Safe from AI is not the same as paid more.

Jobs AI cannot do will not pay more unless not everyone can do them. The standard career advice for the age of AI is to go where the machine cannot follow. Work with your hands. Work with people. Care for the old, serve the customer, fix the boiler. The advice is right about one thing: a chatbot will not change a bandage or calm a frightened patient any time soon. But it answers the wrong question. It tells you which jobs AI will leave alone. It does not tell you which jobs will pay more because of AI, and for most of the work it recommends, the answer is: not much.

This essay is prospective. It asks what generative AI is likely to do to the wages of in-person jobs that need no degree (care aides, shop assistants, waiters, receptionists, the people behind the counter) and why the answer has less to do with what AI can do than with who else can do the job.

Two questions, not one

Almost every forecast of AI and jobs begins with exposure: the share of an occupation's tasks a model could do. Exposure is a useful map of where work will change. It is a poor guide to wages, for a reason older than AI. A wage is a price, and a price moves only as much as supply lets it. If demand for a job rises and new workers can pour in at the going wage, the extra demand becomes extra jobs. If they cannot, it becomes extra pay.

The textbook formula is short. The percentage change in the wage equals the shift in demand, minus the inflow of new workers, divided by the sum of two elasticities: how much employers cut back when the wage rises, and how easily new workers enter. The second of those, the ease of entry, is what the AI debate mostly leaves out. And it is precisely what separates a nurse from a care aide, or an electrician from a shop assistant.

Where the demand comes from

There are good reasons to expect AI to raise demand for in-person work. Three stand out.

The first is an old argument from William Baumol. When one part of the economy becomes much more productive, what it makes gets cheaper, incomes rise, and spending shifts toward what did not get cheaper: services that need a person in the room. A haircut, a meal served at the table, an afternoon of care. The second is demography. Spain is ageing, and care needs hands whatever AI does. The third is subtler. Joshua Gans and Avi Goldfarb show that when tasks are complements, automating the tasks around a job makes the remaining human task the bottleneck, and the bottleneck gains value. When the booking, the paperwork and the follow-up are automated, what is left of a care visit is the visit itself.

AI will also take tasks away from these jobs. Self-checkout and automated front desks already have, and ambient scribes now draft clinical notes. The early trials on physicians find time savings that are real but modest. The net effect on demand for in-person work is still likely to be positive. The question is what that demand buys.

With the default settings, demand for both jobs rises by 10%. In the gated job, wages rise by about 12% and employment by under 4%. In the open job, wages rise by about 1% and employment by over 9%. The open job does not do badly: it hires far more people. But the people already in it see almost none of the gain. Push the inflow up to 10% and their wage stops rising altogether.

The mechanism also runs in reverse, and the United States has just shown it. Between 2019 and 2025, as the pandemic cut the flow of workers into restaurants and shops and the labour market tightened, average hourly pay in food services rose by 39%, against 30% for all private-sector workers. David Autor, Arindrajit Dube and Annie McGrew call it the unexpected compression: for a few years, the lowest-paid jobs gained most. Nothing about those jobs had become harder to automate. What changed was how many people were available to do them.

Where the workers come from

The inflow is the part of the story that AI itself writes. Workers can arrive in open in-person jobs from three directions, and AI pushes along two of them.

The first is from above. In an earlier essay I argued that AI erodes the bottom rungs of professional careers, and early evidence is consistent with that: Erik Brynjolfsson, Bharat Chandar and Ruyu Chen find that employment of workers aged 22–25 has fallen by about 16% in the occupations most exposed to AI, relative to everyone else. Those young workers do not vanish. Some wait. Many take the job that is available, and the job that is available is often behind a counter. Paul Beaudry, David Green and Benjamin Sand documented the same cascade after the tech bust of 2000: when demand for cognitive work stalled, graduates moved down the occupational ladder and pushed less-educated workers further down. The queue at the bottom of the graduate ladder is already long. In the United States, the New York Fed counts 41.5% of recent graduates in jobs that do not typically require a degree. In the United Kingdom, the share of graduates in low- or medium-skilled jobs doubled between 1992 and 2022, according to the CIPD's analysis of the Labour Force Survey. Spain starts from the most extreme position in Europe: 35% of its employed graduates work in jobs that do not need a degree, the highest share in the EU.

The second is from beside. Clerical and call-centre work is among the most exposed to automation, and it is the closest substitute for counter and floor work. A receptionist whose scheduling is automated can become a shop assistant tomorrow. The third, immigration, has nothing to do with AI, but in Spain it is large and flows mostly into exactly these jobs.

The missing rungs, one floor down

The earlier essay was about rungs within a profession: the easy work through which juniors learn. There is a second set of rungs, between floors. For generations, the in-person job was a way in. The shop assistant moved to the back office, the receptionist became an administrator, the waiter became a manager, the care aide trained as a nurse. Each of those moves passed through an entry-level office job: filing, scheduling, invoicing, answering customers by email. That is the work generative AI does best.

If the middle floor thins, the traffic on the stairs changes direction. Fewer people climb up from the counter, because the stair is narrower. More come down to it, because the floor above has fewer places for beginners. The in-person job turns from a starting point into a destination: workers stay longer, arrive with more education than the job needs, and compete with each other for the few routes up.

Who wins inside the in-person economy

None of this means that all in-person work will stagnate. It means that the winners will be sorted by gates rather than by hands.

Gated jobs are those where entry takes a credential or years of training: nurses, physiotherapists, electricians, plumbers. Their demand rises with the rest of in-person work, and newcomers cannot rush in. Some of them gain a second time. David Autor and Neil Thompson show that when automation removes the easier tasks of an occupation, the job that remains is more expert, wages rise and employment falls. Documentation is the clearest candidate in nursing. If AI takes the paperwork, what remains is the clinical judgment. The pay of the clinician should go up, even if fewer are needed per patient.

Open jobs are those anyone can start next week: care aides, shop assistants, waiters, cleaners. Their demand rises too, but they receive the inflow, and their wages will be set less by AI than by the minimum wage. In Spain, where the minimum wage rose sharply after 2019, that floor is already binding for much of this work.

Sales and customer-facing work splits down the middle. Scripted selling (the cold call, the follow-up email, the standard answer at the till) is exactly the work AI does well, and it is the entry rung of the profession. Relationship selling, where the buyer trusts a person over years, becomes the bottleneck and gains value. But trust is built through experience, and experience used to be built on the scripted work. The same split, and the same problem of the missing rung, runs through almost every occupation that deals with customers.

Where wages will go

Put the pieces together and five things can be said about wages, in falling order of confidence.

Real gains will be small at the bottom. AI makes cheaper what can be delivered through a screen: software, advice, translation, customer service. The budget of a shop assistant or a care aide goes mostly on rent, food, energy and in-person services, which AI barely touches. Even if their pay keeps pace with prices, the cheaper goods AI brings are a small part of what they buy. In Spain, where rent takes a rising share of low incomes, whatever AI saves them at the screen, the landlord takes back at the door.

In open jobs, the wage will be set by policy, not by the market. Countries have already chosen very different floors. The UK set its minimum at two-thirds of median earnings. Spain has tied its minimum to 60% of the average net wage, the European Social Charter benchmark, and raised it by 54% between 2018 and 2025. The US federal minimum has not moved since 2009, and has sunk to a quarter of the median wage (many states set their own, higher floors). Where the floor is indexed to average pay, as in Spain and the UK, AI-driven gains in average pay will pass to open in-person jobs automatically. That is likely to be the main channel through which AI reaches the bottom of the pay scale, and it runs through the official bulletin rather than the labour market. The risk is on the other side of the floor: with a binding minimum and workers still arriving, adjustment moves to hours, part-time contracts and informal work.

The distribution will compress at the bottom and open at the top. More workers will be paid at or near the minimum. The entry-level office job loses its premium over the counter. Gated in-person jobs pull away from open ones, and within skilled professions, as an earlier essay argued, the spread widens. The middle of the distribution drifts toward the floor.

The largest loss will not show up in any single year's wage. If the stairs between floors narrow, starting pay in the counter job barely changes. What changes is the slope: the raise that used to come with the move to the office, and the one after that. A snapshot of wages by occupation will miss it. A worker's earnings over a career will not.

The degree premium will spread out. For the graduate who ends up behind the counter, the return to the degree is close to zero. The average premium may hold, carried by graduates above the frontier, while the median graduate's premium falls.

On magnitudes, honesty requires modesty. The evidence so far points to changes of a few percentage points: a precise null in Denmark, an early estimate of about 5% lower wages in the most exposed US occupations relative to the least exposed, and adjustment for young workers through jobs rather than pay. Over the next five years, the safest forecast is relative shifts of a few points between occupations. The large effects, if they come, will be in careers.

What follows

For workers, the advice to go where AI cannot follow needs a second clause. Choose work AI cannot do and that not everyone can do. The moat around an in-person job is not the absence of a robot. It is a licence, an apprenticeship, or a track record.

For policy, the lever is the gate. If in-person demand grows and the pay gains go to gated jobs, the question is who gets through the gates: how long it takes to qualify as a nurse or an electrician, whether vocational training can be done while working, and whether immigrants' qualifications are recognised. Opening the gates spreads the gain. It also lowers the premium of those already inside, which is why it will be resisted.

For research, the argument makes testable predictions with Spanish data. In the labour force survey, employment in open in-person occupations should rise faster than their wages, and the share of graduates among their young workers should rise. In social security records, moves from counter jobs to office jobs should become rarer, and moves from the office to the counter more common. Wages in gated in-person jobs should rise relative to open ones facing the same demand. The share of workers paid at the minimum should rise, and earnings profiles of workers who start in open in-person jobs should flatten across cohorts. If the gate does not matter, the hypothesis is wrong.