The Human Premium
#195 - The companies spending most on AI are hiring the people AI was supposed to replace
I did a manual scheduling error and the email for this post has gone out on Monday vs my standard Saturdays… sorry for spamming!
Hello friends, I hope you had a great week!
Two months ago, in The Abundance Paradox, I did something I usually try to avoid: I made a specific, falsifiable prediction. The total volume of knowledge work will be larger in five years, not smaller. It was a structural bet on Jevons Paradox, the nineteenth-century observation that when something becomes cheaper to use, we end up consuming more of it rather than less. I thought it was the right bet, but a bet was all it was.
Then, over the past few weeks, two things pushed me to revisit it. The first is personal: inside my own professional world, I keep watching a version of that prediction play out in ways I did not quite expect. The second is that real data showed up, a new study from Ramp and Revelio Labs that for the first time links what companies actually spend on AI to what happens to their workforce afterward.
Most of what I found points in the direction the thesis predicted. But one number surprised me, and it suggests that the standard way of modeling AI’s impact on jobs is missing half of the equation. Everyone, me included, keeps running the same calculation: how much do we save if we move this task from a human to an AI? Almost nobody runs the opposite one: how much more could we charge if we kept the human in?
That second question is where I want to end up. But let me start from what I can see with my own eyes.
The view from the inside
The most striking thing about AI at work, at least in my corner of it, is how it changes what each role can reach. I watch product managers who a year ago needed an engineer for every experiment now building lower-stack code themselves, sketching front-end design decisions, structuring the economics of an A/B test end to end. They have superpowers they simply did not have twelve months ago.
The naive reading is that this makes the specialists redundant. If a PM can produce average code and average design, why keep designers and engineers around? What I observe is the opposite. The specialists moved up the stack too. Freed from producing the routine version of their craft, the engineers and designers concentrate on the projects with the highest returns, the ones that used to sit in the backlog forever because everyone was busy shipping the basics. The backlog was never empty; the work inside it was simply priced out.
So the net effect is that companies need more of both. More PMs, because each one can now carry a whole experiment on their own. More engineers, because the frontier of what is worth building expanded faster than the routine work disappeared. This is Jevons doing exactly what Jevons does, and I wrote about that mechanism at length last time, so I will not repeat it here. What I could not tell you until now was whether my little corner of the economy was an exception. Which is why the new study caught my attention.
What 21,559 companies actually did
Until now, almost everything we knew about AI and employment came from surveys and exposure scores: researchers estimating which jobs a language model could theoretically do, then checking employment in those categories. The Ramp and Revelio study takes a more direct route. Ramp processes corporate card and bill payments, so it can observe which firms actually pay AI vendors, how much, and starting when. Revelio Labs tracks workforce records for those same companies. Link the two and you can watch what happens to hiring after a company starts spending real money on AI, across 21,559 American firms.
The headline: companies that adopted AI intensively grew their headcount by roughly 10% over the following two years, compared with similar firms that had not yet adopted. But the detail matters more than the headline. The effect exists only above a spending threshold. Firms in the bottom two-thirds of AI spending per employee, averaging less than three dollars per employee per month (essentially a chat subscription), saw no employment change at all. The gains appear only in the top tier, spending around $34 per employee per month, which in practice means coding agents, API usage, multiple models, AI wired into actual workflows rather than sitting in a browser tab. And the gains are slow: very little happens in the first six months, then the curve bends upward. Buying AI does not seem to do much on its own; the gains arrive when a company learns to work differently around it, and that learning takes a couple of quarters.
Then there are the two numbers I did not expect. Entry-level headcount at high-intensity adopters rose 12%, faster than overall headcount. And customer service headcount rose about 6%. These are the two job categories that nearly every forecast, including my own last post, treated as the first casualties of AI. At the companies using AI most aggressively, they are growing.
What the data cannot tell us
I want to be careful here, because this study will be quoted for years by people who need it to say more than it actually says.
The first limitation is the one that matters most. This is firm-level evidence, not economy-level evidence. The study shows that companies adopting AI intensively grow faster than comparable companies that have not yet adopted. It cannot distinguish between two very different worlds. In one, AI expands the total amount of work to be done, and adopters are simply the first to capture it. In the other, AI adopters are eating the market share of non-adopters, and every job created at an adopter is offset by a job quietly lost somewhere else. Growth of the pie, or reallocation of the pie? The paper cannot say, and the authors admit this openly.
The second limitation is selection bias. AI adopters were already bigger, faster-growing, more technical, and far more likely to be venture-backed before they adopted. The authors handle this about as carefully as the data allows, comparing early adopters only with later adopters of similar intensity, and the pre-adoption trends line up well. Still, the sample is Ramp customers: a tech-forward slice of the American economy, with the clearest gains concentrated in the Information sector. This is a lens on the frontier, not a portrait of the average firm.
And there is a third point, which is really a tension with something I wrote two months ago. In The Abundance Paradox I cited the roughly 20% employment decline among the youngest software developers since 2022 and worried about the junior pipeline. This study finds entry-level hiring up 12% at intensive adopters. Both things can be true at once: the firms that master AI hire more juniors, while the aggregate on-ramp narrows because most firms are not those firms. But I hold my worry more loosely than I did in May. The honest answer is that the junior story is still being written.
The customer service puzzle
Now to the number I keep coming back to. Why would the companies automating most aggressively hire more customer service agents?
The cost-savings spreadsheet has no cell for this. That spreadsheet, the one underneath every consulting deck and every earnings call, computes a single quantity: the incremental cost of not using AI. A human agent costs X per ticket, the AI costs a tenth of that, multiply by volume, present the savings. The human appears in the model exactly once, as a cost line to be reduced.
But a cost line and a product feature can be the same person, priced differently. At that point the rational move may not be to shrink the human team. It may be to grow it, staff it with better people, pay them more, and sell access to them as a feature of the premium tier. Some of the growth in the data is surely more mundane than this (adopters are growing firms, and growing firms have more customers to serve). But the pattern is at least consistent with a strategy that nobody’s spreadsheet contains: automate the commodity part of the job, and sell the human part at a premium.
The human premium
Once you see that inversion, you start noticing how much of the service economy is already quietly organized around it.
Think about what a taxi driver actually sells. To a New Yorker commuting to a meeting, the product is transportation: point A to point B, fast, no conversation required. Waymo will win that customer, and probably should. But to a tourist landing in the city for the first time, the driver sells something else entirely: which bridge has the view, which neighborhood to skip, the feel of a place filtered through someone who lives there. Same license, same car, two different products. When the robot takes over the commodity ride, the guidance ride does not disappear. It becomes visible, separable, and priceable in a way it never was when both were bundled into one meter.
Or take the bar. Vending machines have dispensed coffee for seventy years. By pure task logic, the barista should have followed the elevator operator into extinction. In Italy this thought experiment sounds almost comical: nobody walks into their neighborhood bar because the machine behind the counter is efficient. You go because the guy knows your order, insults your football team, and tells you what happened on the street that morning.
History offers a cleaner version of the same story. In the 1970s, quartz movements made mechanical watches technically obsolete. A 50 dollar Casio kept better time than a Swiss masterpiece, and the Swiss industry collapsed, losing roughly two-thirds of its workforce. Then something strange happened: the survivors repositioned the mechanical watch entirely around the fact that a human made this unnecessary thing by hand, and the industry went on to enjoy the most profitable decades in its history. Fewer jobs than before, better paid than ever, selling precisely the inefficiency that technology had made worthless.
I do not want to romanticize this, because the counterexamples are instructive. Bank tellers and travel agents were also humans in the loop, and customers swore they valued them, right up until the ATM and the booking site revealed that those preferences were softer than stated. The difference, as far as I can tell, is whether the human is the interface to the service or part of the product itself. Interface humans get automated, and no premium saves them, because nobody ever wanted the interaction, only the outcome. Product humans, the ones whose presence is part of what you are buying, are a different economic species. The uncomfortable part is that most jobs are bundles of both, and nobody knows in advance where the line will fall for their own role.
What is new is that companies now have a margin incentive to find out. The automated tier of every service is heading toward commodity pricing: ten competitors, identical capabilities, race to the bottom. The human tier is differentiable, brandable, and priced on willingness to pay. A company that frames its humans as a cost will cut its way into a commodity business, while a company that frames them as a product gets to charge for something its AI-only competitor structurally cannot offer.
The question changes
For three years the debate has circled around one question: what can AI do? Which tasks, which jobs, which percentile of which occupation. That question produced the doomsday forecasts, and so far the forecasts keep missing: radiology first, then software engineering, now apparently customer service.
Maybe they keep missing because capability was never the deciding variable. The Ramp data suggests that adoption is a strategic choice: some firms buy subscriptions and change nothing, others rewire how they work and grow. The human premium suggests the next choice is subtler still. Once the machine version of every service is cheap and adequate, each company has to decide what its humans are for: a cost to be eliminated, or the reason customers pay more. The volume of work, I am now fairly confident, will keep growing. What nobody has modeled yet is the price of the part that stays human. I suspect there is more margin hiding in that question than in all the cost savings we keep celebrating.
Have a fantastic weekend!
Giovanni