Is AI destroying jobs? Two years into that argument, the most careful American evidence says no, or at least not yet, and not in the way the headlines promised. What it is doing is harder to spot in an employment release and much easier to spot on an invoice.
That distinction is the whole finding of Canaries in the Coal Mine?-, the Stanford Digital Economy Lab study by Erik Brynjolfsson, Bharat Chandar and Ruyu Chen, revised in August and built on ADP payroll records covering millions of workers through June 2026. They find no evidence of widespread displacement. They do find a hole at the entry level: employment of 22 to 25 year olds in the occupations most exposed to AI runs 19% below where it would have been had it tracked their less exposed peers, a gap that stood at 13% when they first measured it a year earlier. The adjustment, they note, is arriving through hiring rather than through base pay, and they are careful to call the result an early indicator rather than a causal estimate.
So the money is moving somewhere. We asked people doing the work, and two of them answered with figures.
The $1,000 that disappeared
Ricardo Leon is QA Officer and Content Manager for the Mexican market at Financer.com, a financial comparison platform operating in more than twenty markets. He is from Venezuela, has almost ten years in SEO behind him and freelances when he has spare time. His case is the bluntest we received.
«I have lost 2 clients, and the workload has been pretty diminished with the remaining ones. My income has fallen from $2,000 USD to $1,000, and I feel that I am doing more now to keep my current jobs», says Ricardo Leon.
Two things that figure does not say, and we are not going to pretend otherwise: he does not specify the period over which the fall happened, and the clients are from the freelance work he runs alongside his job. What survives either reading is the halving.
The strange part, for his trade, is that AI made the work faster and the pay smaller at the same time.
«Creating a good content piece could take up to 5 hours; now you can leverage AI to create something even better and humanize it in less than 2 hours», he says.
Volume, in other words, stopped being the constraint. The constraint moved somewhere else, and his diagnosis of where should worry anyone who publishes for a living.
«AI [...] eliminated the incentives to create informational content, since they can provide a better or more complete result in seconds (using your own as a base, BTW); the money is now in transactional keywords and content focused at the bottom of the marketing/sales funnel», says Leon.
The work has not disappeared. It has slid down the funnel, towards the point where somebody buys something.
On training, no miracles and no invoices. Leon says he paid for no courses beyond the one bundled into his AI assistant’s own subscription, and reckons the colleague who showed him the ropes had paid to learn them. The regret is his own.
«I have been using it for almost a year now, and I regret not putting in more effort at first, because I feel a bit behind», he says.
Then there is the part that happens at home, which no productivity study measures.
«At home, our emergency fund disappeared a few months ago after my [partner] faced the same situation with clients and income reduction», he says.
That fund was worth six months of income, built by saving almost 35% of what the household earned. Two incomes falling together is precisely the scenario it existed for, and six months turned out to be shorter than it sounds.
The ten hours that came back
Suzi Prokell is a public relations consultant in the United States with more than 30 years of experience, at the head of Prokell Publicity. Same tool, opposite ledger.
«AI has helped increase my income by making me way more efficient. It has significantly increased the amount of work I can complete in my same 30-hour workweek», says Suzi Prokell.
The mechanism is unglamorous and specific.
«I used to spend about two hours every morning scanning the news for opportunities relevant to my clients. I now have AI set up to do that in the background on an ongoing basis for each client. It monitors and flags potential opportunities while I do other things. I can be at the gym while it’s working, and when I come back I can view the report, decide what’s worth pursuing and start pitching reporters», she says.
Two hours a morning is roughly ten hours a week, a third of the week she actually works. Then comes the number, and the qualification that makes it honest.
«My billable rate ranges from $125-$200 an hour so those 10 hours represent $1,250 to $2,000 in potential billable capacity each week. Not every saved hour becomes a billable one because I’ve deliberately chosen not to fill every available hour with work», she says.
Read that twice, because it is the line most coverage of AI and productivity leaves out. Capacity freed is not revenue banked. Prokell says her income is up, and gives no figure for it, so no figure appears here. What she does quantify is the skill that made the difference.
«I learned AI by using it every day in my actual work. The most valuable skill has been knowing what to ask for and having enough experience to recognize when the answer isn’t good enough», she says.
Thirty years of judgement, in short, is what turns a tool into a margin.
Two tracks, one tool
Put the two cases side by side and the pattern is the one the aggregate data keeps hinting at. AI moves hours first and money second, and which way the money goes depends on what you were selling.
| WHERE | WHAT THE DATA SHOWS |
| United States | Employment of 22 to 25 year olds in AI-exposed jobs runs 19% below less exposed peers |
| United Kingdom | 9.4% of job postings mention AI, a record, while graduate postings are down about 7% in a year |
| Euro area | Firms using AI intensively are about 4% more likely to add staff than others |
| Global | 62% average wage premium in job ads that ask for AI skills |
That last figure comes from PwC’s 2026 Global AI Jobs Barometer, which read more than a billion job advertisements across 27 territories and found the premium for AI skills rising to 62% from 57% a year earlier. The same report describes a labour market splitting into two tracks, with the sharpest rewards going to roles that need human judgement alongside the tooling. It also carries the statistic that reads like Leon’s story told from the employer’s side: since 2019, entry-level adverts demanding prior experience are up 35%, while other entry-level postings fell 10%. The first rung is being raised.
Britain’s graduate squeeze, Europe’s hiring puzzle
Britain shows the squeeze with unusual clarity. Indeed’s Hiring Lab reported in August that 9.4% of UK job postings now mention AI, a record, while graduate postings sat about 7% below the previous year and at their weakest for the time of year since 2020. Postings overall were down 11% since the start of the year, and posted wage growth had slowed to 3.9%, the lowest since February 2022. The exception, inevitably, was IT systems and solutions, top of the table at 7.2%.
The euro area complicates the story in the other direction. In an ECB Blog post- published in March, Laura Lebastard and David Sondermann went through the bank’s SAFE survey and found that firms using AI intensively are about 4% more likely to take on staff than firms that do not, and that only 15% of AI-using firms cite cutting labour costs as a reason for adopting it. Usage, though, is deeply uneven: almost 90% of companies with 250 employees or more use AI, against 60% of those with fewer than ten.
Which is the honest summary of all of it. At company level AI is, for now, more often a hiring story than a firing one. At the level of one person’s invoice, it can halve an income in a year.
How to protect an income that can move
- Count your client concentration before your tools. Leon lost two clients, not a job. Revenue resting on a handful of buyers is the exposure that matters most.
- Decide what the freed hours are for. Prokell’s ten hours are worth something only because she chose, deliberately, what to do with them. Capacity nobody sells is a nicer life, not a raise.
- Size the buffer on outgoings, not salary. Three to six months of essential spending, more if your income is variable, and count the exposure of whoever else pays the bills in your house.
- Learn on live work. Neither contributor paid for training. Both learned by using the tools daily, and the one who started late says he can feel it.
Frequently asked questions
Is AI cutting pay?
- Not visibly, in the aggregate. The Stanford work finds the adjustment running through hiring rather than base compensation, and the euro area data points to AI-intensive firms hiring rather than cutting. Individual cases in content-heavy trades can move much faster than the averages.
Who is most exposed?
- People selling output a model can approximate in seconds, to a small number of clients, and people trying to get onto the first rung. Experienced workers in roles where AI complements rather than substitutes are holding up.
Do hours saved turn into money?
- Only if you sell them. Prokell’s own framing, potential billable capacity rather than income, is the right way to read every claim about AI paying for itself.
How much should I keep in reserve?
- Three to six months of essential outgoings, six or more if your earnings vary. Leon’s household adds the rule the tables leave out: if two incomes are exposed to the same shock, treat them as one.
Prefer this in Italian? We ran the same story, with the Italian numbers, in the Italian version of this article.
