AI in action · Regulation
The Bank of England's Answer to Frontier AI: Test the Models Before and After Deployment, Write the Rules Later
In an essay published on September 30, 2026, Governor Andrew Bailey argued that society must keep the ability to set and revise the boundaries within which frontier AI systems operate, and that rigorous testing before and after deployment, not a new rulebook, is the place to start. The essay names agentic trading and payments as areas where tested models could be deployed safely.
Central bankers do not usually write essays about artificial intelligence, and when the governor of the Bank of England does, it is worth reading what he actually proposes. Andrew Bailey's "Frontier AI and the question of governance", published on the Bank's Insights page on September 30, 2026, begins with a question rather than a rule: "what precisely is the problem we are trying to solve?" His answer is that frontier models combine vast knowledge synthesis with recursive self-learning, which could let them operate outside the oversight of the society that built them. The central challenge, he writes, is "how to make sure that, as these systems become more capable, society retains the capacity to govern them". Asked whether society should keep the power to intervene, he answers "unequivocally yes".
The essay does not call for a ban or for a new regulatory framework now. Bailey says the benefits of frontier AI could be large and warns against premature attempts to regulate a technology that is changing this quickly. What he asks for instead is sequencing: "Understanding, testing and establishing credible points of intervention must come first." A more formal framework may follow, but it is not the starting point.
Testing is the practical centre of the argument. Bailey wants rigorous model testing both before deployment and after, so that vulnerabilities are found, safeguards gain credibility and the results feed into deployment standards. He credits the UK's AI Security Institute with building the scientific base for this kind of assurance and says the work has to speed up as models are used outside controlled settings. He is explicit that testing is not a cure: "Testing will not eliminate failures." Incidents and near misses, in his framing, are data for the next round of tests rather than evidence that the approach has failed.
For financial markets the relevant passages are about cyber risk and agents. The Bank's view, repeated from its Financial Policy Committee's September record, is that frontier AI raises the scale and sophistication of cyber threats to payment systems, market infrastructure and banks, which therefore have to be assessed alongside the technology. The essay also says that because firms will use these models internally, testing them could help the sector deploy AI safely in cyber defence and in agentic trading and payments, and could over time produce shared standards across finance.
The essay sits alongside two other UK positions published within a fortnight of each other: the Financial Conduct Authority telling MPs on October 13 that it will automate its own supervision but will not write AI-specific rules for firms, and the Financial Policy Committee flagging AI-related debt issuance as a growing risk. Read together, the message from London is consistent: no new rulebook yet, heavier testing and supervision now, and an insistence that whoever deploys the model remains accountable for what it does.
Sources: Bank of England — Frontier AI and the question of governance (Andrew Bailey, Sep 30, 2026); Norton Rose Fulbright, Regulation Tomorrow — Frontier AI and the question of governance (summary); UK Parliament Treasury Committee — Bank of England and FCA commit to action on AI following warnings from MPs