Every week brings a new claim about how much of the workforce AI is about to touch. Eighty percent of jobs. Whole professions gone within a decade. Before accepting either number, ask who benefits from you believing it, and whether the data actually supports it.

It mostly doesn't. The real divide isn't between industries, it's between work that sells scarce access to information, which AI is dismantling fast, and work that sells physical skill, liability-bearing judgment, or trust, which it mostly can't touch. Some of that second group is simply insulated. Some is exposed at the task level but protected at the job level. A smaller, more specific group is genuinely exposed both ways, and that's the group the loudest headlines obscure rather than clarify.

Two kinds of business, and only one of them is actually threatened

Some businesses exist mainly because information or a skill was hard to access. A corporate lawyer charging a high fee to draft a routine contract isn't only charging for legal knowledge, now cheap and widely available, they're charging because most people couldn't produce a competent first draft themselves. That part of the business is genuinely exposed.

The rest of the profession isn't, and it's worth being precise about why. A lawyer is also selling accountability: if a contract fails, you can pursue them for malpractice. You can't do that with a chatbot. That's a structural, legal moat, not an informational one. The exposed part of law is routine, templated drafting; the protected part is judgment under real liability, negotiation, novel deal structure, litigation.

Litigators and expert witnesses go further than protected, they're genuinely enhanced: courtroom advocacy, cross-examination, and testimony under oath depend on a real-time, embodied skill with no substitute. AI absorbing discovery and document review just clears more calendar for the skill that was always theirs.

The same split shows up everywhere. A high-end design studio isn't mainly selling access to design knowledge, anyone can find design principles online. It's selling taste, a track record, and a trusted relationship, much harder to commoditize, which is why "creative work is safe" needs a caveat: execution-only, stock-template work is already being squeezed. What survives is curation and trust, not creativity in the abstract.

The real dividing line isn't industry, it's whether a job's paid-for value comes from scarce access to information, which AI erodes fast, or from physical skill, liability-bearing judgment, taste, or trust, which it mostly doesn't touch.

The data everyone skips

The conversation about AI and jobs is dominated by professions that are a small share of actual employment. The Bureau of Labor Statistics (BLS) tracks 22 major occupational groups covering the entire workforce, and sorting the whole thing by the same logic, physical work, liability-bearing work, and routine moat-free work, tells a very different story than cherry-picking a few headline professions.

Physical, hands-on, or in-person occupations, transportation, food preparation, production, healthcare support, construction, repair, cleaning, protective service, personal care, farming, add up to 43.8% of total U.S. employment. Occupations where liability or credentialed judgment protects the core of the job, management, healthcare practitioners, community and social service, architecture and engineering, science, and legal, add another 18.4%.

Office and administrative support, the clearest match for "routine, language-based work with no liability or physical component," is 11.8% on its own, smaller than either category above it. Legal occupations are well under 1% of total employment; starting the conversation with lawyers means starting with a rounding error, not where the actual workforce is.

The remaining 26.0%, sales, business and financial operations, education, computer and mathematical occupations, and arts and design, genuinely straddles both sides and isn't confidently classifiable at this level of detail, spanning junior, exposed roles and senior, protected ones within the same category. Forcing false precision onto it would be less honest than naming the uncertainty.

One methodology note: the percentages are real BLS data; the tier classification is this piece's own, not BLS's, which is why the straddling 26.0% is reported as unclassified rather than tidied up. The Harvard study on junior-versus-senior employment below is secondhand reporting, not the original paper, worth tracing before repeating further.

What the "80% of jobs are exposed" number really measures

The most widely cited statistic here comes from a 2023 study by Eloundou, Manning, Mishkin, and Rock, done with OpenAI: roughly 80% of the U.S. workforce could see at least 10% of tasks affected by large language models, and about 19% could see 50% or more affected.

Two things matter more than the headline number, and the authors say so themselves. Not weighting how important a task is to the overall job produces "curious results", their own example ranks barbers as highly exposed, not because cutting hair is AI-exposed, but because a peripheral task like scheduling touches language. The core, hard-to-replace skill doesn't change.

A separate ILO-affiliated study had domain experts review task-level scores, finding physical and manual tasks scored lower because generative AI can't handle the physical component, while structured, language-heavy tasks scored high. A mechanic's actual job, diagnosing a fault by hand, isn't language work. An email is incidental to it, not the job itself.

The economists themselves are split, and it's not an even split of credibility

The optimistic forecasts about AI's economic impact tend to come from firms with a direct financial stake in AI adoption.

Goldman Sachs's research (Briggs and Kodnani) projects widespread AI adoption could drive a 7% increase in global GDP, almost $7 trillion, over a decade, with annual labor productivity growth rising about 1.5 percentage points. The report estimates roughly two-thirds of U.S. occupations have some AI exposure, with 25% to 50% of tasks within those occupations potentially automatable, not that half of workers lose their jobs. McKinsey Global Institute puts a related number on it: generative AI could add $2.6 to $4.4 trillion in annual value, translating to roughly 0.2 to 3.3 percentage points of additional productivity growth, a wide range reflecting real uncertainty even within McKinsey's own estimate.

Independent academic modeling tells a different story. Daron Acemoglu, the MIT economist who shared the 2024 Nobel Prize in Economic Sciences, built a task-based model and arrived at a far smaller figure: total U.S. GDP growth of just 1.1% to 1.6% over the next decade, about 0.05% in additional annual productivity growth. He's called the larger industry figures "disconnected from evidence." Christopher Pissarides, the 2010 Nobel laureate, has separately argued a genuine AI productivity surge isn't realistic given the data so far.

Acemoglu has raised a concern worth taking seriously: AI labs hiring their own in-house economists could tilt research output toward more favorable conclusions, a real incentive problem in how these numbers get produced. To be fair, this remains a live, unsettled dispute, not a case of one side simply being wrong, Goldman and McKinsey both hedge their numbers with real uncertainty. But the biggest numbers keep coming from firms that benefit most from people believing them, and the most conservative numbers keep coming from researchers with no such stake.

Even Acemoglu's own account narrows AI's impact to office jobs built around data summary, pattern matching, and routine synthesis: journalism, financial analysis, and HR aren't disappearing, specific tasks are shifting.

A case study that already happened

This exact pattern already played out once, before AI, and it's well documented. Travel agents are the clearest historical precedent, and the outcome wasn't simple destruction.

Employment peaked around 340,000 in 2000, then fell 60% to 70% as airlines cut commissions in the 1990s and platforms like Expedia and Priceline gave consumers direct access to the same booking systems agents once monopolized. That's the routine task, finding and booking a flight, automated away almost entirely. But the profession didn't disappear: economist Ernie Tedeschi's analysis found the agents who remained shifted upmarket, coordinating complex trips for high-value clients, the exact judgment-and-trust work an algorithm can't replicate. Employment fell, and wages for those who stayed rose.

U.S. travel agency employment has rebounded since the pandemic, back to around 78,000 jobs by mid-2024, and the BLS projected 20% growth in demand through 2031, faster than average. The trade isn't just surviving its automated routine task, it's recovering because demand for complex, high-value travel keeps growing as people get wealthier.

That's the whole thesis in miniature: the commoditized task got automated and its headcount collapsed, the trust-driven task survived and paid better, and demand kept the profession alive rather than erasing it.

The mechanism nobody's pricing in: rising demand, not scarcity

Here's the part usually missing: trades and hands-on work aren't just insulated from AI, they're positioned to see real demand growth from it.

When a technology makes production much cheaper, people don't consume less overall, they reallocate what they save. The textbook case is food: agriculture fell from about 40% of the U.S. workforce in 1900 to under 2% today, while food consumption itself rose, because people spent the freed-up money elsewhere.

Economists Diego Comin, Danial Lashkari, and Martí Mestieri studied what drives these long-run spending shifts and found rising real income, not changing relative prices, is the dominant force, accounting for over 75% of historical structural change, versus about a quarter from price effects. What people spend rising income on is disproportionately goods and services with high income elasticity: dining out, travel, home renovation, personal and in-person services.

As AI raises productivity and lowers costs in the sectors it touches, real incomes rise, and a large share of that new spending flows toward exactly the services AI barely touches: hospitality, travel, construction, personal care, skilled trades. That's rising demand, not a shrinking pool of scarce workers. William Baumol's older observation, that sectors which can't get more productive see relative costs and wages climb as everything else gets cheaper, still plays a supporting role, but it's the reinforcing effect, not the main driver. The main driver is that wealthier people simply want more of these things.

Putting it together

A more accurate map of who's actually affected looks like this:

Structurally insulated, and likely to see rising demand — 43.8% of the workforce: skilled trades and physical service work, mechanics, electricians, plumbers, construction, food preparation, hospitality, travel. The value here is physical skill and in-person judgment, which AI doesn't touch, and per the income-elasticity research above, this work should benefit further as AI-driven productivity gains raise real incomes people spend on exactly these services.

Exposed at the task level, protected at the job level — 18.4% of the workforce: doctors, lawyers, managers, engineers, and similar credentialed or liability-bearing professions. Routine drafting, summarization, and first-pass analysis are genuinely exposed; the accountable, liability-bearing core isn't, for structural reasons that have nothing to do with model quality. Some roles go further than protected, litigators and expert witnesses are enhanced: freeing up time from routine work redirects more hours toward the courtroom skill that was always the real differentiator.

Genuinely exposed, task and job both — 11.8% of the workforce: work where the entire output is language generation with no liability, physical, or trust-based moat. Office and administrative support is the clearest match, templated drafting, undifferentiated clerical and content work, commodity-quality summarization. It's the smallest of the three defined tiers, and the group the loudest hype tends to obscure rather than clarify.

Mixed, and honestly not classifiable at this level of detail — 26.0% of the workforce: sales, business and financial operations, education, arts and design, and computer and mathematical occupations. Each spans a genuine range, a junior role doing templated, replaceable work and a senior role doing protected, judgment-heavy work, often within the same job title. Software development is the clearest example: the same category contains both the boilerplate work AI compresses dramatically and the complex, liability-bearing systems work it barely touches. Forcing this group into either extreme would be less honest than naming the uncertainty.

That bifurcation isn't hypothetical for developers, it's measured. Stanford's Digital Economy Lab, using actual ADP payroll data, found developers aged 22 to 25 saw employment decline nearly 20% from its late-2022 peak by July 2025, while developers 30 and older in the same high-exposure roles saw employment grow 6% to 12% over the same period. A separate study of 62 million workers across 285,000 U.S. firms found junior employment at AI-adopting companies dropped 9% to 10% within six quarters, with senior employment essentially unchanged. The routine end of the work is being automated at real scale; the complex end is holding steady or growing, the "mixed" classification made concrete rather than hedged.

What this actually means for you

Which category you're in determines what's actually worth doing about it.

• Genuinely exposed: move toward the parts of your field that carry liability, require in-person trust, or can't be templated, the same shift travel agents made. Staying purely in commodity-quality, language-only output is the one position the data doesn't support as stable.

• Exposed at task, protected at job (doctors, lawyers, and similar): the routine share of your work will keep shrinking, and that's fine, it was never your differentiator. Lean further into the liability-bearing, judgment-heavy work AI can't take on rather than competing on drafting and summarization.

• Mixed (software development, business and financial roles, sales, education, creative work): your exposure depends on where you sit within your field, not your job title. Ask whether your daily work looks more like the routine end or the complex end, and move toward the latter rather than assuming the label protects you.

• Structurally insulated: the case for complacency isn't there, but neither is the case for worry. The data points toward rising demand, so invest in what makes you specifically trusted, credentials, reputation, local relationships, to capture that demand rather than assume it arrives automatically.

• Policymakers: target retraining at the genuinely exposed category specifically, not broad sectoral alarm. The disruption is real but narrower than the loudest forecasts suggest.

None of this supports the story where AI quietly guts most of the economy. It supports a narrower, better-grounded one: businesses that thrive on hard-to-access information are genuinely under pressure, and the loudest voices insisting the disruption is bigger tend to be the ones who benefit financially from you believing it.

Put simply: the wealth AI generates doesn't eliminate demand for human work, it redirects it. As AI takes over the routine, commoditized end of intellectual work, the workforce doesn't shrink to match, it shifts. Physical trades and in-person services aren't the casualties of this transition, they're where demand ends up heading, because that's work AI can't do and rising wealth makes people want more of.

Sources

• U.S. Bureau of Labor Statistics, Occupational Employment and Wage Statistics, May 2025: bls.gov/news.release/ocwage.htm
• BLS, "Major occupational groups as a percentage of total employment," May 2024 (22-category breakdown): bls.gov/oes/2024/may/featured_data.htm
• Stanford Digital Economy Lab (Brynjolfsson et al.), ADP payroll data on developer employment by age, via Stack Overflow: stackoverflow.blog/2025/12/26/ai-vs-gen-z
• Harvard study on junior vs. senior employment at AI-adopting firms (secondhand): medium.com/@emirkanbeyaz01/the-great-tech-hiring-freeze
• BLS, Occupational Employment and Wages, May 2023 (office/admin figures): bls.gov/news.release/archives/ocwage_04032024.pdf
• Eloundou et al., "GPTs are GPTs," OpenAI/OpenResearch/UPenn, 2023: openai.com/index/gpts-are-gpts, full paper: arxiv.org/pdf/2303.10130
• ILO-affiliated paper, "Generative AI and Jobs: A Refined Global Index of Occupational Exposure": webapps.ilo.org/wp140
• Briggs & Kodnani, "The Potentially Large Effects of AI on Economic Growth," Goldman Sachs, 2023: goldmansachs.com/insights
• McKinsey Global Institute, "The economic potential of generative AI," 2023: mckinsey.com
• McKinsey's productivity-growth translation, via co-author Alex Sukharevsky: venturebeat.com
• Acemoglu, "Don't Believe the AI Hype," Project Syndicate, 2024: project-syndicate.org
• On Acemoglu's and Pissarides's projections, via MIT Technology Review and WebProNews: technologyreview.com, webpronews.com
• On Acemoglu's concern re: AI labs' in-house economists: metaintro.com
• Tedeschi, "The decline of travel agents," Stripe Economics: stripeeconomics.com
• BLS Occupational Outlook Handbook, Travel Agents (2021–2031 projection): bls.gov/ooh/sales/travel-agents.htm
• Skift, "The U.S. Travel Agency Rebound Post-Pandemic": skift.com
• Aghion et al., "Artificial Intelligence and Economic Growth," NBER: nber.org
• Comin, Lashkari & Mestieri, "Structural Change With Long-Run Income and Price Effects," Econometrica: dcomin.host.dartmouth.edu
• Imas, "What will be scarce?," applying Comin et al. to automation: aleximas.substack.com