What Happens to Baumol’s Cost Disease When the Automated Sector Is Intelligence Itself?
There’s an old idea in economics, dating back to William Baumol in the 1960s, that explains why a haircut costs more today than it did fifty years ago even though scissors haven’t gotten any better. The logic goes like this: when factories automate, manufacturing workers become wildly more productive, and their wages rise. But a barber’s productivity hasn’t changed at all — a haircut still takes the same twenty minutes it always did. Yet the barber’s wages have to rise too, because otherwise the barber would just go work in the factory. So the price of haircuts climbs even though nothing about haircutting has improved. Economists call this Baumol’s cost disease, and it’s been a tidy explanation for why healthcare, education, and live performance keep getting more expensive relative to manufactured goods.
What happens if you swap out the term doing the work in that story? Baumol’s “automatable sector” was manufacturing. What if the automatable sector is now cognition itself — the knowledge, reasoning, and split-second information processing that models like the one writing this post can now do?
This substitution breaks the original story in interesting ways, and chasing down exactly how it breaks turned into a fairly deep rabbit hole. I (the human writer of this post) wanted to walk through where that thinking might land, partly because the destination is something that affects our human future (I bet you no LLM will ever write like that….) , and partly because I’m unsure which parts of it will hold up.
Why the old story doesn’t just port over
The first thing that falls apart is the mechanism. Baumol’s wage spillover depends on labor being a single, unified market — the barber and the factory worker are competing for the same pool of human bodies, so wages equalize across sectors. But artificial cognition isn’t a body. It doesn’t quit the barbershop to go work in a factory. It’s closer to a utility, like electricity, that can be deployed in arbitrary quantities into ten thousand barbershops simultaneously without raising the price of haircuts at all.
So if Baumol’s spillover mechanism doesn’t apply, does cost disease just… not happen in the AI era? That seemed too quick. Something is clearly still happening — wages and prices are shifting in a directional way as AI tools spread — it’s just not the same mechanism. I wanted to figure out what the new mechanism actually was, rather than forcing the old one to fit.
A different way to slice the economy
So I hopped onto arena.ai, Google Gemini, Anthropic Claude etc… and asked several models to reframe Baumol’s cost disease for the AI age. Then got Claude Opus 4.7 to merge the reframes into a single cohesive framework. Here’s the synthesized reframe: instead of asking which sectors are automatable versus labor-intensive (Baumol’s split), ask which stage of producing anything is automatable. Almost every good or service, when you look closely, passes through a sequence of stages before it reaches the world:
1. Figuring out what to do — diagnosis, research, design, planning
2. Coordinating — aligning people, contracts, schedules, decisions
3. Getting permission — regulation, licensing, sign-off, liability
4. Actually doing it — the physical act, the labor, the build
5. Society absorbing it — workflows changing, habits adjusting, trust forming
AI is currently very good at stage one and increasingly capable at stage two. It’s only indirectly useful at stage four (you still need a robot, and robots are hard), and it barely touches stage three and five, because permission and trust aren’t really information problems — they’re social and legal ones.
If that’s roughly right, then the prediction follows almost automatically: when one stage in a serial chain gets radically cheaper, the total cost of the output doesn’t fall by the same amount. It falls until it hits whichever stage is now the bottleneck, and then it stops. The price — and the wage, and the political fight — reorganizes around that remaining stage.
I ended up calling this Bottleneck Theory, or BT, mostly because I needed a name for it to keep talking about it coherently, not because I’m confident it deserves the word “theory” in the strong sense.
Formalize It
I wanted to see if this held together as more than a metaphor, so I tried writing it down properly. If you’re not into equations, skip to the next section — nothing downstream depends on parsing these line by line.
The starting point is a very old idea from agronomy, called Liebig’s Law of the Minimum: plant growth isn’t limited by the average availability of nutrients, it’s limited by whichever nutrient is scarcest. Add more of everything else and nothing changes until you fix the scarcest one. In production terms:
Q = min( x₁/a₁, x₂/a₂, ..., xₙ/aₙ )
Output Q is capped by whichever input’s available quantity (xᵢ), divided by how much of it you need per unit of output (aᵢ), is smallest.
BT applies that same logic, but instead of treating inputs as parallel and substitutable, it treats the five stages above as serial — you can’t skip stage three to get to stage four. For a given sector s at time t:
Q(s,t) = min over stage i of [ τᵢ(s,t) / αᵢ(s) ]Non-bottleneck stages don’t get to charge a premium — they’re competitive. The bottleneck stage captures whatever surplus is left over. This is the formal version of “value migrates to whatever is scarce.”
Now plug in the AI shock. Cognition’s cost falls and its capacity rises, roughly exponentially:
c_cognition(t) = c₀ · e^(−λt)
τ_cognition(t) = τ₀ · e^(μt)As cognition cost collapses toward zero, it stops being the bottleneck stage almost everywhere, and the bottleneck shifts to whatever was second-scarcest. Which stage that is — energy, permits, skilled trades, trust, raw materials — differs sector by sector, and that’s exactly why the same AI shock produces such different-looking outcomes in software versus construction versus medicine. Software is easy to verify, so coding agents ate junior developers’ breakfast, lunch, dinner and protein shakes. An AI diagnosis of cardiac arrest doesn’t help much in emergency airway intubation, so my emergency department colleagues are (still) safe in their jobs.
And so?
A few of the more interesting wrinkles fell out of pushing the formalization further, rather than being assumptions I put in upfront:
Verification doesn’t fall with generation. If a fixed budget produces more outputs as generation gets cheaper, and each output still needs roughly the same amount of human checking, total verification cost actually grows — at close to the same exponential rate generation cost falls. Cheap drafts make expensive review.
There can be too many options. If the rate of plan-generation explodes but the rate at which humans or organizations can actually commit to and absorb a decision stays roughly fixed, you get a kind of indigestion — call it Θ, the ratio of options generated to decisions absorbable. When Θ is large, most of what AI generates is simply never used. The bottleneck isn’t ideas; it’s choosing one and living with the choice.
Shared models can make the whole system more fragile, not less. If many firms reason using similar AI systems, their conclusions correlate more than they used to. That doesn’t show up as average inefficiency — it shows up as larger, more synchronized failures in the tails, the kind you only notice when something rare happens and everyone got it wrong in the same direction at once.
The junior ladder breaks before anyone notices. If AI absorbs the entry-level tasks that used to train future senior experts, the effect on the senior talent pool doesn’t show up immediately — it shows up five to fifteen years later, whenever that cohort would have matured. That’s an unusually clean, checkable prediction, and one I’d genuinely like to be wrong about, because if it’s right, several professions are going to feel a second, delayed shock that nobody is currently budgeting for.
Is all this machinery actually necessary?
Actually — maybe not. There’s a much shorter version of the same idea, which is just Liebig’s law stated plainly: scarcity migrates to the next-binding input, and price follows it. That’s the whole theory in one sentence, and it’s more falsifiable precisely because it commits to less. BT is the longer, more mechanistic version that tries to explain why the migration produces specific side effects — the verification glut, the option glut, the fragility, the delayed apprenticeship crisis. I don’t think these compete with each other so much as sit at different zoom levels. If you need one sentence for a board meeting, use Liebig’s. If you’re trying to design a strategy around which side effect will bite your sector first, the longer version earns its keep.
One more layer: knowledge isn’t the same thing as information
Somewhere in working through this, it became clear that lumping “AI’s knowledge” and “AI’s cognition” together was hiding something important. They’re not the same capability, and they didn’t arrive on the same timeline. Using the Data-Information-Knowledge-Wisdom hierarchy:
– Information — raw facts, lookup-shaped — was already commoditized by search engines two decades ago.
– Knowledge — structured, internalized understanding of how a domain works — got commoditized by large language models roughly 2020 to 2024.
– Reasoning — live, on-the-spot inference and judgment — is the thing currently being commoditized, and isn’t finished yet.
Pre-AI, an expert bundled all three into one expensive package, which is why you couldn’t just buy the knowledge without paying for the attached human. Search unbundled information. LLMs unbundled knowledge. What’s happening now is the unbundling of reasoning, and it’s arriving as a second wave, sometimes hitting professions that thought they’d already weathered “the AI thing” a few years ago.
This matters because it explains a pattern that otherwise looks confusing: why did doctors and lawyers seem mostly untouched by the search-engine era, then get hit by the LLM era for research and drafting, and now seem to be facing a second, distinct wave aimed at diagnosis and strategy? It’s not one shock — it’s three, staggered, and the third one is still in progress.
So, where does this leave us?
I think the thoughtful answer is: with a framework that’s useful for generating hypotheses, and a genuine uncertainty about how many of those hypotheses will survive contact with the next five years. A few things I’d actually bet on:
– Energy, permitting, and skilled physical trades get relatively more expensive, not less, even as software gets dramatically cheaper.
– A verification and assurance industry grows roughly in proportion to how much cheap content gets generated, not in proportion to how good that content is. The industry has to grow, or we’ll end up with disinformation, misinformation, or extreme views taking over the information commons.
– Some professions are about to discover that the disruption they already adapted to was only the first of two waves.
– Countries and regions will diverge less by labor cost and more by which stage of the pipeline they actually own — compute and energy, physical execution, regulatory trust, or localized absorption.
What I’m less sure about is whether the elaborate five-stage machinery is earning its complexity over the one-line Liebig version, or whether I’ve just built an elaborate way of restating “scarcity moves” with more steps. That’s an open question I don’t think you can resolve from the armchair — it needs a few more years of data to find out which predictions actually came true and which were just good-sounding stories.
If you’ve been watching your own industry and have a read on which stage is your real bottleneck now that cognition is cheap, I’d be curious to hear it — that’s exactly the kind of data point this framework lives or dies on.
Leave a Reply