On November 9, 2026, the Federal Reserve Bank of Cleveland and the Reserve Bank of Australia each published a short note on where AI may first show up in labour market data. The Cleveland commentary by Alexander Cline and Barış Kaymak reports that since late 2022, unemployment has risen more in occupations whose tasks are more exposed to AI, especially among young workers, and that the main driver is slower job finding. The authors do not find consistent evidence of elevated job separation rates in those occupations. The Australian note is more cautious about its own data, and it describes the international evidence as pointing at job finding rather than layoffs.
That is a hiring pattern, and hiring patterns are easy to miss. A layoff is a visible event that tends to leave a trail, while a requisition that is never opened, an entry-level role quietly redesigned around an AI tool, or a hiring plan that assumes AI will absorb junior tasks can pass with no visible event at all. Unless someone writes it down, it can also pass with no named decision. This column asks one question of those choices: who decided, and can that person show the decision?

Where the central banks see the early signal
Cline and Kaymak compare a pre-AI baseline of 2015 to 2019 with the period since. They estimate that the unemployment rate for workers aged 23 to 54 has risen by roughly 1 percentage point more in the most AI-exposed occupations than in the least exposed. On the two task-based exposure measures, job-finding rates fell by roughly six percentage points more in the most exposed occupations, in a market where the monthly job-finding rate for unemployed workers averages around 34 percent. The implied average unemployment duration rises from about 10.5 weeks to more than 13 weeks.
The gap widens for the young. Moving from the least to the most exposed group raises the predicted unemployment rate for workers aged 23 to 30 by about two percentage points, and for workers aged 45 to 54 by at most one. The authors read this as consistent with younger workers competing for entry-level positions, where AI tools may substitute fastest for routine cognitive tasks. For those outside trying to get in, they write, “the door appears to be harder to open”.
The authors call their assessment an early and incomplete look at a long adjustment. Their exposure scores mostly measure the potential for AI to substitute for tasks, and the one measure built on actual AI use shows a weaker pattern. These findings are an association. They are not causal proof of displacement.
The Australian note by Mike Major reports little evidence to date that AI has materially disrupted the Australian labour market, with tentative signs of weaker outcomes in more AI-exposed occupations, particularly for younger workers. It finds little evidence of different job-finding rates across exposure groups in Australia. It describes international evidence as showing some signs of weaker employment and job-finding rates, rather than higher layoffs, for entry-level or younger workers in AI-exposed occupations. It warns that exposure measures capture potential task exposure, not realised adoption or whether firms are changing hiring decisions, and that any signs must be read alongside cyclical conditions and other structural change.
RAND’s published summary of a dissertation by Jonah Kushner adds a mechanism from hospital interviews. AI automates simpler tasks through which new medical coders have traditionally learned to code, and the author says this threatens an already weak training pipeline. The same summary reports that occupations with higher AI exposure had higher employment growth from 2005 to 2023, a reminder that exposure does not settle direction.
A role that is never posted has no owner by default
Cleveland’s reading is that firms appear to be waiting to see how AI reshapes their workflows before committing to new headcount, using attrition and reduced hiring to limit labor-cost exposure in occupations where AI may soon substitute for human work. Nothing in that sequence forces anyone to put a name on the choice. A headcount dashboard counts open requisitions, a workforce analytics model scores occupations for AI exposure, and an automated screening tool filters applicants. When those systems check and report on each other, that is Responsible AI, which is machine checking machine. They report what happened, and none of them records who decided that a junior role would go unfilled because AI covers its tasks.
AI Governance asks that question. In Basil’s Checkpoint-Based Governance paper, a named human decides at a checkpoint, and every checkpoint closes on accept, modify, or reject. His earlier column on the on-ramp problem argued that the augmentation thesis assumes the on-ramp away and that measurement has to reach the scale of the person. A hiring plan that leans on AI to absorb junior work is that problem inside a single company.
The decision record to write
Treat any decision to not backfill, cut, or redesign an entry-level role on the basis of AI as a recorded decision with one named owner who accepts, modifies, or rejects it. The record carries four items:
- The tasks AI is assumed to cover, and the evidence for that assumption, such as measured output on the team’s own work.
- Where the skills that role used to build will now be learned, in a named training path.
- The owner’s decision and its date: accept, modify, or reject.
- A review date, with a comparison of actual AI performance and team capacity against the assumption.
Three measures show whether the practice holds:
- The share of entry-level requisition changes attributed to AI that carry a named owner and recorded reasons. The target is every such change.
- Entry-level hires as a share of total hires, tracked against the period before the AI deployment.
- The share of documented training paths that have actually brought a person to the level the absorbed tasks used to build, which separates a written plan from a working one.
What stays open
The central bank findings do not show that AI caused any particular hiring decision. They show a pattern worth recording. The open question is whether a recorded decision changes outcomes: whether owners who must write down the tasks AI covers end up keeping more entry-level roles, or building training paths that work, compared with teams that never had to say. A company that cannot show who decided an entry-level role would stay empty has no way to learn whether the decision was right. This column is general commentary, not legal or employment advice.
Sources
- Cline, A., & Kaymak, B. (2026, November 9). Job finding and unemployment in the age of AI (Economic Commentary 2026-26). Federal Reserve Bank of Cleveland. https://www.clevelandfed.org/publications/economic-commentary/2026/ec-202626-job-finding-and-unemployment-in-the-age-of-ai
- Cline, A., & Kaymak, B. (2026, November 9). Job finding and unemployment in the age of AI [PDF] (Economic Commentary 2026-26). Federal Reserve Bank of Cleveland. https://www.clevelandfed.org/-/media/project/clevelandfedtenant/clevelandfedsite/publications/economic-commentary/2026/ec-202626/ec-202626-job-finding-and-unemployment-in-the-age-of-ai.pdf
- Major, M. (2026, November 9). Monitoring early effects of AI on the labour market (Insight 2026-09). Reserve Bank of Australia. https://www.rba.gov.au/publications/insights/australian-economy/2026/monitoring-early-effects-of-ai-on-the-labour-market.html
- Kushner, J. (2026, November 9). Three essays on artificial intelligence and work in the U.S. (RGSD-A5507-1). RAND Corporation. https://www.rand.org/pubs/rgs_dissertations/RGSDA5507-1.html
FAQ
What did the Cleveland Fed find about AI and unemployment?
The commentary finds that since late 2022 unemployment has risen more in occupations whose tasks are more exposed to AI, especially among young workers. The authors trace this mainly to lower job-finding rates, and they do not find consistent evidence of higher job separation rates in those occupations.
Does the Australian evidence say the same thing?
Not for Australia itself. The Reserve Bank of Australia reports little evidence to date that AI has materially disrupted the Australian labour market and little evidence of different job-finding rates across exposure groups. It sees tentative signs of weaker outcomes for younger workers in more exposed occupations, and it describes international evidence as pointing at job finding rather than layoffs.
Does this prove AI is displacing workers?
No. The Cleveland authors call their assessment an early and incomplete look, and their exposure scores measure the potential for AI to substitute for tasks. The Reserve Bank of Australia warns that exposure measures do not show realised adoption and that cyclical and other structural factors may be at work.
What is the difference between Responsible AI and AI Governance in hiring?
Responsible AI is machine checking machine. Headcount dashboards, exposure scores, and automated screening can count, score, and filter, but none of them records who decided to leave an entry-level role unfilled. AI Governance puts a named human at a checkpoint to accept, modify, or reject that decision.
What should the decision record contain?
It names one owner and states the tasks AI is assumed to cover, with evidence. It names where the skills that role built will now be learned, records the owner’s accept, modify, or reject decision, and sets a review date against actual AI performance and team capacity.
Why does the training path matter?
RAND’s summary of Jonah Kushner’s dissertation describes AI automating the simpler tasks through which new medical coders traditionally learned to code, which the author says threatens an already weak training pipeline. If the tasks that trained new staff disappear, the skills they built have to be learned somewhere else, and the record should say where.
AI disclosure: This column was researched and drafted with AI assistance by HAIA Agents (RAI by GrokBot), using primary sources verified at time of publication. #AIgenerated
Leave a Reply