Blog · Research · 14 September 2026 · 13 min read

Apollo's AI Wage Study Says 6.7%. At the Cutoffs Either Side, Its Own Robustness Check Says 1.89% and 1.74%.

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 Sunday 13 September 2026, CNBC ran a feature on AI and pay. The first study it cites comes from Apollo Global Management: workers in occupations highly exposed to AI saw real-wage growth 6.7 percentage points slower after 2023 than workers in less-exposed occupations, with no statistically significant effect on employment.

CNBC reports that correctly, caveats included. Nothing below is a complaint about the coverage.

This post is about what the same PDF prints further on. Its robustness table re-runs the wage estimate with the exposure line drawn a little lower and a little higher, and gets 1.89% and 1.74%. The paragraph written about that table says 6.7% "sits between these estimates". And the abstract's most striking figure, a 24.3% decline for service workers, rests on a subsample the authors flag themselves.

What CNBC published, and what it got right

The article is "The wages of American workers are under pressure. AI's potential role is drawing more attention", by Trevor Laurence Jockims, published at 9:00 AM EDT on 13 September 2026. The Apollo finding, in CNBC's words:

"Their research found that workers in occupations classified as highly exposed to AI experienced real-wage growth that was 6.7 percentage points slower after 2023 than workers in less-exposed occupations. At the same time, the study found no statistically significant effect on employment."

That matches the paper's results section, which says workers in high-exposure occupations "saw their wages grow about 6.7 percentage points slower than workers not in high-exposure occupations after 2023". CNBC also carries the paper's scale caveat, that "only 321 of roughly 800 BLS occupations could be used, and only 11 met the study's high-exposure threshold", and reports that Ben Zipperer of the Economic Policy Institute "was quick to add that the Apollo study had too small of a sample size to be convincing."

Every figure checked against the paper matches. What the article does not go into is the paper's robustness table or its breakdown by occupation group. Both are public, in the same PDF.

Where the 6.7% comes from

The paper is The Impact of AI on the U.S. Labor Market: Early Evidence from Observed Adoption, by Sania Edlich and Torsten Slok, dated July 2026. It was released on 30 July 2026 through The Daily Spark, the daily data note from Slok, listed there as Apollo's Partner and Chief Economist, on the wealth section of Apollo's website. Apollo describes itself as "a high-growth, global alternative asset manager".

The exposure measure is not Apollo's own. The paper uses the Anthropic Economic Index, which it describes as observed usage, "actual Claude interaction logs", with each occupation scored by "what percent of its tasks have been observed being performed with Anthropic's AI tools". Its limitations section is direct about what that implies: "The exposure measure is specific to Anthropic, the only AI provider to have released usage data for public research, meaning AI exposure per occupation is likely understated."

The design is a difference-in-differences regression across 321 occupations matched between Bureau of Labor Statistics wage data and the index, from 2015 to 2025, with 2023 to 2025 as the post-period. An occupation counts as high exposure at a score of 0.5 or above, a threshold the paper takes from PwC, meaning "at least half of an occupation's tasks (weighted by importance) have been observed being done with AI assistance". Eleven occupations clear it.

The paper puts 5.8 million workers in those eleven, about 3.7% of the labor force. Its Figure 8 breaks that share down by occupation, and one occupation dominates it.

Customer service representatives are 1.73% of workers on the paper's figures, which on its own is nearly half of the 3.7% (the division is ours). Add non-technical wholesale and manufacturing sales representatives, at 0.70%, and two occupations make up about two-thirds. The other nine are between 0.02% and 0.33% each.

The robustness table

The paper tests whether its result depends on where the exposure line is drawn: "To confirm that these findings are not sensitive to the choice of exposure threshold, we conduct a robustness check re-estimating the main specifications at cutoffs of 0.4 and 0.6". The results are Table 6, on page 18 of the PDF. Below, the paper's three columns are laid out as rows, in the paper's order, with every value as printed.

Each column has 3,296 observations, with occupation and year fixed effects and standard errors clustered by occupation. The legend: * p<0.1, ** p<0.05, *** p<0.01. The paragraph the paper writes about this table, quoted whole:

"Table 6 presents the check for robustness with different cutoffs for the definition of a ‘high-exposure’ occupation. For the wage outcomes, the negative effect holds across all three cutoffs, with the coefficient remaining statistically significant at the 0.4 threshold (-1.89%, p≤0.05) and negative but insignificant at the 0.6 threshold (-1.74%), likely reflecting the very small number of occupations above that level. The baseline result of -6.7% at the 0.5 cutoff sits between these estimates, consistent with a gradient in which stricter definitions of high exposure capture a more intensely affected group. For employment, no cutoff yields a statistically significant effect, confirming that the null employment result is not a result of the chosen threshold. Together, these results suggest that the core findings are robust to reasonable variation in the exposure cutoff."

Put the three numbers in order of cutoff. At 0.4 the estimate is 1.89%. At 0.5 it is 6.7%. At 0.6 it is 1.74%. The first and last are 0.15 apart, and 6.7 is not between them: it is about three and a half times the larger of the two (this ratio and the others in this section are our arithmetic from the table).

The gradient the sentence describes fits the first step and not the second. Moving from 0.4 to 0.5, a stricter definition, the estimated decline gets larger. Moving from 0.5 to 0.6, stricter again, it gets smaller, and ends up the smallest of the three. That is the opposite of "stricter definitions of high exposure capture a more intensely affected group".

The paper puts the 0.6 result down to "the very small number of occupations above that level". By Appendix B's printed scores that is seven occupations, against eleven at the baseline (our count). The four that drop out are database architects (0.58), financial and investment analysts (0.57), software quality assurance analysts and testers (0.52) and statistical assistants (0.51). The printed standard error at 0.6 is 0.0143, smaller than the baseline's 0.0181, so in the table itself the lost significance comes from a smaller coefficient, not a wider standard error. With no occupation-level output published, this post cannot say how much of the change those four account for.

At 0.4, Appendix B puts 33 to 35 occupations on or above the line, depending on how two scores printed as 0.40 were rounded (our count), and the estimate is 1.89%, under a third of the baseline.

One smaller mismatch sits in the same paragraph. The text gives the 0.4 estimate as "p≤0.05", while Table 6 gives it three stars, p<0.01. The table's own standard error sides with the text: the coefficient is about 2.4 times its standard error (our division), which on the usual normal approximation is significant at 5% but not at 1%.

None of this says the direction is wrong. Every wage coefficient in the table is negative, and the paper's other headline result, no significant employment effect, holds at all three cutoffs. What the table does not support is the paragraph's closing line, that "the core findings are robust to reasonable variation in the exposure cutoff", if the size of the wage effect counts as a core finding. Moving the line by 0.1 in either direction takes the estimate from 6.7% to 1.89% or 1.74%, on the paper's own readings.

A quartile sentence that does not match its table

The same gap between text and table appears on page 10. The results text says workers in the bottom wage quartile saw real wage growth decline "by around 10.7% relative to low-exposure occupations, compared to 5.4% in the second quartile and 4.0% in the third quartile." Here is that sentence beside Table 4 and Figure 4, which report the same breakdown.

The bottom-quartile figure checks out in all three places, and so does the top quartile's absence of an effect. The third quartile's 4.0% and 4.1% are the same coefficient, -0.0405, rounded two ways. The second quartile does not check out. Table 4 prints -0.0570 with no significance marker, and Figure 4 draws that dot in the grey its legend labels "Not significant". The only -5.4% in Figure 4 is the label under the panel for occupation size.

The 24.3% for service workers

The abstract's most quotable line is that "service workers face a 24.3% decline and the bottom wage quartile a 10.7% decline, while top earners show no significant effect." CNBC did not use the service figure. The results section tells readers to handle it carefully:

"The service occupation estimate (-24.3%) is based on a small subsample (n=239 occupation-year observations) and may be sensitive to the composition of high-exposure occupations within that group. This result should be interpreted with caution."

That caution is on page 10, not in the abstract. Here is what else the PDF lets a reader check.

Size first. The sample covers the eleven years from 2015 to 2025, so 239 occupation-year observations means at least 22 occupations in the service group (our arithmetic). Table 5 says only "Occupation groups defined by 2-digit SOC code". The paper does not publish which occupations it put in which group.

Then the examples. The results text describes service occupations "such as childcare workers, concierges, waiters, police officers, social workers, and others". Appendix B, which lists all 321 occupations with their scores, puts every one of those examples it contains well below the 0.5 line: childcare workers 0.01, concierges 0.19, police and sheriff's patrol officers 0.12, child, family and school social workers 0.01, healthcare social workers 0.09. Waiters do not appear among the 321. None of the named examples is a high-exposure occupation in the paper's own terms.

Then the treated occupations. The paper re-runs its regression separately for each subgroup, and a difference-in-differences estimate needs at least one treated occupation inside the group, or there is nothing to estimate (our reasoning about the method, not a statement in the paper). So at least one of the eleven is in the service group. The paper does not say which.

What can be said is this. In the federal Standard Occupational Classification, the high-level aggregation the Bureau of Labor Statistics recommends groups major groups 31 to 39 as "Service Occupations". Of the eleven occupations in Appendix B, exactly one has a code in that range: medical transcriptionists, 31-9094, with a score of 0.64 and 0.03% of workers in Figure 8. The other ten carry codes beginning 13, 15, 29, 41 and 43. Computer programmers are printed in the appendix as "-1251"; the federal code is 15-1251.

Whether the paper follows that aggregation, it does not say, and its own example list is a reason not to assume so. Social workers carry codes beginning 21 in Appendix B, 21-1021 and 21-1022, and the same BLS aggregation places major group 21 with management, business, science and arts occupations rather than with service. So this post cannot tell you which occupations sit behind the 24.3%. Neither does the paper, which says only that the number "may be sensitive to the composition of high-exposure occupations within that group."

The other study in the same article

CNBC sets Apollo beside an analysis from the Federal Reserve Bank of Dallas by Scott Davis, published 24 February 2026, and says the two reached different conclusions. Across 205 occupations, Davis finds the trend line between AI exposure and wage growth since fall 2022 "nearly perfectly horizontal". Once experience is taken into account, the picture splits. Where experienced workers earn no premium over entry-level pay, "increased AI exposure is associated with a 0.28 percentage point reduction in wage growth"; at the 90th percentile of that premium, with "a 0.2 percentage point increase in wage growth".

The only exposure index the article names is one developed by Edward W. Felten, Manav Raj and Robert Seamans, which Apollo's own literature review classes as theoretical, and the article never mentions usage data. The windows and wage variables differ as well, so neither study can be used to check the other's number.

What this means if you work in one of these jobs

If your occupation is one of the eleven, the direction is the part of the paper that holds up across its own checks. The wage estimate is negative at every cutoff it tested, though not significant at the strictest, and employment shows no significant change at any of them. The size is the part that moves: 6.7% at the line the paper chose, 1.89% and 1.74% at the lines either side.

If you are a childcare worker, a concierge, a police officer or a social worker, the paper names your occupation as an example of its service group, and the matching occupations in its appendix all score as low exposure. The 24.3% is an estimate for high-exposure occupations within that group, and those are not among them.

If you are not sure where your job sits, Appendix B lists all 321 occupations with their scores. Keep in mind what the score is: the share of an occupation's tasks observed being done with one company's AI, which the paper itself says likely understates exposure. A low score is not a guarantee, and a high one is not a verdict. The Dallas Fed result suggests a second question to ask about your own job: how much of your pay rests on experience a new hire does not have.

What this post does not claim

It does not claim that AI is not affecting wages, or that the paper is wrong about the direction of its result. Every wage estimate in Table 6 is negative. It does not claim CNBC misreported anything: the figures it used match the paper. It does not say which occupations the paper placed in its service group, because the paper does not publish that. And it does not explain why the estimate moves so much between cutoffs, because the paper does not publish the occupation-level results that would show it.

The Apollo PDF and its landing page, the CNBC article, an Apollo press release, the Dallas Fed article and two BLS classification pages were fetched and read on 14 September 2026, and every quotation above was checked against the document it came from.