When Everything Gets Cheaper, What Gets Expensive?
Three economic laws, one AI revolution, and what they actually predict about where value goes.
/* This piece started as an internal note for my team, written after I noticed a pattern I couldn’t quite name. Design was getting faster. Output was multiplying. And yet something about the work felt harder, not easier. Certain decisions carried more weight, not less.
I wanted to understand the mechanics underneath that feeling. Not the cultural conversation about AI — that one is loud enough — but the structural logic beneath it. The economic laws that don’t care about the hype cycle, and don’t reverse when the next model drops.
What follows is thinking provoked by observation. We are collectively walking into significant unknowns. But some laws still apply. A few of them walked straight into the AI revolution. */
The Quartet Problem
In the 1960s, economist William Baumol noticed something that shouldn’t have surprised anyone but somehow did: when technology makes some sectors dramatically more productive, the sectors it can’t automate become relatively more expensive — not because they improved, but because everything around them got cheaper.
His example was the string quartet. A factory can produce exponentially more goods per hour through automation. A quartet cannot perform the same piece twice as fast without becoming a different thing. The musicians’ wages rise anyway — not because the quartet got more productive, but because everything else did. They have to be paid enough not to leave for a more productive sector. The concert gets more expensive not from improvement but from the rising productivity tide everywhere else.
This is Baumol’s Cost Disease. For decades, it was easy to identify the work it protected. Physical work. Relational work. The therapy session, the live performance, the trades. You cannot download a plumbing emergency.
Baumol also used haircuts as his go-to example — the idea being that no technology could make a haircut meaningfully faster without it becoming something else entirely. That still holds, mostly. Though if you’ve seen the robot barbers appearing in shopping malls, you might wonder how long “mostly” lasts. The specific example may not be permanently safe. The logic behind it is.
AI shifts the terrain in ways Baumol didn’t anticipate. The things that seemed safely human — writing, designing, coding, analysing, generating — are now subject to the same productivity explosion that manufacturing experienced in the last century. The cost of producing a landing page, a piece of copy, a visual concept, a data analysis: collapsing.
Which means Baumol’s logic arrives from a new angle. The work that can’t be automated — not because it’s physical, but because it’s genuinely complex, contextual, and judgement-dependent — becomes relatively more expensive. Not because it improved. Because everything around it got cheaper.
The Paradox That Makes It Stranger
Before asking what falls into that category, there’s a second law worth understanding: the Jevons Paradox, which explains why abundance doesn’t simply solve the problem.
In the 19th century, economist William Stanley Jevons observed something equally counterintuitive: when a resource becomes dramatically more efficient to use, total consumption tends to increase, not decrease. Efficiency lowers cost. Lower cost expands usage. Expanded usage exceeds the efficiency gains.
Steam engines got more efficient. Britain used more coal, not less. Ford’s assembly line made cars affordable. People didn’t travel the same distances more cheaply — they travelled vastly more, built suburbs, created industries that hadn’t existed, and consumed more fuel than the horse-and-cart era could have imagined.
Apply this to creative and knowledge work. When the cost of producing digital content collapses — images, copy, layouts, code, campaigns, prototypes — the result isn’t less production. It’s vastly more of everything. More content. More campaigns. More variants. More brands in the market with more assets than they could have produced a year ago.
The volume doesn’t stabilise. It multiplies.
Here’s where it gets interesting. The mini skirt was shocking when it appeared. Then it wasn’t. Then it was just a skirt. The same normalisation cycle runs on every new signal. Early AI-generated content triggered immediate recognition — that’s fake, that’s cheap, that’s AI. The line was clear and it functioned as a repellent. Time passes. The outputs get dramatically better. We habituate. The line gets thinner. What was a distinguishing signal becomes background.
This is what Jevons predicts at the cultural level: abundance accelerates normalisation. The half-life of any specific differentiation strategy shortens as production costs fall. Which means the judgement required to stay ahead of the normalisation curve — to recognise when a signal is degrading before everyone else does — is precisely what appreciates in value.
Baumol tells you what becomes scarce when productivity rises.
Jevons tells you what happens to the abundance productivity creates, and how fast it normalises.
Together they describe the current moment with uncomfortable precision: as output gets cheaper, the expensive part becomes deciding what deserves to exist.
What Scarcity Looks Like Now
If output is cheap, what becomes valuable?
The easy answer — “creativity” or “human touch” — isn’t precise enough to be useful. AI produces novelty. It combines things that haven’t been combined before. Human touch is easy to perform and increasingly hard to verify.
The more useful answer comes from looking at what the abundance has actually made visible as scarce.
Judgement. The ability to look at ten outputs and know which one is right — and why — in ways that survive context change. Not aesthetic preference. Structural understanding of what serves the actual goal in this specific situation for this specific audience at this specific moment. This doesn’t scale the way execution scales. Every attempt to systematise it produces a slightly degraded version.
Coherence. Brand systems with clear guidelines have partially solved this at the execution layer — AI can follow a visual language, maintain a tone of voice, stay within approved parameters. The gap that remains is one level up. Guidelines constrain what’s anticipated. Every novel situation is a judgement call about which principle applies, how competing priorities are weighted, what this specific context requires from the brand right now. More critically: a brand system is only as coherent as the thinking that built it. AI can execute a coherent system. It can’t yet recognise when the system itself has drifted out of alignment with the market reality it’s supposed to serve. That recalibration is judgement at the strategic level, not execution.
Trust. When anyone can produce anything at any polish level, the signals that used to indicate quality have become unreliable. New credible signals require genuine investment to produce — which is exactly what makes them credible. Trust becomes harder to build and more valuable when built.
Clarity. In an age of frictionless execution, building the wrong thing with great speed is increasingly expensive. The ability to define what the right thing actually is — before building it — has always been undervalued. It’s now a competitive advantage.
The Work Nobody Sees
There’s a dimension to this that the productivity conversation almost always skips.
The public narrative assumes AI has automated the hard parts of professional work. Diagnosis, lesson planning, content generation, data analysis. These produce visible outputs. They look like the hard parts.
But consider what actually makes a doctor effective — not only the diagnosis itself, where AI is increasingly capable of assisting with pattern recognition, but the read of a patient who is minimising symptoms because they’re afraid of what they might hear. The decision to slow down in a consultation that technically doesn’t require it. The accumulated sense that something doesn’t add up, which sends the investigation in a different direction before anything in the data has confirmed it. None of that is in the chart. It doesn’t bill. It doesn’t appear in the output. But it’s often the difference between what gets found and what gets missed.
Or consider a teacher who restructures how they’re explaining something mid-lesson — not because the curriculum called for it, but because they read the room and recognised that the standard approach wasn’t landing for this particular group of students on this particular day. That adaptation produces no artefact. It doesn’t show up in any metric. But it’s what students remember twenty years later when they try to explain why one teacher changed something for them and others didn’t.
Or the live performance — the musician, the comedian, the theatre company — that adjusts the pacing, holds a silence longer than rehearsed, reads the collective energy of the room in ways that make the difference between a technically correct performance and one that people talk about afterward. The visible output is the same show. The invisible work is why one night feels like something happened and another doesn’t.
This is the invisible infrastructure that runs through every high-stakes human endeavour — the coordination, interpretation, contextual reading, and accumulated judgement that makes everything else function. It doesn’t scale the way execution scales because it’s inherently situated, inherently relational, and inherently dependent on understanding that can’t be fully documented without losing what makes it useful.
AI reduces the cost of production, but not the cost of interpretation. Across these fields, the visible layer has become cheaper and faster. The invisible layer has become more load-bearing as a result — which means more valuable, which means Baumol was right again in a way he couldn’t have anticipated.
The Law That Breaks the Dashboard
There’s a third law worth bringing into the room, and it may be the most unsettling of the three. It explains why optimisation can destroy the thing it claims to improve.
Goodhart’s Law simply states: when a measure becomes a target, it ceases to be a good measure.
Every metric used to evaluate digital performance — click-through rates, conversion rates, engagement scores — is subject to this. The moment a metric becomes the goal, behaviour optimises toward the metric rather than toward what the metric was supposed to indicate. The signal and the substance quietly diverge.
This has always been true. AI makes it more acute. When optimisation is cheap and fast, Goodhartian collapse accelerates. Campaigns optimised purely toward measurable proxies at machine speed can produce metrics that look healthy while the underlying reality they’re supposed to represent degrades.
The judgement required to notice when this is happening — to distinguish a metric moving from genuine value being created — is exactly the kind of contextual, experience-dependent understanding that Baumol says gets more valuable as everything else gets cheaper.
The three laws aren’t independent observations. They form a system.
What Survives the Flood?
Taken together, Baumol, Jevons, and Goodhart predict something specific about the next several years.
Production becomes a baseline, not a differentiator. Competing on production speed or output quality alone is competing in a market where the floor keeps rising and the ceiling keeps lowering.
The work that can’t be made cheap becomes the work that matters most. Judgement, strategic coherence, contextual intelligence, the ability to define the right thing before building it — these appreciate in value as the productivity explosion continues.
Normalisation cycles will keep shortening. What differentiates today becomes expected tomorrow faster than it used to. Reading that curve accurately — before the market has confirmed it — is itself a form of scarce judgement.
Metrics will keep lying, and reading them honestly will keep requiring pattern recognition that doesn’t optimise itself into blindness.
None of this makes AI less important. It makes the human layer of AI-integrated work more important — not as a moral claim, but as an economic prediction.
The laws aren’t new. The terrain they’re describing is.
Three old laws walked into an AI revolution and somehow remained the least confused in the room. They’re still standing.
© Alexandra Mark