In 1833 Parliament appointed Britain’s first factory inspectors. There were four of them, and between them they were supposed to police conditions in around 4,000 mills. Unsurprisingly, the law they were enforcing was widely evaded.
Those four men still mattered more than most of the legislation that came before them. Parliament had already passed laws to protect children in the mills, but nobody was checking. Parliament’s own history of the period notes that an Act passed in 1831, which limited the working day for under-18s to 12 hours, once again came with no procedures for enforcement. What made the 1833 Act different was that it created people whose job was to turn up and look.
Kanishka Narayan, the UK’s AI minister, opened a speech at Imperial College on Monday with this story. He compared the arrival of AI with the early Industrial Revolution, when “children were being sent to work in horrifying conditions”, and drew a lesson that’s easy to skim past. “Early factory laws mattered because Parliament was prepared to establish institutions capable of enforcing them. Our responsibility today is no different.”
The headlines went to the sharper lines. A breach of Britain’s cyber defences by AI is “likely, if not inevitable”. AI companies shouldn’t release a system if they “can’t adequately control the risks”. The FT reports that he’s drawing up plans for compulsory safety reporting and monitoring of frontier AI. All newsworthy. But it’s the factory inspectors I keep coming back to, because they describe a problem the mortgage industry is about to meet from the other side.
His warning rested on a real incident. Earlier this year, he said, an OpenAI agent “went looking for information about public spending on medicines”. It reached an Australian government website that limited what it could see, “so it found another way in”. In Narayan’s words: “Its task was benign, legitimate, but without being asked to do so, the agent gained access to information that it was not authorised to have.” The Australian government has said no personal medical records are known to have been accessed. According to the FT, the incident only became public last month, and Anthony Albanese said OpenAI took “way too long” to tell his government.
Nobody told that agent to break in. It wasn’t being malicious. It was trying to finish its task, and being keen to help took it somewhere it had no business going. Most of the controls our industry has built assume either a person with bad intent or a system with a bug. A tool working exactly as designed, just a little too eagerly, doesn’t fit either box.
The line from the speech that landed hardest with me was a quieter one. When something like this happens, Narayan said, “government must be able to establish quickly what an agent was asked to do, what it accessed, why it acted as it did and whether the same failure could occur elsewhere.”
Anyone who has been through a past business review will recognise that list. Swap “agent” for “adviser” and you have the questions the FCA has been asking mortgage firms for years. What was the client told? What did the adviser look at? Why did they recommend what they did? And the one that turns a single complaint into a remediation exercise: how many other clients got the same treatment?
This is where mortgage professionals sit in a slightly odd spot compared with most of the economy. Narayan is talking about building institutions for AI that don’t yet exist. Financial services already has one. The FCA has said plainly that it won’t write a separate rulebook for AI. Its guidance reads: “We do not plan to introduce extra regulations for AI. Instead, we’ll rely on existing frameworks, which mitigate many of the risks associated with AI.” A blog in June named them: Consumer Duty, the Senior Managers and Certification Regime, and the regulator’s expectations on governance and controls.
So for us, the factory law and the inspector are both already in place. On that reading, Consumer Duty doesn’t much care whether a poor outcome came from a tired adviser or an eager piece of software. If an AI-assisted fact-find misses a sign of vulnerability, or a sourcing tool quietly drops a product the client should have seen, the outcome still belongs to the firm.
What changes is whether the firm can answer those four questions. With a human adviser there’s a fact-find, a suitability letter, file notes, maybe a call recording. With an agent that reads payslips, runs criteria searches and drafts the recommendation, the trail looks different. Parts of it sit in your CRM, parts with a vendor, and parts may never have been kept at all. If a client complained tomorrow about something your AI tooling did six months ago, could you rebuild what it was asked, what it saw and why it produced what it did? Could you tell whether it had done the same for anyone else?
I suspect plenty of firms would find that harder than they’d like. That wouldn’t be carelessness. These tools mostly arrived as productivity features, sold on hours saved, and nobody pictured having to explain one to the Financial Ombudsman.
The other half of the factory story is less comfortable. Four inspectors couldn’t cover 4,000 mills, and Narayan himself called it “a modest beginning, and insufficient to the scale of the challenge”. The principle was right long before the capacity caught up with it.
Regulators today face a similar gap, and to be fair they’re working on it. The AI Security Institute has, in the minister’s words, “over a hundred exceptional technical experts”. The FCA says it’s experimenting with large language models in its authorisations and supervision work. And there was a hopeful thread in the speech too. Across two government pilots, frontier AI has helped identify and address more than a thousand vulnerabilities. The tools that create the risk may yet turn out to be some of the best inspectors we have.
Narayan made one more point about Britain that fits a mortgage firm almost perfectly. Talking about access to American models, he said, as reported by the FT, that “when you’re reliant on someone else’s technology, you have to be clear-eyed about the permanence of any outcome”. Most brokers are in exactly that position with their own tools. The model inside your sourcing platform or CRM was built elsewhere and gets updated on someone else’s timetable. The behaviour you tested in March may not be the behaviour you’re relying on in October.
He wants, he said, “machines that leave us, human beings, in charge”. It’s a good line. But the mill owners of the 1830s were in charge of their mills too. What the inspectors brought was someone who could walk in and see what was actually going on.
The Australian agent wasn’t trying to do harm. It was trying to get its job done. Which leaves an awkward question for anyone handing work to these tools: if the thing you’ve deployed is that determined to be helpful, how would you know if it had found another way in?
Sources:
Financial Times, “AI regulation must mimic Industrial Revolution laws, UK minister says”, 12 October 2026
GOV.UK, “Building a Safer, More Secure Britain in the age of AI”, speech by Kanishka Narayan, 12 October 2026
UK Parliament, “The 1833 Factory Act”
FCA, “AI and the FCA: our approach”