TL;DR. A mixed room in Munich spends Wednesday evening on one question, with no expert panel and no pitch: what does AI mean for my job? This is the background, not the answer: what Anthropic’s own research, Stanford, and the World Economic Forum actually say about AI and jobs in mid-2026, and where my own writing this year has landed, and been forced to change its mind.
On Wednesday evening, a mixed room in Munich, teachers, nurses, engineers, consultants, parents between jobs, will sit down with wine and cheese boards and spend two rounds of conversation on one question: what does AI mean for my job? No expert panel. No pitch. Just people talking it through in small groups, because the best version of this conversation happens when the room is not full of people who already agree with each other.
This post is not the answer to that question. Nobody has the answer yet, and that is the honest starting point, not a disclaimer to get out of the way. What follows is the background: the actual numbers, from Anthropic’s own research and independent economists, as they stand in July 2026, plus where my own writing on this topic has landed after several essays this year, and where it has been forced to change its mind.
The honest starting number
Anthropic, the company that builds Claude, runs an ongoing research project called the Anthropic Economic Index. It looks at real, anonymised conversations people have with the AI and asks a simple question: what is this actually being used for? Its most recent large finding, published in 2026, is that about 49 per cent of jobs have had at least a quarter of their tasks touched by Claude in some way.
That number sounds alarming until you unpack what “touched” means. Anthropic splits usage into two modes: augmentation, where a person stays in the loop and the AI helps them think, draft, or check something, and automation, where the AI simply does the task end to end. For most of 2025 the balance drifted toward automation, from 55 per cent augmented against 41 per cent automated in January to a near-even 55 to 42 split by March. Then it swung back. By November 2025, augmented use had climbed to 52 per cent and automated use had fallen to 45 per cent (Anthropic, Economic Index: Cadences). People are not simply handing tasks over. They are pulling the tool back into a supporting role, which is a very different story from the one usually told in headlines.
Who is actually affected, and it is not who the old automation story predicted
The classic fear about automation, going back to factory robots, was that it would hit manual, blue-collar work first. The Anthropic data says the opposite is happening. Computer programmers show the highest exposure, at roughly 75 per cent of their tasks theoretically within reach of the model. That tracks with what I have spent this year calling agentic engineering: tools like Claude Code let one developer direct an AI agent through a multi-step engineering task end to end, not just autocomplete a line, which is a different order of exposure than a chatbot answering questions. Customer service representatives and data entry roles are close behind. Meanwhile, around 30 per cent of the workforce, cooks, mechanics, lifeguards, bartenders, has essentially zero exposure, because their work is physical and situational in ways a chat interface cannot touch.
There is a second twist. Before ChatGPT existed, workers in the occupations most exposed to AI were already earning 47 per cent more on average than workers in low-exposure occupations, and were far more likely to hold a graduate degree. AI’s first wave is not displacing the most vulnerable workers in the economy. It is reshaping the desks of the relatively well-paid and well-educated first. Whether that stays true, or whether it is simply the order in which the wave arrives, is one of the open questions worth taking into Wednesday’s small groups.
The generation gap is the real story so far
If there is one number from 2026 that deserves to travel further than it has, it is this one. A team led by Stanford economist Erik Brynjolfsson analysed payroll records covering millions of American workers and found a 13 per cent relative decline in employment for early-career workers, aged 22 to 25, in the occupations most exposed to AI, since generative AI tools went mainstream. For the narrower 22 to 25 band specifically, the decline reaches 16 per cent. Employment for older, more experienced workers in the exact same occupations has stayed flat or grown (Fortune, June 2026; Stanford Digital Economy Lab). The effect has not reversed. It has grown by roughly half a percentage point every month since the original data was collected, through nearly four years of data now.
Put plainly: this is not “AI is coming for jobs” in the aggregate sense the headlines usually mean. Aggregate unemployment in the most exposed occupations has barely moved, up around 0.2 per cent year over year as of April 2026. What has changed is who gets to walk through the door in the first place. If you are established in your field, the data so far says you are probably fine. If you are trying to get in, the first rung of the ladder is measurably higher than it was four years ago. A room with both twenty-two-year-olds and forty-five-year-olds in it will experience “what AI means for my job” as two different questions, and the Wednesday format, small mixed groups rather than a single panel discussion, is built for exactly that gap.
Why companies are moving now, not later
I wrote about this in March, when Block, the payments company run by Jack Dorsey, cut close to half its workforce, about 4,000 people, in a single move, and its stock rose 24 per cent the same day (Half the People. All the Output.). The economic logic behind that decision is nearly ninety years old. Ronald Coase argued in 1937 that firms exist because coordinating work inside a company is sometimes cheaper than buying it piece by piece on the open market. When that coordination becomes cheaper still, because AI agents can do work that used to require hiring a person to manage it, the boundary of the firm shrinks. Block’s own engineers were shipping 40 per cent more code per person using the company’s internal AI tooling before the cuts. That is not a one-off. It is the mechanism behind a wider trend: fewer people producing the same or more output, and investors rewarding it.
If you sell your time for a living, the ground already moved
For the freelancers, consultants, and small agency owners likely to be in Wednesday’s room, a second and more immediate shift matters just as much. When a task that used to take forty hours now takes four, someone has to absorb that difference, and increasingly it is the person who used to bill by the hour. Freelance writing rates fell roughly 30 per cent year over year in the commodity segment of the market. Junior software development roles on freelance platforms fell 21 per cent. One analysis of actual corporate spending found freelance marketplace expenditure dropping from 0.66 per cent to 0.14 per cent of total company spend between 2021 and 2025, with every dollar saved on freelancers replaced by roughly three cents of AI cost, a twenty-five-fold efficiency gain for the buyer (Ramp, 2026). McKinsey, the world’s largest consultancy, now earns roughly a quarter of its fees on an outcome basis rather than time and materials, precisely because AI compressed how long the work takes (Your Hourly Rate Is Dead). The billable hour is not dead everywhere, but it is dying fastest exactly where AI can already do the underlying task well.
We have been here before, and the good outcome was never automatic
It is tempting to treat all of this as unprecedented. It is not, and the historical comparison is more useful as a warning than as reassurance. Britain’s handloom weavers, roughly four hundred thousand of them at their peak in the 1820s, saw their entire trade collapse within a single generation once the power loom arrived. The Luddites who smashed those machines between 1811 and 1816 were skilled workers, not fools, and they were not wrong about what was happening to them. They were wrong only about what could stop it. What actually followed was nearly a century of strikes, parliamentary fights, and social invention, the eight-hour day, the weekend, unemployment insurance, before the industrial economy produced anything resembling the middle-class life we now think of as normal (The Ghosts of 1811).
The economists Daron Acemoglu and Pascual Restrepo call the eventual good outcome of that story the reinstatement effect: new technology destroys old tasks but also creates new ones for people to do. Crucially, they are explicit that reinstatement is not automatic. It happened in the twentieth century because unions, public education spending, and policy actively pushed technology toward complementing workers rather than simply replacing them. The World Economic Forum’s own current forecast carries exactly that ambiguity built into it: roughly 170 million new jobs created globally by 2030, against 92 million displaced, a net gain of 78 million, with a further 120 million workers facing medium-term redundancy risk in the meantime (WEF Future of Jobs Report). The net number sounds reassuring until you notice that the 92 million losing their roles, disproportionately clerical workers, cashiers, data entry staff, are not automatically the same people filling the 170 million new roles in AI engineering, cybersecurity, or care work. Whether reinstatement happens for a given person, in a given country, is a policy and institutional question, not a law of economics.
What genuinely nobody agrees on yet, including AI companies’ own users
Anthropic’s most recent survey of its own users, published in June 2026, is worth sitting with precisely because it does not resolve cleanly. Over a third of respondents expect AI to handle “most or nearly all” of their work within twelve months. Yet only about 10 per cent rate losing their own job as likely, a rate slightly below the US economy’s normal annual job turnover. People consistently worry more about their colleagues’ jobs than their own, a well-documented bias that shows up everywhere from car accidents to layoffs. And the people who delegate the most work to Claude, the heaviest automation users, report being the most optimistic about their own future pay, job security, and prospects, not the least. That correlation cuts more than one way: it may mean heavy use genuinely builds confidence and secures a role, or it may simply mean the people already most confident and skilled were the ones who adopted the tool fastest. The data cannot yet tell those two stories apart, and anyone who tells you it can is overselling their certainty.
Why a room, not a panel
This is, in the end, the best case for Wednesday’s format. Everything above is a national or global average, and averages hide almost everything that matters to an individual life: which industry, which country, which employer, which age, which specific tasks make up a specific job. A statistic about “customer service representatives” says very little about what actually happens inside a Bavarian Mittelstand company’s support team versus a call centre in a different country with different labour protections. The people who can tell you what is actually shifting inside nursing, or teaching, or the skilled trades, are in that room on Wednesday, not in a research paper. That is not a weakness of the data. It is exactly why the organisers chose small mixed conversations over a stage and a microphone.
Before Wednesday
Bring the specific version of the question that applies to your actual job, not the aggregate one. What has changed in the last twelve months in your specific role, your specific team, your specific industry? What would you want your employer, your professional association, or your government to be doing right now that nobody is doing? Those are the questions the room can actually answer, better than any report, including this one.
Munich | Claude Conversation. Wednesday, 22 July, Theresienstraße 6, 80333 München. Hosted by Michael Whelehan, Dr. Florian Steiner, Alexander Eiswirth, and Steffi Kieffer, venue partner Make. German and English speaking. Cheese, charcuterie, wine, and something non-alcoholic too. Full details and RSVP on Luma.
The technology works. The access does not always. If you want to pressure-test what AI means for your team, let’s talk.
Forward this to someone asking the same question. Subscribe at drfloriansteiner.com.
Originally published on drfloriansteiner.com