The Jobs Number Everyone Steers By Is a Poll People Stopped Answering. So the Central Bank Built a Shadow Version.

Judge a café by its customer reviews and you will misjudge the busy days, because the busiest customers are the ones with no time to review. Now imagine the café's prices — the reference every kitchen in town copies — are set from those same reviews. That is closer to how a central bank works than most investors realize, and it is why the Bank of England, about the least excitable institution in finance, is quietly building a shadow economy out of web pages and AI.

This week its chief economist, Huw Pill, said official data reliability is increasingly coming into question and that the Bank has begun using AI models and web scraping to pull a usable picture of the economy out of corporate reports, survey responses, and the conversations its field agents hold with businesses around the country. Before you file that under "central bank trivia," slow down: those surveyed numbers are what set interest rates, and interest rates are the discount rate stamped on every stock you might own. If the data underneath them is a poll people stopped answering, then the dial gets turned on a guess.

The poll that stopped being a count

The word "survey" hides the whole problem. "Jobs data" sounds like a count, but in the UK's main Labour Force Survey it is a poll: statisticians knock on a sample of doors and ask people whether they are working, looking, or out of the workforce. A poll is only as good as the fraction of people who answer — and who those people are.

Run a toy version with a hundred people, seventy at work. A perfect poll returns exactly that: a seventy percent employment rate. Now make it realistic. Workers are busy, so suppose only half of them bother to answer, while four out of five people without a job answer because they have the time. The replies that come back are 35 workers and 24 non-workers. In the returned sample the employment rate reads about 59 percent, not 70. Nothing about reality changed; the data just started lying in one direction. That is non-response bias, and it is worse than noise, because normal noise cancels out while a slanted sample pushes the number the same way every month.

The real UK survey has gone through exactly that spin. The response rate fell from roughly 48 percent in 2014 to about 39 percent in 2019, then collapsed to around 17 percent at the start of 2024, and the sample size roughly halved. Because workers are precisely the people least likely to answer, the survey now undercounts employment by more than a million people compared with hard payroll and business data; the published employment rate reads 74.5 percent while the alternative sources imply something above 76 percent. The delivery is wobbling too. In June the Office for National Statistics admitted an operational error had cut the household interviews feeding its July jobs release by around 1,200 — a drop of about 19 percent — and acknowledged the fix might not arrive until 2027.

Counting the receipts instead

So the Bank of England is doing what the café manager does when the reviews stop coming: it stops reading reviews and starts counting the receipts. Its network of agents already holds roughly six thousand confidential one-to-one conversations with businesses a year — a human, on-the-ground survey that is its own thing. On top of that, it is feeding AI models with web scrapes to extract quantitative signals from corporate reports and those agent conversations, watching what companies themselves say in near real time rather than waiting for a monthly poll that arrives already revised. Pill called building these new information sources "at the heart of our strategy" and traced the push to the pandemic, when the in-person surveys became impossible to run.

Map the props and the logic is clean: the customer reviews are the official surveys; the busy customers who never review are the employed people who never answer; the manager counting the receipts, reservations, and kitchen orders is the AI and web scraping. Same player in both worlds — someone who needs to know what is really happening and cannot trust the people being asked to describe it.

Where the shadow breaks

The analogy has now done its job, so here is where it breaks. The receipts are not the truth either — they are just a different kind of picture. Web scraping only sees what is posted online, so it misses the cash-only shop, the worker with no digital footprint, exactly the sort of person whose disappearance from a survey created the hole in the first place. And corporate reports are written by the companies being measured, which is its own bias, just pointed a different way. The shadow measure is faster and does not depend on anyone answering a knock on the door, but it trades one sampling problem for a different one; it fills the gap in time, not in truth.

The other break is institutional and it matters for how you read the news. No single number sets a rate. The Bank's Monetary Policy Committee has nine members, and Pill noted he has voted at forty meetings with the committee unanimous only once — a reminder that a month's jobs figure is one input into a long human argument, not a switch that flips rates. When a headline jobs number moves a market, the move is partly about that number and partly about the widening suspicion that the underlying series cannot be trusted.

Bring the model back to your own screen, because this is not a UK quirk. The U.S. Census Current Population Survey — the household survey behind America's headline unemployment rate — saw its response rate slide from above 90 percent in 2010 to below 70 percent. The same machine runs on both sides of the Atlantic. So the one test worth carrying is this: when a jobs number moves a market this month, ask whether the underlying series is a survey or a hard count. If it is a survey, it is a poll with a bias arrow, not a fact — and the number you are looking at today will likely be revised next month. A single print is a reason to be curious about the range of possible rate paths, not a reason to rebuild a position around one day's wiggle.