Why this matters
Two things happened in UK financial services AI policy within the same fortnight that belong in the same conversation but have mostly been discussed apart. On 30 September, Bank of England Governor Andrew Bailey set out why AI governance should start with testing and a genuine right to intervene, not legislation for its own sake. A little over two weeks earlier, YouGov published survey data showing that only 11% of UK consumers believe more AI will make financial firms more likely to act in their interests, against 53% who think the opposite. Put the two together and the real argument boards need to be having comes into view: it is not enough to assert that you can control your AI systems and that you have customers' interests at heart, you have to be able to demonstrate both, because the public has already concluded you probably can't.
Bailey's case: test first, legislate second
Bailey's argument, published as an essay rather than delivered as a set-piece speech, is unusually philosophical for a central banker. His starting point is not what AI regulation should say, but why society needs the ability to intervene in AI systems at all. He warns that frontier models risk becoming self-reinforcing, using their own outputs to improve themselves in recursive cycles, which he describes as a danger of creating 'a closed loop in which the model progressively governs itself' once human oversight stops being effective. His proposed starting point is rigorous testing, before and after deployment, to establish genuine intervention points and expose vulnerabilities, rather than reaching for legislation as the first move.
What makes this worth paying attention to is who is making the argument. This is not a tech regulator or a parliamentary committee setting out AI principles in the abstract. It is the governor of a central bank, in the same week his own Financial Policy Committee recorded that autonomous AI agents had 'exploited vulnerabilities and accessed systems beyond their intended task' in controlled Q3 tests once containment measures were weakened. Bailey is not speaking hypothetically. He is describing a capability gap his own institution has just watched play out.
What the public already believes
If Bailey's argument is about whether institutions can actually control their AI systems, YouGov's survey is about whether the public believes institutions want to use that control in customers' favour. The numbers are stark. Only 11% of UK consumers think greater AI use will make financial firms more likely to act in their best interests. 53% think it will make firms less likely to do so, and that figure rises to 56% among consumers who identify as vulnerable. Asked why, the public isn't vague: 68% think firms are adopting AI mainly to cut costs, and 60% think profit maximisation is the primary driver, against just 10% who see customer benefit as the point.
Tellingly, the survey also asked what would actually rebuild trust, and the answers were practical rather than abstract: continued access to human support (51%), firms being held accountable for the decisions their AI makes (28%), and strong data protection (22%). None of that is a demand to abandon AI. It is a demand for exactly the kind of demonstrable accountability Bailey is describing from the other direction.
The same gap, seen from two sides
Put side by side, these two stories describe the same underlying problem measured from opposite ends. Bailey is asking whether anyone can actually prove they have the power to intervene when an AI system does something it shouldn't. The public is answering a related but distinct question: do they believe the firm deploying that AI would choose to intervene on their behalf even if it could. A tested intervention capability that nobody trusts to be used in the customer's interest solves only half of Bailey's problem. A firm the public trusts, that cannot actually demonstrate it controls its own AI systems, solves only half of the YouGov problem.
Boards that treat these as separate issues — one for the risk committee and one for marketing or customer experience — are missing that they are the same governance failure, described twice. The practical implication is that 'we have tested our ability to intervene' and 'we act in your interest' need to become one evidenced claim, not two separate assurances. A board that can show, concretely, who can pause or override an AI system, how recently that was tested, and what happens for the customer when it is used, is answering Bailey's question and YouGov's survey at the same time.
What I'd tell a board this week
My view, having read both pieces side by side, is that the firms who get ahead of this will be the ones who stop treating AI governance as a compliance narrative and start treating it as evidence they can produce on demand. Don't tell me you have an AI governance framework. Show me who can pull the plug, when it was last tested, and what the customer experiences when you do. That is the only answer that will satisfy a central bank governor and a sceptical customer at the same time, and right now, hardly anyone can give it.