The Jevons Paradox of Agentic AI

The dominant economic argument for agentic AI is replacement.

If an agent can perform work that previously required a person, the cost of that work falls. Fewer people are required. Revenue per employee rises. Margins improve. That will happen.

But it may not be where the largest companies are created.

When the cost of performing a unit of knowledge work collapses, companies may not simply perform the same amount of work with fewer people. They may consume dramatically more work.

That is the Jevons paradox of agentic AI.

And it changes what investors should be looking for: not how much labor an agent replaces, but how much rationed work it makes rational, where that effect is strongest, and who captures it.

Efficiency does not necessarily reduce consumption

In the nineteenth century, economist William Stanley Jevons observed that improvements in the efficiency of coal use did not reduce coal consumption. They made coal economically useful in more applications, increasing aggregate demand.

The mechanism matters more than the historical example.

When the cost of consuming a resource falls sufficiently, previously uneconomic uses become economical. Demand expands.

But the mechanism has a precondition that most retellings skip. Jevons only fires when demand is elastic enough that consumption grows faster than efficiency cuts cost. Where demand is fixed, efficiency simply shrinks the bill.

Most analysis of agentic AI quietly assumes fixed demand. An agent performs work a human used to perform, so the company needs less human labor. Same work, fewer people.

Why should the amount of work stay fixed?

There is already an enormous reservoir of knowledge work organizations would perform today if human capacity were not the constraint. Most of it isn’t skipped because it is worthless. It is rationed because human attention is expensive.

Agentic AI moves that constraint. The interesting question is not whether induced demand exists. It is where it is elastic, and where it hits a wall.

Where Jevons effects should be strongest

Induced demand is not uniform. It concentrates where four conditions hold at once.

The work was rationed by cost, not by value. The test is the gap between what an organization would do with unlimited capacity and what it actually does. Where that gap is wide, the reservoir is deep. Where teams already do roughly everything worth doing, cheaper work just means cheaper work.

The output is not consumed by someone else’s scarce attention. This is the condition most theses miss. Some work is consumed internally, by the organization’s own systems and decisions: research, preparation, monitoring, reconciliation, coverage. Other work is consumed by a counterparty whose attention did not get cheaper: outreach, pitches, follow-ups, applications. The first kind can expand almost without limit. The second kind expands until it saturates the counterparty, and then it becomes an arms race.

The next bottleneck is far away. When work gets cheaper, the constraint does not vanish. It moves. If every unit of expanded output still needs a human to review, approve or act on it, the human becomes the new ration and the Jevons curve flattens early. The effect is strongest where output is verifiable by systems, cheap to be wrong about, or feeds a decision a human was going to make anyway.

There is a long tail below the human threshold. Every organization has units of work individually too small to justify a person: the low-value account, the minor contract clause, the weak signal, the dormant opportunity. Collapsing cost does not make these units slightly better served. It makes them served for the first time. Threshold effects produce step changes, not increments.

Where all four hold, efficiency produces expansion.

Where the second fails, it produces something else: a Red Queen race. Everyone runs faster to stay in the same place. The work volume rises, the outcome does not, and the gains are competed away into a higher baseline.

That baseline still matters. Table stakes are exactly what earns a budget line. But it is a different thesis from growth, and investors should know which one they are underwriting.

What happens when revenue work becomes cheap

Revenue is a useful test case because it contains both kinds of work inside one motion.

Imagine every seller receives a persistent, customized executive assistant. It knows the seller’s territory, accounts, pipeline, stakeholders, conversations and commitments. It prepares calls, follows deals, maintains digital sales rooms, contextualizes content, scores opportunities, tracks contracts, remembers commitments and surfaces neglected accounts.

The first-order argument is obvious: the EA saves the seller time.

The second-order question is what becomes rational once all of that gets cheap. And here the four conditions split the motion cleanly.

Preparation, account research, stakeholder mapping, commitment tracking, deal hygiene, coverage of dormant accounts: this work is consumed internally. It was rationed by cost. Most of it sits below the human threshold on any given account. Its output feeds decisions the seller was already making.

Every condition holds. This is where the revenue system can become genuinely higher-resolution: every important call prepared, every account researched, every commitment kept.

Outbound volume is different. A seller can now send ten times the follow-ups, but the buyer’s inbox did not get ten times larger. Here efficiency hits a fixed counterparty, and the Jevons effect collapses into a Red Queen race. If everyone automates outreach, outreach becomes noise, and the advantage migrates back to relevance, which is itself built from the internal work.

So the durable expansion in revenue is not more selling at the buyer. It is more knowing inside the seller’s organization, converted into fewer, better moments with the buyer.

That distinction should change what gets built, and what gets funded.

Who captures the induced demand

In Jevons’ England, the volume went to whoever sold the coal.

That is the uncomfortable default for the application layer. In agentic AI, the resource being consumed is cognition. If induced demand runs through a thin product, the expansion accrues to whoever sells the tokens. The application becomes a conduit: more work flows through it, and almost none of the value stays.

So “does this agent unlock induced demand?” is necessary but insufficient. The sharper question is why the expansion would be captured at this layer rather than below it.

Two mechanisms answer that.

The first is pricing. A product priced per seat captures none of the expansion it creates: the seller count stays flat while the work per seller multiplies. A product priced on usage, coverage or outcome scales with the induced demand itself.

The second is context. Tokens are fungible. Accumulated knowledge of a specific organization’s accounts, commitments, permissions and history is not. When the expanded work depends on context that lives inside the product, the value of each additional unit of work accrues to whoever holds the context, not whoever supplies the cognition.

Pricing determines whether a product can capture the expansion. Context determines whether it deserves to.

And context, accumulated long enough, turns into something else.

Infrastructure is an achieved state

The AI market currently collapses very different companies into the same category: wrappers, point solutions, platforms, agents. They are not the same economic objects.

But even those distinctions are insufficient. “Platform” describes product architecture. Infrastructure describes what the product has become to the customer.

Slack is a useful example. Organizations may evaluate Slack against competitors, negotiate price or consolidate vendors. But mature customers are no longer reconsidering whether workplace communication deserves to exist as an organizational capability. The category has budget presumption.

Nor did generative AI make Slack conceptually obsolete because AI can write messages. AI operates through communication infrastructure.

That gives us a test for agentic companies:

If AI automates the human activity your product currently supports, does your product disappear, or does AI begin operating through it?

The second position is considerably more durable.

Infrastructure survives the automation occurring around it because the organization has begun assuming its existence. Workflows depend on it. Permissions depend on it. Context accumulates inside it. Governance is built around it. Other systems integrate with it.

That is more than switching cost. It is organizational embeddedness.

The question is no longer how painful migration would be. It is how much of the organization’s operating architecture would have to be reconsidered if the product disappeared.

From agent to organizational assumption

Return to the executive assistant for every seller.

Initially, it is purchased as an AI capability. But imagine it works. Trust grows. It receives access to more context, then more systems, then more workflows. Its jurisdiction expands.

The seller begins relying on it continuously. Managers begin assuming sellers have that capacity. RevOps designs processes around it.

Enablement assumes knowledge can be delivered through it. Leadership resets expectations for account coverage, preparation and follow-through, because the old human capacity constraint has moved.

Eventually the operating assumption becomes: every seller works with persistent synthetic capacity.

At that point, removing the product does not just make sellers less productive. The organization has to reduce the amount or resolution of work it attempts, or redesign the revenue motion.

This is where the Jevons layer and the infrastructure layer meet. The induced demand is what the organization rebuilds itself around. The product that carries it becomes the thing it cannot remove.

The three bubbles an agentic company must survive

Durable agentic companies have to survive at least three
transitions.

The AI bubble. Eventually “AI” stops creating novelty value by itself. Customers stop buying because something is agentic and return to buying because something is necessary. Infrastructure survives because its value no longer depends on category excitement.

The build-versus-buy bubble. As models and tooling improve, enterprises will increasingly ask why they cannot build agents themselves. For a shallow product, that question is fatal. For infrastructure, it changes shape. The customer may be able to recreate the original feature. Recreating the accumulated context, integrations, permissions, governance and organizational dependencies is a different proposition. The question becomes less “can we build this?” and more “do we want to own and continuously evolve this operating layer ourselves?”

The headcount-anchor bubble. Today, many agentic products are priced and valued against the salaries they replace. That anchor holds only while synthetic labor is novel. Once it is ordinary, “replaces three FTEs” stops commanding a premium, the way “replaces the typing pool” once did. Products whose value is expressed in headcount saved will see that value deflate with the price of cognition. Products whose value is expressed in work the organization can now attempt, and could not before, are anchored to something that grows.

The investment question

This changes diligence.

Do not ask only how autonomous the agent is, how many people it can replace, or how impressive the demo looks. Ask where the work sits on the Jevons map:

• Was this work rationed by cost, or simply not worth doing?
• Is the output consumed internally, or by a counterparty whose attention is fixed?
• When the work gets cheaper, where does the next bottleneck land, and how far away is it?
• Is there a long tail of units that were below the human threshold and are now served for the first time?

Then ask whether the product captures the expansion:

• Does pricing scale with induced demand, or with seats that stay flat?
• Does the expanded work depend on context that lives inside the product?

And finally, the test that settles it:

If you removed the product after two years of successful adoption, would the customer merely lose a useful tool, or would it have to redesign how work gets done?

Call it the Removal Test. The second answer means the product has become infrastructure.

The problem is that the Removal Test is a lagging indicator. It confirms winners after they have won, and diligence happens earlier. So watch for its leading signals, which show up within the first year:

Jurisdiction creep. Access to new systems and workflows is granted after deployment, not negotiated before it.

Expectation reset.lang=EN-US> Managers change coverage, preparation or response standards because the capacity is assumed.

Process rewiring.lang=EN-US> Operations teams redesign workflows that route through the product, rather than around it.

Usage outruns headcount. Work per user rises after the novelty period, not just at launch.

The build question goes quiet. After deployment, the customer stops seriously discussing building it internally, and can say why.

Each of these is the Removal Test arriving early.

The agentic infrastructure thesis

The causal chain runs in one direction.

Efficiency lowers the cost of work. Where that work was rationed by cost, consumed internally and far from its next bottleneck, demand expands. The organization attempts more work at higher resolution. It builds new operating assumptions around that capacity. Products that hold the context the expanded work depends on, and are priced to grow with it, accumulate trust and jurisdiction. Those products become organizational dependencies. Organizational dependencies become infrastructure.

Where the work is consumed by someone else’s fixed attention, the same efficiency produces a race instead of an expansion. Useful to win. Not the same thing to own.

So the biggest opportunity in agentic AI may not be replacing human labor. It is expanding the economically viable surface area of knowledge work, in the places where that surface can actually expand, and owning the infrastructure the expansion runs through.

The winners of the agentic era may be surprisingly difficult to recognize from today’s AI categories.

Five years from now, the strongest of them may no longer be thought of as AI companies at all.

They will simply be part of how the company operates.

And that may be the clearest signal that they won.