You’ve probably watched this happen, even if the decision wasn’t yours: a team rolls out an AI assistant, resolution time drops, the queue shortens, customers wait less. Then a few months later, during next year’s budget planning, someone asks whether the two open positions that have been sitting unfilled really need filling. Maybe the team can manage without them. The better outcome and the smaller headcount can come out of the exact same rollout, and nobody in that room has to be lying for both to be true.
That tension points at a belief you probably carry around without ever examining it: if you learn to work faster and better, your position has to improve. The belief has its own logic. A better tool in your hands means you get more done in the same time, so you bring more value. For decades that held up - the computer, the spreadsheet, a good design tool all made you more productive, and a more productive you was a more wanted you. There’s no reason to throw that intuition out. The problem starts once you check what it actually assumes about the market that’s supposed to turn your extra sharpness into something more.
A better outcome, a different arithmetic
How much you get done per hour, how many hours the work actually takes in total, and how many people it takes to do it - these are three different things, and it’s easy to collapse them into one. If your output per hour grows faster than demand for the thing itself, you need fewer hours. If the drop in cost pulls in enough new demand, total hours can rise even though each one is more productive. Knowing how much faster you’re working tells you nothing, on its own, about which of those two paths you’re on.
Here’s the simplest version of the arithmetic, purely illustrative: ten people produce a hundred units of a service a week, everyone’s output doubles, and orders climb to a hundred and fifty units. Hours unchanged, that now takes seven and a half roles instead of ten - the company does more while employing fewer people. No real team splits that cleanly - your actual role has bottlenecks and obligations this arithmetic can’t see. What the example proves is narrower: that rising output and shrinking demand for labor can sit next to each other without any contradiction.
In practice, I’ve watched the scale run well past anything a textbook example captures. A team I’m building right now for a similar digital transformation would have needed three or four times as many people before 2022, before ChatGPT became an everyday tool, to do the same work. I don’t say that comfortably - planning a smaller team means planning around fewer people, and I know which side of that decision I’m usually standing on. I see the difference directly every time I plan a team’s structure, and if you’ve built or joined a team in the last two years, you’ve probably felt the same thing without putting a number on it.
What happens to the gain from that higher productivity is a decision, not a law of physics.11What happens to the gain from that higher productivity is a decision, not a law of physics. Here’s something budget conversations rarely say out loud: headcount decisions are rarely made because of productivity numbers alone. They’re almost always the side effect of some other process already underway - a restructuring, an acquisition, an external audit, a cost cut that was coming regardless of what the dashboards said. Higher productivity rarely triggers a cut by itself. More often it just hands someone the argument and the number they reach for once the decision is already being made for a completely different reason.
I’ve watched an even more blatant version of this. The promise that AI would raise a team’s output showed up in redundancy conversations before anyone had so much as gotten access to the tool. Productivity that hadn’t happened yet was already doing the work of justifying a decision that had already been made. I’ve also watched the same thing from the other side: tracking metrics - hours per project, cases handled, gains from AI-driven process work - existed for exactly one purpose, to put a paper trail under a decision made earlier, next to a budget cut and a headcount target someone had already signed off on.
The same mechanism runs the other way too. When a team scales up, or quality improves, that’s usually part of some bigger plan too, rarely a direct response to work having sped up. What actually happens to your team depends less on the productivity number you can point to, and more on whatever other process happens to be running in the background of the organization right now.
Where the demand shows up
The data don’t agree with each other, and that disagreement tells you more than any single number would. A study of an AI assistant rollout at a large customer service operation, covering more than five thousand people, found an average fifteen percent jump in cases resolved per hour.22Brynjolfsson, Li & Raymond, Generative AI at Work. 5,172 customer-service workers. Published in the Quarterly Journal of Economics, 2025. Look closer, though, and the variation matters more than the average: less experienced people gained the most, in both speed and quality, while the most experienced barely sped up at all, and their quality dipped slightly. That study tells you something about how one task changes. It tells you nothing about how many people get employed next year.
A Danish study looked at eleven occupations exposed to large language models, using administrative records and surveys covering about twenty-five thousand workers.33Humlum & Vestergaard, Still Waters, Rapid Currents. Denmark, 11 exposed occupations, administrative data through 2024. The researchers found no significant average effect on pay or recorded hours, and their specification rules out anything bigger than a two percent effect. At the same time, they saw tasks getting reorganized inside firms, and some amount of people moving between jobs. It’s one country, early in adoption, so it doesn’t tell you much about how big the effects eventually get. It’s enough, though, to puncture the story you sometimes hear about an immediate, economy-wide collapse in jobs.
Put the two next to each other and they’re measuring different things at different levels: one shows you what changes inside a task when AI does part of it, the other what happens to an entire country’s labor market when adoption has barely started. Neither answers the question you actually care about: will you be needed in this same role five years from now?
The risk to your employment doesn’t have to look like a layoff email to be real. US data show that employment of twenty-two to twenty-five-year-olds in occupations more exposed to generative AI grew noticeably slower than it did for their less exposed peers - about nineteen percent below where the trend would put it.44Stanford Digital Economy Lab & ADP, Canaries in the Coal Mine? US data, workers aged 22-25, through June 2026. The researchers point out there’s no matching gap among experienced workers, and that this isn’t proof of broad displacement across the economy. What you get instead is a narrower door. A team that already has its experienced people might not fire anyone, and still quietly stop opening spots for people trying to get in. If you’re starting out, that mechanism hurts just as much as a layoff, even though nobody ever announces it.
There’s a second bar rising underneath this one, and the statistics haven’t caught up with it yet: knowing how to actually use AI tools is stopping being your edge and becoming your entry ticket. If you can’t open Cursor, Antigravity, or VS Code today and get an idea running on localhost, you’re competing at a growing disadvantage. Your talent hasn’t changed - the ground under it has, and it moved fast enough that almost nobody got fair warning.
The strongest pushback to everything above deserves a real hearing: a lower cost of doing the work can create demand that genuinely wasn’t there before, because paying someone for it used to be a cost a client actually felt, and now it isn’t. This is the actual mechanism behind the gentler version of the future, and it deserves more than a dismissive footnote. You can see it clearly in the creative industries: tools got cheap enough that producing text or graphics costs next to nothing, and small businesses that used to think twice before hiring a copywriter or a designer now just make that work themselves. What they make is often full of mistakes, the kind of thing the industry now calls AI slop, but it’s good enough for that audience, because nobody there was asking for perfection. The same shift that opens this new demand wipes out entire careers on the other side of it, because the clients who used to pay for that work simply stop ordering it.
A second, quieter version of this runs inside the profession itself, alongside the new clients who weren’t being served before. People who use AI well are scaling up and taking clients from people who don’t, inside the exact same job title. A translator, a video editor, an illustrator, or an SEO consultant with a real handle on prompting can run a client list today that used to need a whole team. Neither of these two forces - new demand, or someone else scaling past you - reaches the person who just lost the work automatically.55Neither of these two forces - new demand, or someone else scaling past you - reaches the person who just lost the work automatically. Whether either one arrives in time for you depends on timing, scale, and where you happen to be standing.
Who gets the benefit
Every study above describes a world where AI is one tool among several, adopted slowly and only in parts. A model built by a team at the Anthropic Institute asks what happens under much bigger assumptions - it traces three different speeds and scales of AI adoption through the US economy out to 2030.66Korinek, Jones, Sacher, Cotter & McCrory, Economic Scenarios for Transformative AI. Anthropic Institute Working Paper No. 2026-02. US economy only; three conditional scenarios, no probabilities assigned to any of them. Worth flagging up front: this only covers the US, and its three variants are conditional scenarios, each dependent on a different bundle of assumptions, with no probability attached.
In the mildest of the three, knowledge-sector wages in 2030 sit four-tenths of a percent above where they’d be without AI, and unemployment across the economy is up by one-tenth of a point - a difference you’d barely notice. In the middle scenario the picture flips: knowledge-sector wages are three-tenths of a percent below where they’d be without AI, and unemployment climbs to 4.6% from a normal 3.8%. In the extreme scenario the gap opens wide: knowledge-sector wages come in eleven and a half percent below where they’d be without AI, unemployment inside that group hits nearly eighteen percent, and across the whole economy, close to twelve. Every one of those numbers is measured against a hypothetical no-AI path over the same years. “Eleven and a half percent lower” means your wage grew more slowly across those four years than it otherwise would have, measured against that path in 2030.
What the model makes visible is the gap between two things you’d otherwise lump together: how much richer the whole economy gets, and how much a specific group of workers takes home. In the extreme scenario, output is a third higher than it would be without AI, and knowledge-sector wages are still lower against that same baseline. The economy can get richer while your group loses ground - what decides the split is the balance of power in negotiating the gains, not the sheer size of the growth.77The economy can get richer while your group loses ground - what decides the split is the balance of power in negotiating the gains, not the sheer size of the growth.
You can see the same direction outside the model, in what the labs themselves are actually optimizing for. OpenAI and Anthropic are racing each other on two fronts at once: who builds the more capable model, and who gets the cost per token down faster, precisely so they can automate their own internal work and keep their own headcount costs flat while output keeps climbing. The companies building the tool are running the exact rule this letter opened with, on themselves: better outcome, fewer people needed.88The companies building the tool are running the exact rule this letter opened with, on themselves: better outcome, fewer people needed.
The second pushback you’ll hear: the best will be fine, because they’ll just take a bigger slice of the market. Probably true, for some of them. It doesn’t tell you how many seats are left for the profession as a whole, or whether the premium for being the best goes to the person doing the work or gets swallowed by whoever controls the client relationship. Someone else’s success in a shrinking market tells you nothing about how many seats are disappearing from it.
None of these three scenarios is fate. They’re conditional, hinging on how fast AI capability actually grows and how widely people use it, and, as the model’s own stress tests show, on how easily the economy can add capital and how fast wages adjust. Change those assumptions and the number changes with them. Take any single figure with real caution, and take the mechanism behind it seriously anyway.
Getting better at your craft happens inside a market whose appetite for that craft can shift for reasons that have nothing to do with how hard you worked to get better. That doesn’t erase the point of learning, or the value of doing better work. It just adds a second question next to “am I doing this faster and better now”: how much of this work will still be needed, and who’s going to end up pocketing the difference once it’s cheaper.
You can judge your future purely by your own speed and the quality of what you ship. Or you can add a second question to that judgment - how much demand actually stands behind the work, and who captures the value once it costs less to do. That second question is the one that actually explains why anyone still pays you, once the work itself has gotten cheaper than it’s ever been. I’m asking myself that same second question about my own work, for what it’s worth. None of us gets to sit this one out.