Blog · Research · 26 July 2026 · 8 min read

Even Amazon's AI Team Isn't Safe. Here's What Actually Predicts Exposure

AI-generated. This article was written by an AI system and published without substantive human review. Facts are drawn from primary sources and linked inline — verify anything you intend to act on.EU AI Act Art. 50(4)

On 22 July 2026, Amazon confirmed layoffs inside its Artificial General Intelligence organization — the unit building the company's own frontier AI — affecting teams under vice presidents Adeeb Shanaa and Vishal Sharma, perCNBC. Amazon has not disclosed how many people were affected. Those in the US were offered 90 days of pay and benefits, outplacement support, transitional healthcare, and severance eligibility.

The most repeated piece of career advice of the last three years is that getting good at AI keeps you safe. That advice just failed its own stress test. The people cut here were not adjacent to AI, or AI-curious, or AI-upskilled. They were building it.

So if proximity to AI is not the variable, what is? The research has an answer, and it is more specific — and more checkable against your own calendar — than the headlines suggest.

What the payroll data actually shows

The strongest evidence comes from Stanford's Digital Economy Lab. In"Canaries in the Coal Mine?"(Brynjolfsson, Chandar and Chen, 13 November 2025), the authors analysed high-frequency ADP administrative payroll records — actual pay data, not survey self-reports — and found a16% relative decline in employment for workers aged 22 to 25 in the most AI-exposed occupations, after controlling for firm-level shocks. Employment for more experienced workers in the same occupations remained stable or continued to grow.

The finding that matters most is not the 16%. It is the shape underneath it. The authors report that declines are "concentrated in occupations where AI is more likely to automate, rather than augment, human labor," and that adjustment happened"primarily through employment rather than compensation" — headcount moved, pay did not. Two roles with identical titles and identical salaries can sit on opposite sides of that line depending on what the work actually consists of.

That automate-versus-augment split is not a metaphor. It comes fromAnthropic's Economic Index(10 February 2025), which classified roughly a million anonymised Claude conversations against the US Department of Labor's O*NET taxonomy of about 20,000 work tasks. It found42.6% of usage was automation — the model performing the task — against57.4% augmentation, the model collaborating with a person. Stanford borrowed that classification to sort occupations. Which is also the first thing to be honest about.

Why we are not treating that as settled

A study is only as good as the limits its readers know about, so here are this one's.

It is correlational, and the authors say so. The paper documents an association between AI-exposure measures and employment outcomes. It does not claim AI caused the decline, and neither do we. The effect also only reaches statistical significance once full controls are applied from 2024 onward — earlier apparent declines could reflect interest-rate-driven hiring slowdowns, post-pandemic remote-work reversion, or a tech hiring slump that predates the current generative-AI wave.

The data has coverage limits. ADP payroll records avoid self-report bias, which is a real advantage, but they only cover employers who use ADP: skewed toward mid-size and larger US employers, and likely undercounting gig work, very small businesses, and government.

And the exposure classification is borrowed, not independently derived. It rests on Claude usage being a reasonable proxy for generative-AI usage generally. Anthropic's ownJune 2026 reportstates plainly that its survey base "is not representative of the general population" — Computer and Mathematical occupations are 30% of respondents against roughly 4% of actual US employment. Any skew in Claude's user base can ripple into Stanford's classification.

None of that makes the finding wrong. It makes it a strong signal rather than a verdict, and anyone selling it to you as a verdict is overselling.

The gap between announcements and causes

Here is where most coverage of this topic goes wrong, and it is worth separating carefully.

Gallup surveyed 23,717 employed US adults in February 2026 (margin of error ±0.9 points). Among workers who had actually been laid off,1% named AI or automation as the primary cause. The reasons they gave instead were organisational restructuring (15%), budget and cost cutting (11%), and economic conditions (11%). Meanwhile a large share of corporate layoff announcements now mention AI.

Both things can be true, because they measure different objects. "Companies cite AI when explaining a layoff" is a statement about corporate communication. "AI caused this person's job loss" is a statement about causation. Treating the first as evidence for the second is the single most common error in this entire discourse.

The same Gallup release contains the figure most often quoted as proof that AI skills protect you: tech workers using AI less than monthly had an 18% predicted layoff riskagainst 6% for those using it at least monthly. Three times the risk. We would not read that as protection. Workers who adopt new tools early tend to be adaptable, engaged, and in more resilient roles — characteristics already associated with lower layoff risk long before AI arrived. The correlation is real. The mechanism is unproven.

The split that does hold up

The most useful framing we found this year comes fromPwC's 2026 Global AI Jobs Barometer(15 June 2026), built on more than a billion job advertisements across 27 countries. It separates two things that get bundled together as "AI is changing this job."

Same technology, opposite consequences, decided by whether the expertise stays load-bearing. PwC also put the average wage premium for AI skills at 62%, up from 57% the prior year — the market is paying more for the combination, not less.

The entry-level picture in that data deserves its own line. Across 2.4 million US entry-level postings, PwC found AI-exposed entry-level roles areseven times more likely to require traditionally senior-level skills — judgment, leadership, creativity, face-to-face interaction. Those "seniorised" entry rolesgrew 35% since 2019, while other entry-level roles shrank 10%. The bottom rung is not vanishing. It is being raised, which is a harder problem for the people standing under it.

What to actually do with this

Three things follow, and none of them is "learn AI."

Audit your week by task, not by title. Take the last ten working days and split the work into two piles: tasks a competent model could complete end-to-end given your context, and tasks where it could only draft while you carry the judgment and the accountability. The ratio is your actual exposure. It will not match your job title, and it is the only number here that is about you specifically.

Move work across the line deliberately. The professionalised pattern is not luck. It is what happens when the expert remains the one who decides, and the tool absorbs the throughput. If your role is drifting toward being a fast producer of model-shaped output, that is the democratised track, and it pays worse over time.

Stop reading announcements as data. When a company says it is cutting roles because of AI, that is a sentence written by a communications team. When 1% of laid-off workers say AI cost them their job, that is a measurement. Calibrate to the second.

Amazon's AGI team was as close to the frontier as it gets, and closeness did not help. What the evidence keeps pointing at instead is duller and more actionable: the structure of the work itself. That is measurable, it is partly within your control, and unlike your industry, it does not require anyone's permission to change.