AI Job Loss: The Next Search Takes Longer
A strong jobs month can hide a longer wait for the person whose role disappeared. The Goldman figure is a likelihood score, not a count of AI layoffs.
Christian Kosiek•Published •11 min read
A national employment report answers a national question. The person whose job just disappeared needs a different one: how long until a stable next job, and will that job use the skills already built? New research from Goldman Sachs suggests the wait itself deserves as much attention as the headline total.
We have already separated tasks that change inside a job from a job that disappears, and a headcount cut from proof that AI replaced each role. This article is the next question. Once the role is gone, how long is the return?
A payroll total does not answer that person
On 30 October 2026, Fortune reported a Goldman Sachs client note by economist Pierfrancesco Mei. Workers displaced from AI-exposed jobs since 2025 took about a month longer to find a new job than other displaced workers. They were also about 10 percentage points more likely to still be unemployed a year later. Mei described those effects as comparable in magnitude to earlier waves of technological change, and called the damage from those earlier waves long-lasting.1
Axios, on 3 November, reported the same order of result for workers most likely displaced by AI in 2025: roughly a month longer than other laid-off workers, and roughly 10 percentage points more likely to remain unemployed a year after the loss.2 The two accounts do not use identical wording for the window (since 2025, or in 2025). They describe the same research note. Neither is a government count of people whose employer demonstrably replaced them with AI.
The age split is part of the same reporting. Workers aged 35 and older took about five weeks longer than other displaced workers to get back to full-time work. Workers under 35 took about three weeks longer.12 Mei wrote that younger workers may be able to adapt more flexibly to AI disruption. Fortune also reported that Goldman found no significant gap between the scored groups for job losses before 2023.1
| Comparison, as reported | Figure | What it is not |
|---|---|---|
| Likely AI-displaced since 2025, versus other displaced workers | About one month longer to a new job | A census of layoffs caused by AI |
| Still unemployed a year later | About 10 percentage points more likely | Proof that AI was the employer's stated reason |
| Aged 35 and older, back to full-time work | About five weeks longer | A result for every AI-exposed worker |
| Under 35, back to full-time work | About three weeks longer | Evidence that younger workers were unharmed |
| Found work in 2025 after an AI-exposed job | About three-quarters moved to similar or greater AI exposure | A guarantee the next role is safer |
The last row is from Axios. Among displaced workers from AI-exposed jobs who found new work in 2025, roughly three-quarters landed in occupations with similar or greater AI exposure, up from just over half in 2023.2 A new job can put the person back in the same kind of exposure.
The longer wait is a score, not a census
The Current Population Survey’s Displaced Worker Supplement does not ask whether AI caused the loss. Fortune reports that Goldman analyzed more than 15,000 displaced workers who lost jobs between 2016 and 2025, including the new 2023 to 2025 supplement. Goldman built a score from the occupation’s AI exposure, the industry’s AI adoption rate and whether the position was eliminated, then treated the top 30% as likely AI-displaced.1 Axios describes the same ingredients: occupation, industry adoption and the circumstances of the loss, including whether the position was eliminated.2
That method is a reason to take the result seriously and a reason to refuse a slogan. "Likely AI-displaced" means a high score on a constructed measure. It does not mean an investigator reviewed each redundancy letter. Fortune says Goldman still considers it too early to assess the impact along education and geography.1 The client note is not a public document. The figures above are press accounts of it.
The historical comparison is a different paper, and it should stay labelled as history. Goldman Sachs Research published "How Automation Impacts the Lives and Careers of Displaced Workers" on 11 September 2026, not in October. It looks at earlier waves of automation. Technology-displaced workers took roughly one month longer to find a new job than people in stable or expanding occupations, and they suffered three percentage points higher real earnings losses on reemployment when moving between full-time jobs. A decade on, their real earnings were still lower than those of comparable workers who lost jobs in occupations that were not disrupted. Workers who spent at least four weeks in retraining within three years of the loss achieved roughly two percentage points more cumulative real wage growth over the next decade than peers who did not retrain, and were less likely to return to unemployment.3
That retraining figure is an association in historical data. It is not a randomized demonstration that a particular course caused the gain. It is still a practical reason to treat transition support as something to measure, rather than as a farewell gesture.
A strong October does not close the gap
The two stories can be true in the same week. On 6 November 2026 the U.S. Bureau of Labor Statistics reported that total nonfarm payroll employment rose by 165,000 in October, higher than the average monthly gain of 48,000 over the prior 12 months. The unemployment rate changed little at 4.1 percent. The number of unemployed people, 7.1 million, also changed little. Employment increased in professional and business services, in health care, and in transportation and warehousing.4
A payroll release describes employment across the economy. Mei’s comparison describes people who have already lost jobs and were scored as likely AI-displaced. A strong aggregate month does not prove that AI created the gains, and it does not disprove a longer search in particular occupations. New jobs in health care or warehousing are not automatically a next step for an accounts clerk or a junior analyst.
Goldman had already separated substitution from augmentation in an April research note. The public article of 24 April 2026 says AI reduced monthly U.S. payroll growth by roughly 16,000 jobs over the past year and raised the unemployment rate by 0.1 percentage point, while jobs with augmentation potential added about 9,000 to monthly payroll growth. Elsie Peng writes that the aggregate impact was likely smaller than those estimates, because they do not fully capture hiring for data-center construction or extra demand from productivity gains.5 Fortune’s 6 April account of the same note adds the split the public page does not print: substitution of about 25,000 jobs a month, with augmentation adding about 9,000 back.6 Both pieces say the result comes from a model, not from a census of AI-caused layoffs.
The useful message is the coexistence of loss and new hiring, not the exact residual. A company that automates invoice matching may hire customer advisers or data specialists. Neither vacancy automatically fits the person who knew the exceptions, the clients and the regulatory edge cases. We made the same point when new AI infrastructure jobs turned out not to be a transfer plan.
The overlooked risk is the occupational ladder
A transition is not a software tutorial. An administrative assistant may have judgment about exceptions, confidentiality and customer relationships, while the hiring market now advertises a narrower technical role. A junior analyst may see entry-level drafting disappear before there have been enough chances to learn how to challenge an assumption. An experienced worker may hold domain knowledge and still lack a form of evidence a new employer will accept.
Axios’s reemployment figure belongs here. If about three-quarters of people leaving AI-exposed jobs in 2025 moved into occupations with similar or greater exposure, then "found a job" is not the same as "left the risk behind".2 The September research describes an older version of that slide: technology-displaced workers were more likely to move into work that was more routine and less complex, because the change that cut the job had also cut the market value of the skills.3
The practical response is not a generic instruction to become an AI expert. It is a short map: what the person can already do, what evidence proves it, and what two or three nearby roles actually require.
What a small employer can do first
Consider a hypothetical service organization of about 70 people that introduces AI-assisted document processing. Six coordinators check submissions, correct records, talk to clients and escalate unusual cases. If processing time falls, leadership can reduce staffing, take more demand, shorten turnaround or retrain coordinators for quality checks and client support. The better choice depends on workload, error rates, money and what employees want. This scene is an illustration. It is not a surveyed company.
Measure tasks before you announce a headcount target. For four to six weeks, record volumes, exception rates, rework, escalation time and customer outcomes. Separate time that truly disappeared from work that moved to another team. Faster document preparation is not the same thing as faster case resolution.
Then build a transition map for each affected role. List demonstrated skills: data validation, process knowledge, client communication, regulatory judgment. Compare them with the tasks and the evidence two or three realistic destination roles require. A quality-control role may need sampling and audit notes. A client role may need negotiation and account experience. Write the gaps down. Do not call every gap "AI skills".
Offer paid practice on real work, with supervision that matches the risk. A short prompting workshop can raise confidence. It does not show that someone can review an AI-produced record against a quality rubric, document a recurring error or handle a simulated client exception. Those outputs are a portfolio. A certificate of attendance is not.
Watch whether the new role lasts. Track internal placement, retention at six and twelve months, wage changes, whether the move was a choice, and quality outcomes. Moving someone to a new title for a few weeks is not a completed transfer. If redeployment is impossible, job-search help and time for training are still real support. Say clearly what the employer cannot promise.
If the organization is in the EU and uses an AI system to choose who is dismissed, or to filter who is selected for a remaining role, that purpose can fall under Annex III, point 4, of the AI Act. Point 4(a) covers recruitment and selection, including filtering applications and evaluating candidates. Point 4(b) covers decisions that affect the terms of a work relationship, including promotion or termination, and the monitoring of performance. The Commission’s AI Act Service Desk says the Annex III high-risk rules apply from 2 December 2027.78 That date is not legal advice, and a skills map is not itself that system. It is a reason to keep a human record of what people can do before any tool ranks them.
A course cannot invent demand
Social organizations, workforce boards and local training providers can do what a single small firm often cannot: convene employers around a shared transition, check short demonstrations of skill, and connect people with supervised projects. That matters most for workers who cannot fund months of unpaid retraining or move to another labor market.
A well-designed course still cannot manufacture demand for an occupation that is shrinking. Publish placement rates, wages and the jobs graduates actually get. Judge the course by those outcomes, not by the number of AI modules in the catalog.
How hilighter can help
The decision this research supports is narrower than "will AI create more jobs than it destroys". Which tasks are changing here, who carries the risk, what adjacent work exists, and what would make those employees credible candidates for it?
hilighter helps a team map skills by role, compare people with those requirements in Skills & Talent Insights, keep proof of capability in Skills Evidence & Recognition and set personalized growth paths. The AI Job Impact view is indicative. It is not Mei’s score, and it is not a prediction that a role will disappear. A free plan covers up to three job roles, with unlimited employees.
Sources
- Fortune, Nick Lichtenberg, Goldman says AI layoffs are starting to look long-lasting, 30 October 2026. Account of a Goldman Sachs client note by Pierfrancesco Mei. The note itself is not public.
- Axios, Courtenay Brown, AI job displacement is looking extra painful, 3 November 2026.
- Goldman Sachs Research, How Automation Impacts the Lives and Careers of Displaced Workers, 11 September 2026. Historical automation, including the retraining comparison. Read 9 November 2026.
- U.S. Bureau of Labor Statistics, The Employment Situation, October 2026, news release USDL-26-1632, 6 November 2026. Payrolls rose by 165,000; the unemployment rate changed little at 4.1 percent.
- Goldman Sachs Research, The Jobs AI Is Likely to Boost, and Those It May Disrupt, 24 April 2026. Net payroll estimate and the augmentation figure. Read 9 November 2026. The official headline uses an em dash; this citation does not.
- Fortune, Nick Lichtenberg, AI is cutting 16,000 U.S. jobs a month, Goldman Sachs says, 6 April 2026. The substitution figure of about 25,000 is in this account of Elsie Peng’s note.
- European Commission, AI Act Service Desk, Annex III, point 4. Read 9 November 2026.
- European Commission, AI Act Service Desk, When does enforcement start?. Annex III high-risk rules apply from 2 December 2027. Read 9 November 2026.
Sources last reviewed: 9 November 2026. The October duration figures are press accounts of a client note, not a public microdata release. The April payroll figures are a model. The 70-person organization is illustrative. This is not legal advice.