These Are the 30% of Jobs AI Can’t Reach in 2026 (According to Anthropic’s Own Data)

Everyone is talking about what AI is coming for. Fewer people are talking about what it genuinely cannot touch.

In March 2026, Anthropic, the company behind Claude, published a landmark research paper titled “Labor Market Impacts of AI: A New Measure and Early Evidence.” Most headlines zeroed in on the scary numbers: computer programmers at 75% task coverage, customer service reps at 70%, data entry workers at 67%.

But buried in that same report is a finding that deserves far more attention.

Roughly 30% of American workers have zero measured AI exposure. Zero. Not low exposure. Not partial exposure. Their jobs don’t even register in Anthropic’s framework because the tasks involved show up too rarely in how people actually use AI tools at work.

That’s not a rounding error. That’s tens of millions of jobs that AI hasn’t meaningfully touched — and based on how those jobs are structured, may not touch for a very long time.

Here’s what that 30% actually looks like, why it matters for your career decisions, and how to use this data to your advantage right now.

What Anthropic Actually Measured (And Why It’s Different)

Most AI-and-jobs research takes a theoretical approach. Researchers ask: “Could AI theoretically do this task faster?” If yes, the job gets flagged as exposed.

Anthropic did something more useful. Their economists looked at what people actually use Claude for in professional settings, then weighted fully automated uses more heavily than tasks where humans are still involved. They called this “observed exposure” — and it produces very different results than the theoretical model.

The gap is significant: AI is far from reaching its theoretical capability, with actual coverage remaining a fraction of what’s feasible.

Computer and math jobs show 94% theoretical exposure but only 33% observed. Office and admin sits at 90% theoretical and 25% observed. Business and finance shows 85% theoretical and 20% observed.

What this tells you: even in the fields most associated with AI disruption, actual real-world automation is a fraction of what’s technically possible. The gap between “AI could do this” and “AI is doing this” is enormous.

For workers in the truly low-exposure categories, that gap doesn’t exist. Their jobs barely appear in AI usage data at all.

Here’s what most people don’t realize: employers now expect multiple technical competencies, not just one specialization. The days of being “just a marketer” or “just an analyst” are over. You need AI skills, project management, data literacy, and more. Building that skill stack one $49 course at a time is expensive and slow. That’s why unlimited access makes sense:

The 30%: Who Actually Has Zero AI Exposure

Anthropic’s data found that 30% of workers have zero coverage, as their tasks appeared too infrequently in the data to meet the minimum threshold. This group includes cooks, motorcycle mechanics, lifeguards, bartenders, dishwashers, and dressing room attendants.

Look at what those jobs have in common. They all require:

• Physical presence in unpredictable environments
• Real-time judgment calls that change by the minute
• Direct human interaction that context can’t be fully scripted
• Manual dexterity and situational awareness

An LLM cannot flip a burger, rewire a circuit panel, or pull a drowning swimmer out of a pool. These aren’t limitations that better prompting can fix. They’re fundamental constraints of what language models actually are.

Trades like construction and plumbing remain resilient because they require physical manipulation of unpredictable environments and manual dexterity that digital models cannot replicate.

This is the core insight hiding inside Anthropic’s data: AI’s primary limitation isn’t intelligence. It’s embodiment. The moment a job requires a physical body operating in the real world, the calculus changes completely.

The Occupational Categories With the Lowest Observed Exposure

Beyond the zero-exposure group, several broad occupational categories show dramatically lower AI penetration than the white-collar fields getting all the attention. If you’re evaluating where to invest your career development energy, these are the fields worth understanding.

Skilled Trades and Construction

Construction, installation, and repair work sits at the bottom of Anthropic’s exposure charts. Anthropic determined that AI’s labor market impacts show limited evidence that AI has affected employment in trades to date.

The work itself explains why. An electrician diagnosing a wiring problem in an older home is operating in an environment full of variables: outdated blueprints, non-standard materials, code changes over decades, and physical conditions that vary room by room. No AI system can send a robot to handle that job reliably in 2026.

Check out our breakdown of the highest paying trade jobs for 2026 if you want specifics on what these careers actually pay — the numbers might surprise you.

Some of the lowest-exposure skilled trade occupations include:

• Electricians — diagnosing, installing, and maintaining electrical systems
• Plumbers and pipefitters — working with water, gas, and drainage systems in unpredictable conditions
• HVAC technicians — installing and repairing heating and cooling systems
• Welders and fabricators — precision physical work requiring tactile feedback
• Construction managers — coordinating crews, materials, and real-world conditions on job sites
• Auto mechanics and diesel technicians — physical diagnostics and repair work

Healthcare and Direct Patient Care

The least exposed occupations tend to require physical abilities. This shows up clearly in healthcare. Administrative healthcare roles like medical record specialists are highly exposed (67% in Anthropic’s data). But roles involving direct patient care tell a very different story.

Nursing, physical therapy, occupational therapy, and emergency medicine all require the kind of embodied judgment that AI genuinely cannot replicate. A physical therapist working with a post-surgical patient is reading body language, adjusting pressure in real time, building trust through human connection, and making split-second decisions based on physical feedback. That’s not a workflow you can automate.

The healthcare hiring boom is real and ongoing. The Bureau of Labor Statistics projects strong growth across direct care roles through 2034, even as AI-exposed occupations in the same sector face slower growth.

Low-exposure healthcare roles worth considering:

• Registered nurses — especially in procedural and emergency settings
• Physical and occupational therapists — hands-on rehabilitation work
• EMTs and paramedics — emergency response requiring real-time physical intervention
• Surgical technologists — direct involvement in operating room procedures
• Dental hygienists and dental assistants — physical patient care and procedural work
• Home health aides and caregivers — personal care requiring human presence and judgment

Food Service and Hospitality

Cooks and chefs appear explicitly in Anthropic’s zero-exposure examples, and the reasoning holds across most food preparation and service roles. The physicality of cooking — temperature management, texture evaluation, timing coordination across multiple dishes, adapting to equipment quirks in real time — involves sensory feedback that language models can’t replicate.

Fine dining and complex culinary work adds human creativity and aesthetic judgment on top of the physical challenge. A sous chef improvising during a busy service isn’t following a script. They’re making dozens of judgment calls per minute based on sensory input that no AI has access to.

Personal Care and Appearance Services

Hairstylists, barbers, estheticians, and massage therapists all make Anthropic’s low-exposure list for similar reasons. The work is:

• Touch-dependent — requires physical contact and tactile feedback
• Client-specific — every session involves reading a unique individual’s needs and preferences
• Environment-dependent — adapting to tools, space, and conditions that vary constantly

These roles also depend heavily on the social and emotional dimensions of service. People develop loyal relationships with their stylist or massage therapist. That relationship is part of what they’re paying for.

Emergency and Protective Services

Lifeguards appear explicitly in Anthropic’s zero-exposure examples. Firefighters, police officers, and emergency responders belong in the same category. The work involves unpredictable physical environments, life-and-death time pressure, and the kind of situational awareness that can’t be scripted.

These roles also carry legal and ethical dimensions that make pure automation a non-starter regardless of technical capability.

The Counterintuitive Finding About Who’s Actually at Risk

Here’s the piece of Anthropic’s research that most people missed entirely.

The workers in the highest-exposure group are more likely to be older, female, more educated, and higher paid. They also earn 47% more on average than the unexposed group, while workers with graduate degrees are much more concentrated in the exposed bucket. That is a useful correction to the lazy narrative that AI risk is mainly about low-skill work.

This flips the conventional wisdom completely. The narrative has been that low-skill, low-wage workers face the most disruption from automation. Anthropic’s actual data says the opposite: it’s the educated, higher-paid, often female-dominated white-collar workforce that shows the highest observed AI exposure right now.

The truly protected group? Workers in physical, lower-credential roles that the media has always characterized as precarious.

For career planning purposes, this matters enormously. The instinct to “stay in white-collar work because it’s safer” is not supported by Anthropic’s data. The skills-based hiring playbook is increasingly rewarding demonstrated physical and interpersonal competence over credentials alone.

What the “Gap” Means for Your Timeline

One of the most useful findings in Anthropic’s research is the gap between theoretical AI capability and observed exposure — and what it implies about timing.

Why might actual usage fall short of theoretical capability? Some tasks that are theoretically possible may not show up in usage because of model limitations. Others may be slow to diffuse due to legal constraints, specific software requirements, human verification steps, or other hurdles.

For jobs in the zero-exposure category, these hurdles aren’t temporary software gaps. They’re physical reality. You can’t patch your way to a robot that installs plumbing under a house or provides wound care to a post-surgical patient.

The big question is how fast the gap between observed AI exposure and theoretical exposure will close. It will vary a lot between different professions. The idea that the same levels of automation that hit software developers is about to hit every other knowledge worker in the next 12 to 18 months seems off. It’s going to take substantially longer.

This gives workers in low-exposure fields a realistic planning horizon — not just “you’re safe for now,” but a structural advantage that’s embedded in the nature of the work itself.

The Hiring Signal That Matters Most

Occupations with higher observed AI exposure are projected by the BLS to grow less through 2034. The research also found that hiring of younger workers has slowed in those same high-exposure fields.

Specifically, Anthropic’s data shows a 14% drop in the job-finding rate for workers in exposed occupations among the 22-to-25 age group since late 2022. That’s not mass unemployment. But it’s a signal worth taking seriously if you’re early in your career and choosing a field.

The hiring slowdown in exposed fields isn’t happening because companies are firing people. It’s happening because they’re not backfilling roles when people leave. AI is doing the work of the next hire before that hire ever gets made.

In low-exposure fields, that dynamic doesn’t exist. Trade jobs that pay well are experiencing genuine labor shortages because demand for the work is growing while the supply of qualified workers isn’t keeping pace.

How to Use This Data in Your Career Planning

Knowing that a field has low AI exposure is useful. Knowing how to act on that information is more useful.

If you’re early in your career:

• Skilled trades, direct patient care, and physical service roles offer structural protection that white-collar entry-level positions increasingly don’t
• The credential-to-income ratio in trades is often better than it appears — apprenticeships and licensure programs cost far less than four-year degrees
• AI tools can still help you be more effective within a low-exposure role (scheduling, client communication, administrative tasks) without threatening the core work

If you’re mid-career and evaluating a change:

• Moving toward more physical, human-centric work is a defensible long-term strategy, not a step backward
• Hybrid roles that combine low-exposure physical work with AI-augmented coordination and planning are particularly well-positioned
• The human skills that AI can’t replicate translate across industries — developing them now pays dividends regardless of where you land

If you’re advising someone entering the workforce:

• The framing that “trade jobs are for people who couldn’t do something else” is both wrong and increasingly expensive to believe
• Anthropic’s data makes the career case that most guidance counselors haven’t caught up to yet
• Check the top 25 jobs that won’t get eliminated by AI for a broader list of roles with structural protection

The Honest Caveat

None of this means low-exposure jobs are completely immune to change. AI will continue reshaping how work gets scheduled, documented, quoted, and managed, even in fields where the physical labor itself remains human.

An electrician in 2030 will probably use AI tools to pull permits faster, generate quotes, and troubleshoot diagnostic codes. A nurse will use AI for documentation and pattern recognition in patient data. The physical, judgment-heavy core of the work stays human. The administrative wrapper around it gets more efficient.

That’s actually the best-case outcome for workers in these fields: more of their time on the work they’re trained for, less on paperwork.

Anthropic found no systematic increase in unemployment for highly exposed workers since late 2022, and the rapid diffusion of AI is generating a wave of research measuring and forecasting its impacts on labor markets. For workers in the zero-exposure category, there’s no unemployment signal to find at all.

The Bottom Line

Anthropic built their research to find out where AI is actually making inroads — not where it theoretically could. When their own data shows 30% of workers with zero measured exposure, that’s not a consolation prize. That’s a real structural feature of the labor market that belongs in every career planning conversation happening right now.

The jobs AI can’t touch share a common thread: they require a human body, operating in the physical world, making real-time judgments that vary with every client, every environment, and every day. No language model has a body. That constraint isn’t going away.

If you’re investing in a career, investing in skills, or advising someone who is, Anthropic just gave you the most reliable map available of where that investment is protected. The skills-based hiring playbook and the AI skills you need for 2026 both have a place in modern career planning. But for long-term structural security, the 30% is worth understanding on its own terms.

The physical world is still hiring. And right now, it’s one of the few places AI genuinely hasn’t been able to follow.

Here’s what most job seekers miss: recruiters Google your name before they ever schedule an interview. Having a great LinkedIn profile isn’t enough anymore. A personal website proves you can do the work, not just claim it…

Sources:

• Anthropic: Labor Market Impacts of AI — March 2026
• Fortune: How AI Will Reshape Work
• Euronews: How AI Will Reshape Work — Anthropic Report
• Axios: Anthropic Launches AI Job Destruction Detector

---