A financial adviser who never sleeps, never asks for a fee negotiation and never second-guesses her own judgement sounds, on paper, like exactly what a stretched public service should want. Sheldon Mills's review for the Financial Conduct Authority does not dismiss that appeal. It simply asks who answers for the adviser when something is wrong. That question, more than any of the review's specific recommendations, is the one worth carrying across from financial regulation into healthcare, because the NHS is already living inside a version of it.

The Mills Review's central anxiety is that as decisions move from tools that assist a human to agents that act within a goal, accountability becomes harder to trace. Roughly a fifth of UK adults are already estimated to be using AI capable of some autonomous action, according to the review, while trust in how that AI is governed lags well behind its adoption. Consumers, it notes, rarely check whether a given tool has passed through any formal regulatory route before relying on it. Replace consumer with patient and adviser with ambient scribe, and the description holds almost without editing.

Ambient voice technology, the systems that listen to a GP consultation and produce a structured clinical note, has gone from a curiosity to a fixture of primary care in barely two years. NHS England's own guidance now treats a scribe that summarises rather than merely transcribes as a medical device requiring at least Class I registration, and a supplier registry published in January lists the products cleared to operate under that standard. Yet the pattern the Mills Review flags in finance, tools deployed faster than the public or the workforce using them can verify their regulatory status, has already forced NHS England into a corrective posture once this year, when the Chief Clinical Information Officer issued a notice ordering the removal of non-compliant AVT products from clinical use. The problem was not that the technology had failed. It was that adoption had outpaced anyone's ability to say with confidence which version of it had been properly checked.

What the FCA and the MHRA are converging on, from opposite sides of the economy, is the recognition that principles-based regulation assumes a human is still making the final call. By design, agentic systems undermine that presumption. The MHRA's National Commission on the Regulation of AI in Healthcare, chaired by Professor Alastair Denniston, received more than 770 submissions to a call for evidence that closed in February, many of them wrestling with exactly the liability question the Mills Review poses for finance: who is responsible when an autonomous or semi-autonomous system produces an error that a human did not catch, because the system was designed to reduce the amount of human catching required.

There is a political dimension neither regulator can avoid stating plainly, even if both are inclined to hedge it. Agentic AI is attractive to institutions under fiscal strain precisely because it promises output without headcount. For a government that needs visible evidence of NHS productivity without the capital spend that deeper reform would require, ambient AI and its successors are a low-cost, high-visibility answer. That is not an excuse to oppose technology. It is a reason to be sceptical of regulatory frameworks built to keep pace with political urgency rather than clinical or financial risk.

The lesson from the Mills Review that NHS leaders and the MHRA's Commission should take most seriously is not about finance at all. It is that a regulator scaling up its capacity to supervise autonomous systems, as the FCA now proposes for its own AI Lab, is an admission that the old supervisory model, built around approving a product once and monitoring a human's use of it, no longer matches how these systems behave once deployed. The NHS's regulatory architecture for AI is being rewritten this year on almost identical grounds. Whether that rewrite produces genuine accountability or merely a faster path to procurement will determine whether the next compliance notice arrives before or after a patient is harmed.