Short answer: there is no job AI can never touch, but there are jobs where the work that matters is much harder to automate than the lists suggest — and the difference is not the job title, it is the task mix. A job is a bundle of tasks, and replacement happens one task at a time: the routine, screen-based tasks go first, while tasks that need a body in the room, a person someone trusts, or a judgment call with incomplete information hold up far longer. This page gives you a way to test any job yourself instead of trusting a list: a three-question audit you can run on your own week, ten families of work where the hard tasks dominate, the jobs people most often get wrong in both directions, and what to do if your own job turns out to be mostly routine. The test itself is free — use it to see which kind of hard work your interests already point toward.
Check your fit — free 4-minute test →
Most lists of jobs AI can't replace work the same way: a title, a one-line reason, and a confident tone. The problem is that the title is not what gets automated. The tasks inside the title are. A radiologist's week is not one thing called "radiology"; it is reading scans, talking to patients, deciding what an ambiguous image means, and writing up findings. Software is already good at the first part and bad at the middle two. A teacher's week is not one thing called "teaching"; it is grading, planning, mentoring, and managing a room full of teenagers. Software is already good at grading and useless at the last one. When a list says a job is safe, it is really saying that the hard tasks inside that job outweigh the easy ones — and when a job disappears, it is usually because the easy tasks were most of the week and the hard tasks were not enough to keep the role whole.
That changes how you should read every list, including this page. The honest question is never "is this job safe?" but "how much of this job's week is made of tasks that resist automation, and how much is routine work that software already does?" A job where 80% of the week is judgment, trust, and hands-on work in changing places is in a different position from a job where 80% of the week is filling in standard forms — even if both jobs sound equally "human" in a headline. The World Economic Forum's Future of Jobs Report 2025 makes the same point at the level of the whole economy: it projects roughly 170 million new roles created and 92 million displaced between 2025 and 2030, a net gain — but the displacement is concentrated in routine, rules-based work, not spread evenly across every title.
So the method on this page is deliberately different from a list. You will audit a job task by task, using three questions. Then you will see the ten families where the hard tasks dominate, with what AI has already absorbed inside each one — because the honest picture includes the erosion, not just the walls. If you want the same question from the claim-testing angle, see AI proof careers; if you want it as a resistance ranking, see careers safe from AI.
Take a real job — yours, or one you are considering — and break a normal week into its tasks. Not the job description's version, which lists the impressive parts, but the actual week: the meetings, the forms, the calls, the driving, the fixing, the waiting. Then run each task through three questions. A task that passes two or three is hard for AI to take; a task that passes none is already exposed, whether or not anyone has automated it yet.
Question 1: does this task have to happen in person, in a place that keeps changing? Software works best where the setting is controlled: the same screen, the same data, the same room. A task that happens in a different building, a different home, a different body, or a different emergency every time is a different problem on every occurrence. An electrician opening a wall never knows what is behind it until the wall is open. A paramedic never knows what the scene looks like until they arrive. If the task could be done from a laptop in any city, it fails this test. If it needs a body, hands, and eyes on site, it passes.
Question 2: does it depend on someone trusting you, or on you answering for the result? Some work only functions if another person believes you are paying attention and care about the outcome: a patient telling a nurse the symptom they left off the form, a client signing because they trust their adviser, a student admitting to a teacher that they are lost. And some work exists because a named person has to stand behind it — a surgeon, a pilot, an engineer who signs off on a bridge. A system can simulate the words of empathy and it can draft the report, but it cannot be the person the patient trusts, and it cannot be the person who answers when the result goes wrong. If trust or accountability is the product, the task passes. If nobody in particular has to answer for it, it fails.
Question 3: does it need a judgment call with incomplete information? The easiest tasks to automate are the ones with a rulebook answer: check the box, apply the formula, file the form. The hardest are the ones where the data runs out before the question does — two plausible diagnoses, a negotiation where both sides are bluffing, a classroom where the right move depends on reading a teenager's mood. Software can surface evidence and recommend a course of action, but the decision, and the responsibility for it, still sits with a person. If the task is mostly judgment, it passes. If it is mostly following a procedure, it fails.
Now count the week. If most of your tasks pass two or three questions, your job leans on real protection. If most pass one or none, the title may sound safe while the work is not — and the honest move is to change the task mix, not to argue with the list. A quick worked example shows the difference. A medical coder's week is mostly reading records and assigning standard codes from a screen: it fails all three questions, and it is already being automated. An electrician's week is mostly diagnosing faults in buildings that are all wired differently, fixing them by hand, and signing off on work that has to be safe: it passes all three, and demand is projected to grow 9% from 2024 to 2034, against 3% for all occupations on average (BLS 2024–34 projections). The two jobs are not in the same world, and the audit tells you why in about ten minutes.
The table below is not a promise about any single career — demand moves, and every family below has routine tasks inside it that software is already absorbing. What the families share is that the hard tasks are most of the week, not a garnish on top of routine work. For each family, the table shows why it holds, what AI has already absorbed inside it, and the usual way in. Check demand for any specific occupation in the Bureau of Labor Statistics Occupational Outlook Handbook before you commit to a direction.
| Family | Why it holds | What AI already does inside it | Usual way in |
|---|---|---|---|
| Nurse practitioners & physician assistants | Hands-on care, judgment under uncertainty, and patient trust — all three audit questions at once. Projected growth is well above the all-occupations average of 3%: nurse practitioners 40% and physician assistants 20% from 2024 to 2034 (BLS). | Documentation, scheduling, intake forms, routine screening, and progress notes | Graduate degree plus license |
| Electricians | Every site is different, the work is physical, and a licensed person signs off on safety. Projected growth 9% from 2024 to 2034 (BLS). | Diagnostics support, parts lookup, route planning, and quoting | Apprenticeship plus license |
| Plumbers | Unpredictable, in-person diagnosis and repair; no two jobs are the same behind the wall. | Dispatch, quoting, and inventory lookup | Apprenticeship plus license |
| Teachers (K–12) | Mentoring, motivation, and classroom judgment depend on a relationship with real students. | Grading, lesson-plan drafting, and routine feedback | Bachelor's degree plus teaching credential |
| Therapists & counselors | The product is trust: clients say things to a person they believe is listening, not to a system. | Scheduling, reminders, intake paperwork, and progress notes | Master's degree plus license |
| HVAC technicians & mechanics | Diagnosis in messy, uncontrolled environments — a machine that has never been serviced the same way twice. | Diagnostics support, parts lookup, and scheduling | Trade school or apprenticeship |
| Chefs & cooks | Taste, timing, and running a kitchen under pressure — judgment that has to happen in the moment, by hand. | Recipe generation, menu costing, and ordering | Culinary training or on-the-job experience |
| Paramedics & firefighters | No scene has a script; the work is physical, urgent, and accountable. | Dispatch support, routing, and reporting | Certification plus academy or field training |
| Social workers | Judgment in ambiguous human situations, where ethics and trust matter more than procedure. | Case-note drafting, scheduling, and form filling | Degree plus license |
| Veterinarians | Diagnosis plus hands-on treatment plus an owner who has to trust you with an animal they love. | Record-keeping, scheduling, and routine lab interpretation | Doctor of Veterinary Medicine degree |
The pattern in the third column is the same everywhere: the absorbed part is the repeatable, screen-based work around the hard core. The hard core stays human — but a job that keeps only the core and loses everything around it can still shrink, with fewer people doing the same high-value work.
Two mistakes show up again and again when people read lists like the one above. The first is trusting a title that sounds safe. Radiologists are the classic example: the job sounds deeply human — a doctor studying images and deciding what they mean — but image pattern-matching is exactly what software does well, and the routine reading is already being automated. The job survives on the judgment and the patient conversations around the image, not on the image itself. Junior legal work has the same shape: document review and first-pass research are routine, screen-based tasks, and they are going fast, while the negotiation and the courtroom judgment hold up. The lesson is not that these careers are doomed; it is that the safe part of the job is smaller than the title suggests, and the audit tells you how small.
The second mistake is the opposite: writing off a job that sounds exposed. Teaching looks exposed from the outside — lectures can be recorded, quizzes can be graded by software — but the part that matters most, keeping a room of teenagers engaged and noticing the one who has gone quiet, is not a screen task at all. Cooking looks exposed too — recipes are easy to generate — but running a kitchen during a rush, tasting, adjusting, and keeping a team moving is judgment under time pressure, in person, by hand. In both cases the list-writer looked at the title and the routine parts and missed the hard core. If you are choosing, run the audit on the actual week, not on your impression of the title. And if you are still mapping your own strengths, how the Holland Code works explains the six interest patterns that point toward hands-on, social, and judgment-heavy work — the patterns the hard tasks tend to need.
If your audit came back thin — most of your week passes one question or none — the move is not to panic and not to hunt for a permanently safe title, because there is not one to find. The move is to change the task mix, in four steps. First, audit your current week honestly and name the tasks that already pass two or three questions; those are your foothold, and most jobs have more of them than their owner thinks. Second, take on more of the hard tasks: volunteer for the judgment calls, the client conversations, the on-site problems — the work that needs trust and a signature — and let the routine parts go to the tools. Third, learn the tools instead of competing with them: the people who stay valuable through a shift are usually the ones who use AI on the routine parts of their work and keep the judgment, the relationships, and the accountability for themselves. Fourth, check the demand side before you commit to a direction — the Bureau of Labor Statistics Occupational Outlook Handbook shows projected growth for specific occupations — and check your own fit with a free test, so you are moving toward work that suits your interests, not just away from a list. If you want a second opinion on your own job, our eight-question check scores the tasks in your current role and tells you which of the three audit questions your job is weakest on.
One last honest note: the audit is a snapshot, not a verdict. Moats move. Tasks that looked hard to automate five years ago are routine now, and the line between "needs a human" and "a human is nice to have" keeps shifting. The durable strategy is not a title that never changes; it is a position with more than one hard task, skills that travel to nearby roles, and a habit of re-running the audit once a year. If your job passes the audit today, check again next year — and if it does not pass, start moving the mix now, while you still have the routine work to trade away. For the longer version of that strategy — how to keep re-checking as the tools improve — see future proof jobs.
No job is fully immune. Replacement happens task by task, not title by title: the routine, screen-based tasks inside a job go first, while tasks that need physical presence, trust, or judgment hold up longer. The jobs people call safe are the ones where the hard tasks dominate the week.
Three things slow replacement down: the work has to happen in person, in a place that keeps changing; it depends on someone trusting you or on you answering for the result; or it needs judgment with incomplete information. A job where most of the week passes all three tests is much harder to automate than one where most tasks fail them.
The families where hard tasks dominate are hands-on trades (electricians, plumbers), care and health roles (nurse practitioners, therapists), teaching, emergency response, and skilled work in unpredictable places. Check demand for any specific occupation in the Bureau of Labor Statistics Occupational Outlook Handbook before committing to a direction.
It cannot predict the future, but a free Holland Code test can show your interest pattern — whether your strengths lean toward hands-on, social, or judgment-heavy work — and which nearby roles use the same strengths if your field changes. The test itself is free, and your result is computed from your answers.
Change the task mix instead of hunting for a safe title. Take on more judgment calls, client conversations, and on-site problems; use AI on the routine parts of your work; and build skills that travel — diagnosis, communication, and hands-on problem-solving carry across fields when a single job starts to shift.
Related: AI proof careers · Will AI take my job quiz · Future proof jobs · Careers safe from AI · Holland Code test