On 3 July I said the machine story lets everyone off the hook. A box in a Bank report does not change that.

Four weeks ago I argued that the entry-level market really has collapsed, and that the cause we have settled on is the one letting the most people off the hook. This edition has found a box in a Bank of England report and treated it as the corroboration that was missing. It is not, for four reasons.

Start with what Agents' intelligence is. It is a summary of conversations. Bank staff talk to company contacts, write down what they hear, and aggregate it. There is no sampling frame, no denominator, no control group and no verification. If every firm in Britain told the Agents the same wrong thing, the box would report it with complete confidence.

Second, ask what a firm gains by naming automation.

A company that has cut its graduate intake has three explanations available. It over-hired and is correcting. Employer National Insurance and wage floors have made junior staff cost more than they produce. Or the work is now done by software. The first makes the management look foolish. The second makes the company sound as though it is blaming the government for its own margins. The third makes it sound modern and disciplined, and it is the only one of the three a finance director can offer a central bank without it reading as a complaint. That is not lying. People believe the flattering account of themselves, which is precisely what makes it unreliable.

Third, read the Bank's own ordering, which this edition quotes accurately and then walks slightly past. Employment intentions are broadly flat, the Agents say, reflecting weak and uncertain demand, then labour cost pressures, then increased use of automation. Third. Behind two things that have nothing to do with AI, in a document that also describes an energy shock, squeezed margins, falling construction output and consumers who have stopped buying anything expensive. An economy in that condition would be shedding junior roles if the transformer had never been invented.

Fourth, the outreach section Ada builds on is made up of people who turned up to a Bank engagement event. Those who feel the labour market is going badly are more likely to attend a meeting about the labour market than those who do not.

Now the concession, and it is a real one.

The graduate recruitment line is specific in a way the rest is not. The Agents did not ask firms to rank causes and happen to get automation third. Professional services contacts volunteered that they are recruiting fewer graduates, and named the tasks that have gone: document preparation, invoice processing, basic analysis. Those are recognisable jobs, described by the people who stopped offering them. It is the strongest single piece of evidence this country has, and the editor is right that it did not exist a month ago.

It is still not a measurement, and the space between the two is where the next five years of bad policy will be made. A government that believes AI is taking entry-level jobs will spend money on retraining. A government that believes employer taxes are taking them will spend money on payroll relief. Those are opposite policies of similar cost, and this box cannot tell you which one is right.

The editor says the finding turned up sideways, in somebody else's paperwork. Agreed. My point is that sideways evidence is exactly how a country becomes confident about the wrong thing, and a month spent being careful about the causal story is not a reason to stop being careful in the week something finally appears to confirm it.