By Mufy Pachorawala, founder of getAIwork · Last updated: August 20, 2026
Key facts
- Tasks are automated before jobs are. Most roles are a bundle; AI takes the routine parts and reshapes the rest
- 72% of the AI listings we screen are technical — AI is creating skilled demand at the same time it removes routine work
- 281 live platform programmes on our board are people being paid to supervise AI — the correction layer is itself a job category
- Exposure is not the same as replacement. High-exposure jobs often change shape rather than disappear
- The safest move is adding a resistant property to what you already do, not switching to a ‘safe’ job you have no standing in
What makes a job hard for AI to replace?
Job-title lists age badly because automation does not arrive job by job — it arrives task by task. A more durable way to think about it is to ask which properties of a piece of work make it expensive or impossible to hand to a model. Four properties do most of the work, and the more of them a role has, the more resistant it is.
| Property | Why AI struggles with it | Work where it dominates |
|---|---|---|
| Physical & unpredictable | Robotics in unstructured environments remains far harder and more expensive than software. A model can plan a repair; it cannot crawl under the sink. | Trades, maintenance, field service, care work, logistics on the ground |
| Accountable | Someone must be legally, professionally or morally answerable for the decision. Accountability cannot be delegated to a system that cannot be held responsible. | Medicine, law, safety engineering, audit, licensed professions |
| Relational | The value is the human relationship — trust, persuasion, care, negotiation with a person who wants a person. | Therapy, teaching, sales, negotiation, senior client work, nursing |
| Ambiguous | The problem is not yet well defined. Deciding what should be done, with incomplete and conflicting information, is different from executing a defined task. | Leadership, strategy, research direction, novel problem-solving |
Notice that none of these are about intelligence. Models are already better than most people at many well-specified cognitive tasks. The friction is elsewhere: bodies, liability, trust and the mess of undefined problems.
Which jobs are safest from AI?
The most resistant roles stack several properties at once. A paramedic is physical, accountable and relational simultaneously — three separate reasons the work stays human. That stacking, rather than any single property, is what makes a role durable.
- Skilled trades and field work — electricians, plumbers, HVAC technicians, lift engineers. Unpredictable physical environments plus liability plus on-site judgment. Demand pressure here is demographic as much as technological.
- Hands-on healthcare — nursing, paramedics, physiotherapy, dentistry, care work. AI is a genuinely useful diagnostic and documentation aid in these fields, which changes the day-to-day without removing the person.
- Licensed and accountable professions — surgeons, structural engineers, senior auditors. AI is doing more of the analysis; the signature and the responsibility stay with a named human.
- Frontline education and mental health — teaching, counselling, social work. The relationship is the mechanism, not the packaging.
- Complex negotiation and senior relationship roles — enterprise sales, diplomacy, dispute resolution, executive leadership.
- Skilled craft and bespoke work where the buyer specifically values human authorship or one-off physical execution.
The honest caveat: “safest” means the work is least likely to be eliminated, not that it will be unchanged. Every role on that list will have AI inside its workflow within a few years, and the people who use it well will out-compete the people who refuse to.
Answer 6 questions and we match you against 816 screened live listings by skill, level and country.
Take the 2-minute match quiz →
What is actually being displaced?
The pattern in what we screen and in what employers advertise is consistent: pressure falls hardest on structured information processing that can be specified in a document — and it falls first on the entry rungs of those ladders, which is the part that gets least attention and matters most.
Most exposed: routine content production, first-line support scripts, basic data entry and reconciliation, template document assembly, first-pass translation, simple scheduling and coordination, and junior tasks in analysis and research whose output is a summary. What these share is a defined input, a defined output, and a tolerance for review.
The consequential effect is not mass unemployment in these fields but the erosion of the apprenticeship rung. Professions have historically trained seniors by paying juniors to do the routine work. When the routine work is automated, the training path needs rebuilding — and that is a real, current problem for anyone entering law, accounting, analysis, translation or content today.
There is a counter-current worth naming, because it is the thing our board measures directly: supervising, correcting and evaluating AI output is itself now a substantial category of paid work. 281 live platform programmes on our board are exactly that, and 72% of all our live listings are technical roles that exist because of AI rather than in spite of it. New work is being created in the same motion that removes old work — it is simply not distributed evenly, and it does not appear in the same place or to the same people.
Why “safe job title” is the wrong question
Two errors follow from thinking in job titles. The first is false comfort: a title on a “safe” list can still lose 60% of its daily tasks, which changes headcount, pay and career structure without the job vanishing. The second is false panic: a title on an “exposed” list often just gains a tool and a raised expectation of output.
The better question is about your own week. Take the tasks you actually do and sort them by how well-specified they are. The parts that could be written as a clear instruction with a checkable output are the exposed parts — regardless of what your job is called. The parts that require your body, your licence, your relationships or your judgment about what matters are the durable ones.
That exercise also tells you where to invest. If most of your week is in the exposed column, the move is not to abandon the field — it is to grow the durable column inside it.
What should I actually do about it?
- Audit your week honestly. Split your recurring tasks into “specifiable” and “not specifiable”. The ratio is your real exposure, and it is more informative than any published index.
- Add a resistant property to what you already do rather than starting over. A copywriter who owns client relationships and results is in a different position from one who fills briefs. A bookkeeper who advises is not a bookkeeper who reconciles.
- Become the person who supervises the automation in your field. Domain knowledge plus AI fluency is a scarcer combination than either alone, and it is the transition path that does not require throwing away your experience.
- Get demonstrably good at verification. Models are fast and confidently wrong; the person accountable for catching that is doing work that cannot be handed to the thing being checked.
- If you are early-career, seek roles with a real apprenticeship — where you are trained rather than only tasked. That rung is thinning, so it is worth optimising for deliberately.
For where the paid demand currently sits, AI jobs in 2026 covers role types and what listings ask for, our live market statistics show the skill and level splits on the board, and best AI certifications covers which credentials are worth the time.
See which of the live, human-approved listings fit the experience you already have.
Find your match in 2 minutes →
Frequently asked questions
What jobs can AI not replace?
Work that is physically embodied in unpredictable environments, legally or professionally accountable, built on human relationships, or requires judgment under genuine ambiguity. Roles that stack several of these — paramedics, electricians, teachers, licensed engineers — are the most durable. The properties matter more than the job titles.
Will AI take my job?
More likely it will take some of your tasks and change what the rest are worth. The useful question is what share of your week could be written as a clear instruction with a checkable output. That share is your exposure, and it varies enormously between two people with the same job title.
Which jobs are most at risk from AI?
Structured information work with defined inputs and outputs: routine content production, scripted first-line support, data entry and reconciliation, template document assembly, first-pass translation, and junior analysis whose output is a summary. Entry-level rungs in these fields are under the most pressure.
Is AI creating jobs as well as removing them?
Yes, though not evenly and not always for the same people. 72% of the live AI listings we screen are technical roles that exist because of AI, and 281 live platform programmes are people paid to supervise and correct model output. Creation and displacement are happening simultaneously in different places.
Are creative jobs safe from AI?
Partly, and less than creative workers hoped. Commodity creative production is under real pressure; work where the buyer wants a specific human’s authorship, taste or accountability is far more durable. Within creative fields the split usually runs along ownership of the client relationship and the outcome.
Should I switch careers because of AI?
Usually not as a first move. Adding a resistant property to work you already have standing in tends to beat starting from zero in a field where you have none. Switching makes sense when your current role is mostly specifiable tasks and there is no path to the durable parts of it.
How long do I have to adapt?
Nobody credible can date this, and forecasts in this area have a poor record in both directions. What is observable now is that adoption is uneven by sector and that entry-level routine work is moving first. Treating it as a multi-year shift you should start on now is more useful than betting on a timeline.
Mufy Pachorawala · Founder, getAIwork
My AI scans thousands of AI-job posts — 37,792 screened so far — and I personally approve every listing before it reaches the board. I write these articles by the same rules: pay quoted only as listed, and no income promises. Read our editorial rules.
Related: AI jobs in 2026: every role type · Live AI job market statistics · Best AI certifications
