Until February a machine could not refuse your loan on its own. Now it can, and the rule about when a person must be involved has never been written down.

On 9 July we published a hard number. The ONS recorded financial and insurance activities down seventy-eight thousand jobs across 2025, more than any other sector, in the same fortnight the Financial Conduct Authority set out a five-rung ladder ending at the word observer. Leah wrote then that nobody had shown the two were connected and nobody in authority had asked.

Six weeks on I want to put a second fact next to that one, because on its own the first is easy to misread.

Underwriting is the purest white-collar job there is. It is judgement about a stranger and nothing else. No physical component protects it, the work is already numerical, and the data needed to automate it belongs to the industry doing the automating. If any office job should have gone by now, it is this one.

It has not gone. What the firms describe automating is the reading. Aviva's tool, launched for life cover in November and extended to critical illness in March, works at the front end, taking in the file so an underwriter can decide faster. Zurich describes its rollout as taking admin away from underwriters. Even the one vendor selling an agent that can quote and bind end to end still writes that the carrier's underwriters remain accountable.

So the sector shed seventy-eight thousand jobs and kept the deciders. Both those things are true and I do not think they are in tension. What went was the processing around the decision. What stayed was the person who answers for it.

Then the part that made me sit up. The reason the person stayed is not that the machine cannot do it. Nobody in the industry says that. The reason is that a rule required a human to be inside any significant automated decision about a British person. And on 5 February that rule was replaced.

The old position was a prohibition with narrow exceptions. The new one permits the machine to decide alone, with safeguards after the fact, everywhere except decisions using special category data such as health. Ines sets out below what that predicts, and it predicts Aviva's choices almost exactly.

The whole regime now turns on whether a decision had meaningful human involvement. Parliament did not define that phrase. It handed a Secretary of State the power to define it by regulation, and as of this morning no such regulation exists. Elena has the shape of that below and it is not a party matter, which is rather the point.

I am not going to tell you the old rule was better. That is an argument with two decent sides and it is not mine to settle in your inbox. What I will say is that if the thing standing between a judgement job and a machine is a legal default rather than any limit on the machine, then the job is safe exactly as long as the default holds. Capability does not go backwards. Defaults change on a Thursday.

Standing disclosure, as always. This note is mine and carries my name. It is drafted with the same AI that writes the rest of this paper, and every analyst desk here runs on models built by Anthropic. A paper about what machines do to human work does not hide that it was made with them.

— M.