How Exposed Is Your Job? A Framework for Reading Your Own Risk

Career Landscape Reports: 3 of 5 in a series

The first two pieces in this series made two claims that sit in tension with each other. The first showed that AI making a task cheaper tends to expand the total amount of work in a field, not shrink it, so the aggregate story is more optimistic than most doom-focused coverage suggests. The second showed that the traditional path into that expanded work, the junior role where you learned judgment by doing routine tasks, is closing for a specific and identifiable group of people even as the field around them grows. Both of those things are true at the same time, and neither one answers the question most people actually want answered when they read either piece: not what’s happening to software developers or consultants as a category, but what’s happening to their own job.

That question deserves a better answer than the “AI-proof careers” listicles currently flooding search results, most of which rank occupations by a single automation-risk score with no explanation of what the score measures or how reliably it predicts anything. Before getting to a framework worth actually using, it’s worth understanding why those rankings are less trustworthy than they look.

Why Most Exposure Rankings Overstate the Risk

Anthropic’s Economic Index, which tracks how its Claude models are actually being used across the economy by analyzing millions of real conversations against the Department of Labor’s O*NET occupational task catalog, makes a distinction that most public rankings collapse into one number: theoretical exposure versus observed exposure. Theoretical exposure asks what a model is technically capable of doing, given a task description. Observed exposure asks how often people are actually using AI to do that task in practice, right now.

The gap between those two numbers is large and consistent. Across occupational categories, theoretical exposure runs 50 to 65 percentage points higher than observed exposure, and the two are correlated but far from identical: theoretical capability explains roughly 85% of the variation in real-world use, which leaves a meaningful share of the picture explained by something other than what the technology can technically do. Computer and mathematical occupations, for instance, show roughly 94% theoretical exposure but only about a third of that in observed use. The task is capable of being automated. It isn’t necessarily being automated yet, and the reasons why turn out to be more informative than the raw capability score.

That gap is also why a huge share of the economy still shows essentially no measurable AI usage at all. Of the roughly 18,000 individual work tasks the Index tracks, over 92% show zero measurable use in the data, and 54% of the occupations tracked show no observed exposure whatsoever, concentrated almost entirely in physical, hands-on work: cooks, mechanics, lifeguards, bartenders, skilled trades. If you’ve seen a headline ranking your job’s “automation risk” at 90-something percent, it’s very likely reporting theoretical capability, not what’s actually happening to people doing that job today. That distinction matters enormously for how much weight you should put on any single number.

Four Things That Actually Predict the Gap

Once you set aside single-number rankings, four factors show up consistently across the research explaining why some technically automatable work stays firmly in human hands while other work gets absorbed quickly. The first three are properties of the role itself. The fourth is different: it’s a property of the specific person doing it, and it’s the one most exposure rankings miss entirely.

Whether the work requires a physical presence. This is the cleanest predictor in the data and the one the Anthropic Index confirms most directly: the entire zero-exposure category is dominated by work that happens in a specific physical location, doing something with your hands, in an environment that isn’t fully predictable. A language model can draft a contract or summarize a meeting from anywhere. It cannot yet fix a furnace, plate a dish, or physically restrain someone in crisis. This factor is durable but not permanent. Robotics is a different technology on a different timeline than large language models, and physical presence protects a role from the current wave of AI disruption without protecting it from automation indefinitely.

How costly an error actually is. Thomson Reuters’ 2026 Future of Professionals report captures this directly: firms in law, tax, and audit are adopting AI tools quickly for drafting and research, but adoption slows sharply wherever the output shapes a legal judgment, a regulatory filing, or formal client advice, because in the words of the report’s authors, “almost right” isn’t good enough when real liability is on the line. This shows up beyond the professions with formal licensing requirements, too. Broader research on enterprise AI adoption finds the same pattern: AI moves fastest into modular, document-based work with fast feedback loops, and moves slowest into roles requiring implicit knowledge, high-stakes decisions, and heavy regulatory oversight, precisely because mistakes in those roles carry consequences that are hard to reverse. Insurance data reflects the same shift from the other direction: AI-related errors have become a leading driver of professional indemnity claims, which is the market’s way of pricing exactly how costly those mistakes are turning out to be.

Whether the labor market for that specific role is already tight. This factor gets the least attention in most exposure rankings, and it matters more than people assume. Registered nursing faces a documented national shortage of well over 200,000 workers. Skilled trades face a projected shortfall over two million workers by 2030. In roles facing that kind of structural shortage, employers are frequently adopting AI and automation tools to extend the productivity of the workers they already have, rather than to replace headcount they can’t fill anyway, because the constraint isn’t cost, it’s finding a body to do the job at all. A bipartisan policy analysis of coming workforce gaps put this distinction explicitly: automation offers the clearest upside in high-skilled, white-collar fields with ample labor supply, like legal and financial services, precisely because those fields don’t face the same scarcity pressure that protects nursing or the trades regardless of how automatable individual tasks might be on paper.

How deep and hard-to-transfer your own accumulated judgment is, independent of your job title. This factor sits underneath the first three rather than alongside them, and it’s the direct link between this piece and the second one in this series, which looked at how the traditional path to building that judgment is narrowing. Economists studying this pattern have started calling it seniority-biased technological change: a technology that substitutes for entry-level execution while complementing expert judgment, rather than treating all workers in an occupation as equally exposed. The evidence for it is specific and consistent. Research by Hosseini and Lichtinger found that firms adopting generative AI reduced junior employment by 9% relative to non-adopting firms, while senior employment at the same firms was essentially unaffected. Separate research by Brynjolfsson and colleagues found a 16% relative decline in early-career employment across highly AI-exposed occupations, again with no corresponding decline among senior workers in those same occupations. Two independent research teams, looking at different data, found the same shape of result: the job title and the technical exposure score can be identical for a junior and a senior person in the same field, and the actual outcomes still diverge sharply along seniority lines. The reason, according to economic research on how expertise gets built and passed on, is that a meaningful share of what senior people know is tacit rather than codified: judgment built from years of pattern recognition across unusual cases, the kind of knowledge that resists being written down and is normally transferred only through direct mentorship, not through a document AI could be trained on. That’s a property of the person and the specific depth of what they’ve accumulated, not of the occupation they hold.

Why These Factors Interact Rather Than Simply Add Up

None of these factors works in isolation, and a role can score protectively on one and exposed on another at the same time, which is exactly why single-number rankings flatten something that isn’t actually flat.

Software development is the clearest example, and it connects directly back to the first piece in this series. It requires no physical presence and the stakes of an individual error are often low and quickly correctable, both traits that should predict high exposure. And by observed-use measures, software-adjacent work does show meaningfully high AI usage. But the labor market for people who can design systems, not just write code to a specification, remains structurally different from the market for narrower programming tasks, which is the entire distinction the Jevons Paradox piece in this series was built around. High technical exposure on the first three factors didn’t translate into employment decline for experienced developers, because the fourth factor, the depth of judgment that separates a system designer from someone executing a specification, did the real protective work.

Anthropic’s labor-market data adds a useful confirming detail here: workers in highly exposed occupations currently earn roughly 47% more on average than workers in zero-exposure occupations, and are nearly four times more likely to hold a graduate degree. High exposure and high value aren’t opposites. They’re frequently the same role, viewed from two different angles, which is a more useful way to think about your own position than a single risk score that treats “exposed” as a synonym for “expendable.”

The same research adds a sharper, more targeted number that connects directly to the second piece in this series. Looking specifically at workers aged 22 to 25 in highly exposed occupations, the study’s authors found the job-finding rate has fallen roughly 14% compared to 2022, before the current wave of generative AI tools existed, with no equivalent decline for workers over 25. The authors themselves flag the result as only barely statistically significant, which is worth taking seriously rather than treating as a dramatic headline number. But it’s a specific, age-targeted early signal sitting inside a broader picture that otherwise shows no detectable rise in unemployment for exposed workers as a group, and it lines up with exactly the entry-level narrowing the Narrowing Ramp piece in this series documented from an entirely different dataset.

Applying This to Your Own Role

Rather than looking up your job title in a ranked list, it’s more useful to sit with four plain questions about the actual work you do day to day.

Does a meaningful part of your job require you to be physically present, doing something with your hands or your body, in a way that can’t be fully scripted in advance? If so, that part of your role is genuinely protected for now, though worth revisiting periodically as robotics and physical AI systems mature on their own, separate timeline.

If your output is wrong, who bears the cost, and how hard is that cost to undo? A drafting error caught in review is a different category of risk than a regulatory filing, a medical judgment, or a client-facing recommendation that’s already been acted on. The more your work sits in the second category, the more the human judgment layered around any AI-generated first draft is likely to remain the actual job, not a temporary bridge until the technology catches up.

And is your specific role, not your industry in general, one where employers are currently struggling to find enough qualified people? A tight labor market for your exact skill set changes the calculus employers are making even when the underlying tasks are technically automatable, because the alternative to automating isn’t necessarily human labor at the old price. Sometimes it’s simply no labor at all.

And how much of what makes you good at your job would be genuinely hard to hand to someone else, even someone with your same title and years of experience on paper? The answer isn’t your job title or your tenure. It’s the specific, accumulated pattern recognition you’d struggle to fully explain to a replacement even if you tried, the kind of judgment that’s normally passed on only through direct, sustained contact with someone who already has it. The deeper that judgment runs, the more it functions like a fourth, personal layer of protection layered on top of whatever the other three factors say about your role in general.

None of these questions produces a single reassuring number, and that’s the point. The honest answer for most real jobs is “partially protected, on some but not all of these dimensions, and worth checking again in a year,” which is a less satisfying conclusion than a ranked list but a considerably more accurate one.

Where This Leaves the Series So Far

Put the first three pieces together and a coherent picture starts to form. The first showed that AI expanding the total work available in a field is a real pattern, not wishful thinking. The second showed that the traditional entry point into that expanding work is narrowing for people early in their careers. This piece adds the tool for figuring out where your own role actually sits between those two stories, rather than assuming either the optimistic or the pessimistic version applies to you by default. The next piece in this series takes on a different problem entirely: even once you know where you stand, the ground has a habit of moving. Companies that cut aggressively on AI have reversed course within months more than once already, and knowing your exposure level doesn’t tell you how to make a decision while that volatility is still playing out.

---

Sources

This piece drew on the following research and reporting:

• Maxim Massenkoff and Peter McCrory, “Labor Market Impacts of AI: A New Measure and Early Evidence,” Anthropic, March 5, 2026. Primary data on theoretical versus observed AI task exposure, the zero-exposure occupational category, and wage and education patterns by exposure level.
• Thomson Reuters, “2026 Future of Professionals Report,” 2026, featuring commentary from CEO Steve Hasker on liability standards in legal, tax, and audit AI adoption.
• Workera, “AI Adoption Will Remain Uneven in 2026. Here’s Why, and How to Fix It,” 2026, by Taylor Sullivan. Analysis of why AI adoption is faster in modular, low-stakes work and slower in high-stakes, regulated roles.
• Insurance Edge, “AI Errors Could Be Professional Indemnity Risk,” August 2026.
• Bipartisan Policy Center, “New Report Addresses U.S. Workforce Shortages Over the Next Decade,” 2025.
• Lighthouse, “Worker Shortage: Causes, Impacts, and Solutions in 2026,” 2026. Data on registered nursing and skilled trades labor shortfalls.
• Voxbooster, “Labor Shortage Statistics 2026,” 2026. Additional labor shortage figures by sector.
• Enrique Ide, “Automation, AI, and the Intergenerational Transmission of Knowledge,” IESE Business School, working paper, 2026. Source for the “seniority-biased technological change” concept and the cited findings from Hosseini and Lichtinger (2025) and Brynjolfsson et al. (2025a) on junior versus senior employment effects at AI-adopting firms.

I use AI tools to help research sources and draft these posts. I review, edit, and take responsibility for everything published here, including any errors.