AI is not yet removing jobs at scale. It is changing who gets hired first.
In short
The latest US evidence does not show a wave of layoffs in occupations most exposed to AI. It does show an earlier signal: weaker entry into roles whose tasks LLMs are already performing in practice. For e-commerce, the conclusion is not “replace the team with AI”. It is: before adding headcount to a repetitive digital process, check whether the process itself should be redesigned.
Contents
- What Anthropic actually measures
- No layoff wave yet
- The first signal appears in hiring
- Newer Census data strengthen the signal
- Which roles are most exposed
- What this means for e-commerce
- Five processes to inspect before the next hire
- What these data do not prove
1. What Anthropic actually measures
In March 2026, Anthropic published a US labour-market analysis built around a new measure called observed exposure. It combines the theoretical ability of an LLM to accelerate a task, real-world Claude usage, and whether that usage is work-related and automated rather than merely augmentative.
Exposure is not displacement. It means that some tasks within a role are already being performed with AI in ways that could affect demand for labour.
2. No layoff wave yet
Using Current Population Survey data, Anthropic found no systematic increase in unemployment among workers in the occupations most exposed to AI since late 2022. The estimated difference versus workers with little or no observed exposure was statistically indistinguishable from zero.
That does not mean AI is having no labour-market effect. It suggests that the first measurable change may appear somewhere other than mass redundancies.
3. The first signal appears in hiring
Among workers aged 22–25, Anthropic observed a divergence from around 2024. The monthly job-finding rate in occupations with no observed exposure remained around 2%, while entry into the most exposed occupations fell by roughly 0.5 percentage points.
The average estimate for the post-ChatGPT period is a 14% lower job-finding rate for 22–25-year-olds entering highly exposed occupations relative to 2022. The authors stress that the result is only just statistically significant, and they do not find the same decline among workers over 25.
14% does not mean “14% fewer jobs”. It is an estimate of the decline in the job-finding rate for young workers entering highly exposed occupations. It should not be presented as a general fall in employment.
4. Newer Census data strengthen the signal
In September 2026, the Center for Economic Studies at the U.S. Census Bureau published a working paper using administrative records for college graduates. In the most AI-exposed decile of college majors, the likelihood of initial employment fell by 5 percentage points, while initial full-quarter earnings fell by 13% after generative AI became widely available.
The authors attribute roughly half of the earnings decline to lower earnings within the sectors employing these graduates. The rest comes from a shift in the industry mix towards lower-wage sectors such as restaurants and retail. The effects weaken as graduates move further from labour-market entry, but remain substantial for the most exposed majors.
This still does not prove that AI is the sole cause. It is, however, an independent signal consistent with the idea that generative AI's labour-market effects may show up first at the point of entry.
5. Which roles are most exposed
Anthropic's measure places digital, repeatable and machine-readable work near the top. Computer Programmers have about 75% task coverage. Customer Service Representatives, Data Entry Keyers and Financial Analysts also rank among the most exposed roles.
At the other end, 30% of workers have zero coverage in this specific measure. Examples include cooks, motorcycle mechanics, lifeguards, bartenders and dishwashers. That does not mean those jobs are “safe from AI”. It means their tasks appeared too infrequently in Claude usage data to meet the threshold for observed coverage.
| Work pattern | LLM exposure | E-commerce example |
|---|---|---|
| Text + rules + structured inputs | High | product copy, classification, first-line CS drafts |
| Analysis across digital sources | Increasing | reporting, alerts, channel monitoring |
| Judgement with business context | Partial | pricing decisions, operational exceptions, priorities |
| Physical and on-site work | Low in this measure | warehouse work, packing, physical service |
6. What this means for e-commerce
Many e-commerce processes have exactly the characteristics associated with higher exposure: they are digital, repetitive, rule-based and performed across large numbers of similar records. The practical question is therefore not “how many people will AI replace?”. A better question is: which tasks should we stop scaling by adding more human hours?
This is the difference between cutting a team and redesigning an operating model. AI can take on preparation, classification, monitoring and first drafts, while people retain control of exceptions, decisions and accountability.
7. Five processes to inspect before the next hire
- Listing production. Generate and adapt descriptions, attributes and channel-specific content from a single product-data source.
- Product-data enrichment. Classify products, map categories, detect gaps and normalise data before publication.
- First-line Customer Service. Classify tickets, retrieve order context and prepare responses for approval.
- Channel monitoring. Detect feed errors, status changes, anomalies and exceptions instead of manually checking dashboards.
- Reporting and decision preparation. Gather data, explain variances and prepare the material an operator uses to make the decision.
In each case, the objective should not be “AI instead of a person”. The objective is to remove manual work that does not require human accountability, experience or situational judgement.
8. What these data do not prove
Anthropic's study concerns the US labour market and uses Claude usage data. It is not a direct measure of the Polish or wider European labour market, and it does not capture all AI usage. Observed exposure measures task coverage, not the probability that a job will disappear.
The Census study concerns US college graduates and is a working paper. It should be read as current research evidence, not as settled scientific consensus.
For an e-commerce operator, the useful signal is simpler: before hiring another person into a process that scales linearly with SKU count, channels or tickets, check whether the real constraint is the process design itself.
AI Automation at Pleroma Commerce
We design automation around real e-commerce operations: product data, marketplaces, Customer Service, monitoring and reporting. We map exceptions and accountability first, then automate.
Sources
- Anthropic, Labor market impacts of AI: A new measure and early evidence, 5 March 2026.
- U.S. Census Bureau, Center for Economic Studies, Graduating into Disruption: Labor Market Outcomes for AI-Exposed College Majors, September 2026.
- Anthropic, What 81,000 people told us about AI economics, 22 April 2026.