The FCA Mills Review Puts Agentic Finance on the Record: Who Is Accountable When an AI Agent Moves the Money

On July 6, 2026 the Financial Conduct Authority published the Mills Review, led by executive director Sheldon Mills. The FCA called it the first review of its kind initiated by a financial regulator anywhere. It names four shifts AI is driving in retail financial services, makes seven recommendations, and cites FCA-commissioned research finding that roughly 11 million UK adults are likely to use AI that acts autonomously within pre-set goals. One of the seven recommendations is to "enable the foundations for agentic finance."

I want to stay on one thread inside a wide-ranging document: the accountability gap the Review opens and does not close. Once an AI agent acts inside a customer's pre-set goals, a firm has to be able to prove, per action, which identity authorized it, what policy applied, and what the agent asked the model to do. That evidence has to exist at the moment of the call. I am not going to write about consumer duty or market concentration, which the Review also covers and which sit outside what I can speak to usefully.

What the Review says

The Mills Review names four shifts AI is driving across retail financial services: changes to firm operations, changes to consumer journeys, changes to competition and market power, and amplification of fraud and cyber risk. It makes seven recommendations, among them enabling the foundations for agentic finance and building an AI-enabled agentic supervisory model inside the FCA itself. Norton Rose Fulbright's Global Regulation Tomorrow summary is a useful secondary read on the recommendations.

The number that anchors the whole thing is the 11 million UK adults the FCA's commissioned research expects to use AI that acts autonomously within pre-set goals. That is agentic finance stated as a near-term consumer reality rather than a lab concept. An agent that operates inside a customer's stated goal, moving money or opening products within limits the customer set, is the case the Review puts on the regulatory record.

The accountability gap the Review opens

Agentic finance moves the point of action away from the human. A customer sets a goal, and an agent executes inside it across many steps and many days. The Review encourages this direction while asking who is accountable when the agent acts.

The gap is evidentiary. When a human clicks, the firm has a record of the human's action. When an agent acts inside a pre-set goal, the firm needs an equivalent record for each thing the agent did: which identity authorized this specific action, what policy was in force at that moment, and what the agent actually asked the underlying model to do. Without that, a firm faced with a complaint or an FCA inquiry can describe its agent's general behavior but cannot show what happened on the specific action in dispute. General assurance about how the agent was designed is a different artifact than a record of what it did at 14:32 on a Tuesday.

Why the evidence has to exist at call time

The failure mode is reconstruction after the fact. A firm that logs agent activity loosely, or relies on the application that ran the agent to also account for it, is trusting the system under review to write its own record. That record can be incomplete, can be written by a component that crashed before it committed, and comes from the same system whose behavior is in question.

Per-action accountability means the record is created at the moment the agent acts, bound to the agent's identity, and held independently of the application. Once an agent has taken 400 actions inside a customer's goal over a week, there is no reliable way to reconstruct per-action authorization afterward if it was not captured at each call. The evidence either exists at call time or it does not exist in a form an FCA inquiry will accept. This is the same action lineage requirement that NIST's agent identity work describes, arriving now through a UK conduct regulator.

Where this sits against the other 2026 finance regimes

The Mills Review is UK, forward-looking, and centered on conduct and the consumer. It reads differently from the two other finance-AI developments moving this year. DORA's ICT third-party risk regime, which EU banks are mapping toward the register and exit-strategy requirements, is EU operational-resilience law. The ECB's July 7 letter to 110 supervised banks is a eurozone supervisory filing with an October 31 deadline. The Mills Review is none of those. It is a UK review that puts agentic finance on the record and asks the accountability question the other two do not center.

What the three share is an underlying evidence requirement. Each one, from its own angle, ends at the same operational need: a firm has to be able to show what a specific AI action did, under which identity and policy, at the moment it happened. The vocabulary differs across the three regimes. The record that satisfies them is the same record.

DeepInspect

This is the gap DeepInspect closes for the accountability question the Mills Review raises. DeepInspect is a stateless proxy at the AI request boundary. When an agent acts inside a customer's pre-set goal, each call the agent makes to a model passes through DeepInspect, which binds it to the agent identity the application supplies, applies the policy in force at that moment, and commits a signed per-decision record before the action completes.

For a UK firm building toward agentic finance, that produces the per-action evidence the Review's accountability question demands: which identity authorized this action, what policy governed it, what the agent asked the model, and what the outcome was, captured at call time rather than reconstructed. It does not make the FCA's consumer-duty or competition judgments for you, and it does not design your agent's decision logic. It records what the agent did to the model, per action, in a form an inquiry can inspect.

If you are building agentic finance products and want per-action accountability in place before the FCA asks for it, let's talk today.