The Jobs Most Exposed to AI Demand the Most Reasoning: A 737-Occupation Analysis

How tight is that pattern? We measured it. IQ Career Lab is a cognitive assessment platform that studies how intelligence maps to careers and earnings. For this analysis we joined our own reasoning-demand index, the one behind our occupational wage dataset, to the peer-reviewed AI Occupational Exposure index. The merged table covers 737 matched U.S. occupations. The result: the reasoning a job demands tracks about half of the variance in its AI exposure [IQ Career Lab analysis, 2026]. Every number below can be rebuilt from public data.

The Staircase Nobody's Career Advice Predicted

Sort all 737 jobs into five equal groups by reasoning demand, then ask what share of each group is more AI-exposed than the median American job. The answer climbs like a staircase. It never dips [IQ Career Lab analysis, 2026].

The same pattern holds however you cut it. The 20 most reasoning-intensive jobs in America sit, on average, at the 84th percentile of AI exposure [IQ Career Lab analysis, 2026]. That list is physicists, mathematicians, actuaries, and their neighbors. The top reasoning quintile averages +0.97 standard deviations of exposure; the bottom averages -0.98 [IQ Career Lab analysis, 2026]. The gap is almost two full standard deviations. Narrow the lens to language models alone, using the same team's ChatGPT-era update (Felten et al., 2023), and the correlation is still r = 0.65 [IQ Career Lab analysis, 2026].

This is the opposite of how most people still model AI risk. The old intuition, that machines eat the bottom of the skill ladder, was trained on a century of automation that began with the 1913 assembly line. But language-model AI flipped it. The abilities it reproduces best, reading, writing, inference, and structured judgment, are the same ones that reasoning-heavy careers lean on hardest.

How We Built the Number

Our reasoning-demand score is the one we published with the reasoning-and-wage dataset: the mean of three ONET 29.1 ability levels, Inductive Reasoning, Deductive Reasoning, and Mathematical Reasoning, on the 0-to-7 ONET scale, joined to BLS wage and employment data for May 2024.

The AI-exposure side is not ours, by design. The AI Occupational Exposure index (AIOE) links 52 occupational abilities to ten AI applications and scores all 774 SOC occupations on a standardized scale (Felten et al., 2021). It is peer-reviewed, public, and widely used in labor economics, which means our result can be checked end to end by anyone.

The join required one technical step: AIOE is published on 2010 occupation codes, so we mapped it to 2018 codes with the federal crosswalk, averaging the sixteen occupations where old codes merged. That matched 737 of our 739 occupations [IQ Career Lab analysis, 2026].

The headline correlation of r = 0.71 held up in every check we ran [IQ Career Lab analysis, 2026]. Restricting to the 721 occupations with a clean one-to-one crosswalk moves it to 0.712. Weighting each occupation by employment, so big jobs count more, gives 0.702. Each of the three reasoning abilities correlates with exposure on its own: deductive at 0.71, inductive at 0.67, mathematical at 0.63. Whichever slice you take, the conclusion survives. Reasoning demand and AI exposure are two views of the same terrain.

Does the Ranking Pass a Face-Validity Check?

Yes, at both ends. The jobs that score highest on reasoning demand carry exposure far above the national average [IQ Career Lab analysis, 2026]. The jobs at the bottom of the reasoning scale are the economy's least exposed [IQ Career Lab analysis, 2026].

The extremes also match the published record. Felten and colleagues reported that their most-exposed occupations were white-collar roles requiring advanced degrees (Felten et al., 2021). Genetic counselors, financial examiners, and actuaries topped their list; fitness trainers and plasterers' helpers anchored the bottom. Our merged table reproduces that ordering from independent inputs [IQ Career Lab analysis, 2026].

The Entire Shelter Has Nine Rooms

Here is the finding we did not expect to be this stark. Take the top fifth of occupations by reasoning demand, 148 jobs. Ask how many of them sit below the median for AI exposure, the closest thing the data offers to "demands serious thinking, but AI barely overlaps with it."

Nine. Not ninety, nine [IQ Career Lab analysis, 2026].

And the list explains itself. Every single one pairs heavy reasoning with hands, bodies, or physical space: the diagnosis happens in a mouth, a cockpit, or a field, not in a document. If you are hunting for cognitively demanding work where AI exposure stays low, the menu is short, and most of it runs through hands-on fields such as procedural medicine and aviation.

Exposure Is Not a Pink Slip

The task-level work points the same direction. Researchers at OpenAI estimated that about 80% of U.S. workers could see at least 10% of their tasks affected (Eloundou et al., 2024). Roughly 19% could see half or more of their tasks affected (Eloundou et al., 2024). And exposure rose with wages, not against them (Eloundou et al., 2024).

So what happens inside exposed jobs? The early field evidence points to augmentation. In a randomized experiment on professional writing tasks, ChatGPT cut completion time by 40% and raised output quality by 18% [Noy and Zhang, 2023]. Among 5,172 customer-support agents, an AI assistant lifted issues resolved per hour by 15% (Brynjolfsson et al., 2025). The gains were largest for the least experienced agents (Brynjolfsson et al., 2025). Denmark is the sharpest test: chatbot adoption there is among the world's highest, yet matched employer-employee records through 2024 contain no detectable effect of chatbots on earnings or hours [Humlum and Vestergaard, 2025]. The confidence intervals rule out effects larger than 1% [Humlum and Vestergaard, 2025].

That is not a promise of safety. Early-career workers in the most exposed occupations have seen relative employment declines around 16% since late 2022 (Brynjolfsson et al., 2025). That warning comes from the same team's "Canaries in the Coal Mine" paper, and it lands where AI automates tasks instead of augmenting them. The honest summary: exposure marks where work is changing fastest. Incumbents with strong reasoning skills are being amplified so far, but the entry-level rungs of the ladder are creaking. We covered the hiring-side consequences in how AI is reshaping hiring in 2026.

Exposure Does Not Move Pay, Reasoning Does

One more result from the merged table. AI exposure correlates with occupational pay at r = 0.52 against log median wage [IQ Career Lab analysis, 2026]. That sounds like exposure commands a premium. It does not survive a control. Reasoning demand alone explains 60.4% of log-wage variance across these 737 occupations [IQ Career Lab analysis, 2026]. Adding AI exposure to the model nudges that figure to 60.6% [IQ Career Lab analysis, 2026]. Exposure predicts pay only because it travels with reasoning demand. In the 2024 wage data, the labor market prices the thinking, not the exposure, a pattern consistent with what cognitive demand does to occupational wages.

The usual caveat applies with full force. These are occupation-level correlations, not personal guarantees. The gap between job-level and individual-level effects is the difference between a 50% and a 4% figure (Strenze, 2007). No index can tell you whether you, personally, will thrive in an exposed occupation.

What This Means If You Are Choosing a Career

The "find a job AI can't touch" strategy has a math problem. The untouchable jobs that also demand serious thinking number just nine of the 737, and most of them require a decade of clinical or flight training. For everyone else the real question is not whether your work overlaps with AI. It is whether you bring the reasoning that exposed work rewards. In every field study above, the workers who turned AI overlap into an advantage were doing the judgment-heavy parts the tools cannot: framing the problem, checking the output, deciding what matters. That is fluid reasoning, and unlike task knowledge, it transfers across occupations.

Knowing where your reasoning stands, rather than guessing, is the single most useful data point this analysis can't give you.

How to Cite and Reproduce

Cite this as the IQ Career Lab AI-exposure and reasoning-demand analysis, 2026. It is built from the IQ Career Lab reasoning-demand index (ONET 29.1 Abilities plus BLS OEWS May 2024) and the Felten-Raj-Seamans AI Occupational Exposure indices, the 2021 AIOE and its 2023 language-modeling update, joined on 2018 SOC codes via the federal crosswalk. The JSON file carries the exact statistics: r = 0.71 (95% CI 0.68 to 0.74), quintile means, the nine-occupation low-exposure list, and all robustness runs. To reproduce, average the three named ONET ability levels per occupation, crosswalk the published AIOE scores to 2018 SOC, join, and correlate. Two occupations (Data Scientists and Therapists, All Other) have no 2010-code ancestor and are excluded.

Every input is public. If the number is wrong, it can be caught, and that is the point.