There’s a story technologists like to tell about AI agents: software will do most cognitive work humans currently do. Transaction costs, the friction of coordinating economic activity, collapse toward zero. Markets become faster, smarter, more efficient. Everyone benefits.

It’s tidy but conveniently missing about half the economics.

The moment you treat agents as an economic object, the familiar frameworks stop fitting. Agents aren’t workers with lower wages. They aren’t software with better features. The market they’re creating doesn’t behave like any market we’ve built institutions around.

In an endpoint when transaction costs drop so far that the logic of firms and employment starts to dissolve, the “Coasean singularity” named by the MIT and Harvard researcher, the supply and demand in Economics 101 and what people take for granted in labor economics become… strange.

Demand has two regions

Demand for AI agents splits into two structurally different regions:

Substitution demand: agents doing things humans currently do: drafting, summarizing, screening, supporting. This has a natural price ceiling: no buyer pays more for an agent than the human alternative costs. The whole substitution, or a more fashionable word today, displacement market trends toward commodity, and it is deflationary. As agents displace workers, human wages fall, and the ceiling they’re priced against falls with them.

Frontier demand: agents doing things humans couldn’t do at scale regardless of cost. Monitoring every transaction across a financial system in real time. Personalizing every customer interaction for millions of people simultaneously. Operating in hazardous mining environments. There was no human version to substitute. These get priced against the value of the outcome, not the cost of an alternative, hence no ceiling, no floor.

This echoes MIT’s research that identifies the two scenarios for agent deployment: making decisions of similar quality at dramatically lower cost, or making higher-quality decisions than humans. The first maps to Substitution demand. The second is Frontier.

The boundary between the two regions is not fixed. As agents improve, tasks that once required human judgment slide into the substitution category. The frontier keeps moving. The people on the wrong side of that line don’t get much warning.

Region 1 — Substitution demand
Agents doing what humans do
Hard price ceiling
Priced against the human alternative. No buyer pays more for an agent than the human costs. As agents displace workers, wages fall and the ceiling falls with them.
ceil = human wage (falling) future $low $high low replaceability high →
Examples
Drafting emails Screening CVs Customer support Summarising docs Basic data entry Report formatting
Recently crossed in
Legal doc review Radiology screening Code review
Moving
boundary
deflation spiral displace wages↓ ceiling↓ price↓ boundary shifts left over time
one-directional*
*assumes agents keep improving
Region 2 — Frontier demand
Agents doing what humans couldn't
No price ceiling
No human version existed at scale regardless of cost. Priced against outcome value — not a human alternative. No ceiling, no floor. Deflationary pressure is absent.
undiscovered discovery zone priced below true value $low $high low replaceability high →
Examples
Real-time fraud monitoring 1:1 personalisation at scale Hazardous environment ops Continuous system watch Drug interaction screening
Moving toward boundary
Advanced legal reasoning Surgical assistance
X-axis: capability vs. human baseline
how well agent matches human output quality at lower cost
shared dimension
replaceability →
X-axis: superior scale vs. human ceiling
tasks humans couldn't do at any cost — priced against outcome value
Assumptions:
Buyers can evaluate qualityWithout this, the ceiling mechanism doesn't close and substitution pricing breaks down
Migration is one-directionalTasks move Frontier → Substitution only, not back — assumes continued capability improvement
Frontier demand is discoverableBuyers eventually learn to see the outcome value — early frontier tasks are systematically underpriced
Marginal copy cost stays near zeroRegulatory or compute shifts could change this and introduce a supply-side floor
The boundary is always crisp at any moment — but it never stays put. Tasks once priced against the value of impossibility slide into the deflationary spiral with no warning.

Supply has a gap where the middle used to be

In a normal labour market, supply is continuous. Buyers choose along the spectrum based on need and budget.

AI agent supply doesn’t work this way.

At one end: the commodity tier. Once an agent is trained, copies deploy at near-zero marginal cost. Supply, at least in terms of quantity, is effectively unlimited.

At the other end: the differentiated tier. Agents with genuine domain depth, trained on proprietary data, fine-tuned on years of professional feedback. These are scarce, not because copying is expensive, but because the inputs that make them valuable are slow to accumulate. You can’t rush the training data and pile garbage data that makes a legal agent reliable in contested litigation.

Between these two tiers, there is a chasm. AI does not optimize for competent generalist. An agent that is almost as good as the best one is nearly worthless if buyers can identify the best one, exactly the same as why Google dominates the search engine market and confirmed by this NBER paper.

Tier 1 — Commodity
Infinite supply, near-zero marginal cost
Unlimited copies
Once trained, the model deploys at effectively zero cost. No capacity constraint — supply curve is nearly flat. Price pressure is purely downward.
Price Quantity supplied ~0 marginal cost floor future $$$ $ low unlimited → traditional supply curve
What it takes to build
Compute Open weights Prompt engineering API access + compliance cost + integration labour
The chasm
Middle tier collapses
market value capture 0 quality rank #1 ~75% #2 ~18% #3 near-zero value gravity
When buyers can identify best, second-best captures a fraction. Below that: nearly nothing.
Tier 2 — Differentiated
Scarce inputs, not scarce copies
Artificial scarcity
Domain depth from proprietary data and accumulated feedback. Copying the model is cheap — but the inputs that made it valuable took years. You can't compress that accumulation.
Price Quantity supplied $ $$$ few many → data moat strong now fragile 5–10yr? supply limit
What scarcity actually comes from
Proprietary data Expert feedback Domain fine-tuning Institutional trust synthetic data risk distillation / theft portability reg.
Supply curve: nearly flat (horizontal)
no capacity constraint — price floor set by compute + compliance
no continuous
middle tier
Supply curve: steep (upward sloping)
constrained by data and time — steepness reflects accumulation barrier
Assumptions:
Near-zero marginal costInference compute, compliance, and integration labour could raise the true floor
Buyers can evaluate qualityWithout this the winner-takes-most mechanism doesn't close; enterprise lock-in can sustain local middle-tier survivors
Data moat is durableSynthetic data, distillation, and portability regulation each erode it; moat is strong now, fragile over 5–10 years
Winner-takes-most, not allTop agent captures ~75% of value; second-best survives at thin margins; below that, near-zero
AI does not optimise for the competent generalist. The middle tier doesn't erode gradually — it collapses. What remains is a flat commodity floor and a steep differentiated peak, with nothing viable between them.

Zero friction is real.

Agents are, for most practical purposes, location-less. An agent trained in Tashkent deploys in Copenhagen at no incremental cost. For cognitive tasks, there is no Geneva wage premium anymore. There is just a global market.

Labour economics has always depended on friction. Workers can’t instantly move between markets. That friction sustains wage differentials and creates floors. Agents collapse the friction on the cognitive side entirely. Wage arbitrage that took decades in manufacturing, a.k.a. move production to cheaper labour, takes months or even shorter for knowledge work.