervises the first conduct cases under the model, and whether the seven recommendations turn into enforceable rules before the review's 2030 horizon arrives.