When Intelligence Gets Cheaper, What Becomes Scarce?
Summary. AI is rapidly reducing the cost of analysis, prediction, monitoring, software creation and some forms of decision-making. As those capabilities expand, other limitations become more consequential. But those limitations are not all the same. Some are capacity constraints that can be replenished quickly. Others depend on physical infrastructure, specialized expertise or institutional change that may take years. Some are demand constraints: the enterprise can produce more, but the market does not need more. Others are created by management itself through policies, incentives or organizational design. The strategic challenge is therefore not simply to ask what becomes scarce next. It is to identify what now limits economic value, understand what kind of constraint it is, determine how persistent it will be and who controls it, and redesign the system accordingly.
Consider what initially looks like a straightforward AI productivity story. HSBC
Those numbers matter together. Better detection alone could create another problem. If a more capable system generates more suspicious activity for humans to investigate, the organization may simply exchange a detection constraint for an investigation constraint. Local productivity improves while the surrounding system becomes overloaded.
HSBC appears to have improved both the signal and the system required to absorb it.
Now compare that with the infrastructure supporting AI itself. In the PJM electricity market, the average time from an interconnection application to commercial operation increased from less than two years in 2008 to more than eight years in 2025. The technology creating demand for electricity may advance in months while the infrastructure required to supply it advances in years. (RMI)
A bank investigation queue and an electricity grid appear to have little in common. Economically, however, they expose the same management problem.
Technology can expand one productive capability faster than the surrounding system can adapt.
What happens next depends on what actually becomes constrained.
The economics are established. The management problem is changing.
No need to invent new economics to explain this phenomenon.
Alfred Marshall introduced the concept of quasi-rents to describe returns generated by man-made productive factors whose supply is temporarily fixed but can expand over a longer period. The time horizon matters. What looks scarce today may cease to be scarce once supply has time to respond. (OUP Academic)
David Teece added a different question in his 1986 paper, Profiting from Technological Innovation. An innovator does not necessarily capture the economic returns from an innovation. When imitation is relatively easy, returns can accrue instead to the owners of specialized complementary assets required to commercialize it. (ScienceDirect)
Operations theory approaches the problem from another direction: identify what limits the system's performance, improve it, and recognize that the limiting factor can change as the system changes.
AI brings these ideas together at unusual speed and breadth. The capability becoming cheaper is not merely storage, communications, or one manufacturing process. AI affects analysis, prediction, pattern recognition, language production, software development, monitoring, diagnosis, and increasingly some forms of decision and execution.
In earlier Alpha Decisions work, I have described the resulting phenomenon as Constraint Migration: when technology changes relative scarcity, relaxing one constraint can make another economically consequential.
But that statement is only a starting point. Saying that “the bottleneck moves” does not tell management enough.
The more useful questions are: What now limits economic performance? What kind of constraint is it? How long is it likely to remain consequential? Who controls it?
Those distinctions matter considerably.
The constraint set changes
Constraint Migration can easily be imagined as a serial process:
Constraint A → Constraint B → Constraint C.
Real enterprises rarely behave so neatly.
Several constraints can become near-binding at the same time. One may adapt quickly while another persists. A third may become visible only after the organization reaches greater scale. Sometimes the new limitation is not a shortage of productive capacity at all.
The better unit of analysis is therefore the constraint set surrounding the economic outcome the organization is trying to produce.
When AI materially expands a capability, the next limitation generally falls into one or more of four categories.
Capacity constraints occur when some part of the system cannot absorb or execute the additional work. Examples include electricity, testing, production capacity, specialized expertise, human review, or distribution.
Demand constraints occur when additional output can be produced but cannot be economically absorbed. The organization has made production more abundant without increasing demand.
Authority or institutional constraints occur when the technical capability exists, but the organization cannot safely, legally, or legitimately allow the system to exercise it.
Management-design constraints arise from the system management itself: objectives, incentives, organizational boundaries, policies, handoffs, or operating rules that prevent the new capability from becoming economic value.
These categories are not mutually exclusive. An enterprise may encounter several simultaneously.
But distinguishing among them matters because they require very different responses. More investment may relieve a capacity constraint while doing nothing about demand. Better governance may expand usable authority without creating another customer. Changing an incentive or workflow may solve a management-created constraint without adding any new productive capacity.
That produces an important reminder for AI strategy:
Capacity created is not value realized.
Sometimes the market is the constraint
Generative AI has made content creation extraordinarily inexpensive. That does not mean the world has created additional attention at the same rate.
HubSpot's 2026 survey of more than 1,500 marketers found that 71% said AI helps them create significantly more content. At the same time, 53% reported difficulty making that content stand out in an AI-saturated market, and 52% believed AI had made content so easy to create that it had become less effective overall. (HubSpot Blog)
Gartner found a related signal from the demand side. In a March 2026 survey, 49% of U.S. consumers said generative AI had made the quality of available content worse; that figure rose to 57% among Gen Z and millennials. Gartner summarized the problem succinctly: AI is increasing the volume of media consumers encounter, but not necessarily its value. (Gartner)
These surveys do not prove a causal economic relationship between AI-produced content and declining marketing returns. But they illustrate an important possibility.
AI relaxed the production constraint.
It did not replenish human attention.
If content becomes abundant while attention remains finite, producing still more content may make differentiation, credibility, distribution, and relevance more valuable. The appropriate management response is not to replenish content-production capacity. That capacity is already abundant.
This is why the first question after an AI productivity improvement should not be, What do we automate next?
It should be, What prevents this additional capability from becoming additional economic value?
The next constraint is not necessarily the durable constraint
Software development provides an unusually useful example.
Google's 2024 DORA research found that a 25% increase in AI adoption was associated with improvements in documentation quality, code quality, and code-review speed. Yet that same increase was associated with an estimated 1.5% reduction in software-delivery throughput and a 7.2% reduction in delivery stability. Improving development activity did not automatically improve software delivery. (Google Cloud)
It would be easy to conclude that AI made code abundant and therefore made human verification scarce.
The 2025 evidence makes the story more interesting. DORA later found that AI adoption became positively associated with software-delivery throughput and product performance, while its relationship with delivery stability remained negative. Google interprets this as organizations learning to integrate AI, but greater development velocity continues to expose downstream weaknesses in testing, version control, and feedback systems. (Google Cloud)
Meanwhile, the market began attacking parts of that constraint set. In December 2025, Cursor announced an agreement to acquire Graphite. Cursor's explanation was unusually revealing: as writing code got faster, reviewing changes, merging them safely, and collaborating effectively increasingly became bottlenecks to production-grade software. (Cursor)
This does not establish that “verification” was replenished within a year.
It establishes something more useful: verification is not one thing.
Code review, automated testing, integration, security validation, architectural judgment, and accountable risk acceptance are different capabilities. They need not replenish at the same rate.
That preserves a narrower but more consequential proposition:
Some human verification is transitional scarcity.
Where verification consists primarily of repeatable checks against observable criteria, software and AI can attack the constraint. Where it depends on tacit expertise, ambiguous judgment, legal accountability, or authority over irreversible consequences, replenishment may be much slower.
The fact that humans become the next bottleneck does not necessarily mean humans become the moat.
That is an empirical prediction, not a permanent claim about human superiority. If AI systems increasingly automate review, testing, and verification while consequential judgment remains hard to scale, we should see the composition of residual human work shrink but become systematically harder.
The same technology produces different results in different systems
DORA also provides something company anecdotes usually cannot: comparative evidence across organizations.
Its 2025 research drew on nearly 5,000 technology professionals and examined which organizational capabilities change the relationship between AI adoption and performance. Quality internal platforms amplified AI's positive influence on organizational performance. A strong user-centric focus amplified AI's effect on team performance; without that focus, greater AI adoption could have a negative effect. Mature version-control practices and small-batch work similarly changed how effectively organizations absorbed greater AI velocity. (Google Cloud)
This is not a controlled matched-pair experiment showing two otherwise identical companies making different constraint decisions. We should not pretend that evidence exists.
But it is stronger than collecting successful corporate narratives.
The technology itself does not determine the result. The surrounding system changes the technology's economic effect.
That is precisely the problem Constraint Migration is intended to make visible.
Investment response is not capacity response
Once you identify a genuine capacity constraint, another mistake becomes easy.
A valuable scarcity attracts capital. That tells us people have recognized the constraint. It does not tell us how quickly they can remove it.
Power infrastructure makes the distinction vivid.
AI-related electricity demand has stimulated enormous investment in generation and grid equipment. Siemens Energy announced a $1 billion U.S. expansion in 2026, including additional gas-turbine and grid-equipment capacity. Yet the delivery time for large gas turbines had stretched beyond five years. Developers were reserving manufacturing slots years before projects were otherwise ready to proceed, even as manufacturers expanded capacity. (Reuters)
This distinction is fundamental:
Investment response is not capacity response.
Capital can move at the speed of financial markets. Usable productive capacity may move at the speed of manufacturing, construction, permitting, credentialing, organizational learning, or physics.
For management, this suggests a more precise concept than simply calling something “scarce.”
I use replenishment rate as shorthand for how much usable capacity can be brought online, by whom, at what cost, within the relevant decision horizon.
The word usable matters.
Another $100 million committed to turbine manufacturing is not electricity available tomorrow. A hundred new job requisitions are not a hundred experienced engineers. A new policy document is not necessarily institutional confidence to delegate a consequential decision to a machine.
The important diagnostic, then, is not simply whether capital or competitors are attacking a constraint. What matters economically is how responsive usable capacity is to additional resources over the period.
A six-month scarcity requires a different strategy from a constraint likely to persist for a decade.
Better operators anticipate where constraints will emerge
Autonomous mining offers another lesson, but not the one a simple bottleneck story would suggest.
It would be tempting to tell Rio Tinto's experience as a sequence: automate trucks, discover another bottleneck, then create a remote operations center.
The historical record shows something more sophisticated.
When Rio announced its Mine of the Future program in January 2008, it presented driverless trucks, autonomous trains, remotely operated drills and a Remote Operations Center in Perth as building blocks of an integrated mine-to-port operating system. The center was intended to oversee mines, processing plants, rail, ports and even power. Rio explicitly contrasted its approach with competitors it said were focusing on discrete technologies rather than modernizing the whole operation.
Rio was not simply chasing a bottleneck after it moved.
It was anticipating what successful automation would do to the surrounding production system.
The lesson is stronger:
Good operators do not merely respond to Constraint Migration. They design for where constraints are likely to migrate if the technology succeeds.
For Rio, the relevant system boundary was not the truck. It was the mine-to-port production system.
The same principle applies to knowledge work. If AI radically accelerates software creation, the transformation boundary may need to include review, testing, architecture, deployment, product management, and customer adoption. If AI changes customer service, the relevant system may extend into fulfillment, pricing, retention, and customer economics.
Optimize the mine, not the truck.
Governance can become productive capacity
Some successor constraints are institutional rather than physical.
An AI system may technically be capable of taking an action while the enterprise is unwilling to authorize it. In that situation, the limiting factor is no longer intelligence alone. It is the organization's ability to allocate machine decision rights safely.
MIT CISR's 2026 research provides useful evidence. Based on 30 interviews with 27 executives at nine global companies, researchers developed a framework for allocating decision rights between humans and AI according to ambiguity and risk. Their detailed example is One New Zealand, which had more than 50 AI agents operating in early 2026. (MIT CISR)
One NZ's treatment of those agents is not uniform. Its AI agents resolve more than 60% of in-app customer questions. For consequential actions such as acting on behalf of a customer, leaders define explicit risk tolerances and guardrails. The company introduces agents to small cohorts, validates outcomes and adjusts human involvement according to the consequences of error. (MIT CISR)
More importantly, decision rights evolve.
Marketing teams initially review AI-generated audiences before campaigns launch, but MIT CISR reports that autonomy can increase as confidence grows. In network operations, One NZ has introduced progressively more closed-loop automation. The company combines these changes with a Responsible AI policy, lifecycle guardrails, a standardized “Trust Trail,” and a named business owner accountable for every deployed agent. (MIT CISR)
This is closer to the mechanism we need.
Governance is not simply preventing the AI from acting. Governance helps determine which actions the enterprise becomes willing to delegate.
That supports a more precise proposition:
Governance becomes productive capacity when stronger controls let the enterprise delegate a wider range of consequential machine action without proportionally increasing human gating or unacceptable risk.
The claim has limits. One NZ does not prove that every investment in governance creates economic value, nor does it establish a universal causal relationship between stronger governance and greater autonomy.
But it demonstrates the mechanism. Decision rights, risk classification, observation, accountability, and staged delegation can change the operating envelope available to AI.
That reframes a familiar debate. Security and governance are usually described as constraints on AI. Under some conditions, they are also capabilities that enable more productive AI use.
A constraint is not the same thing as an asset
Identifying a durable scarcity creates another potential error.
Scarcity is not automatically valuable to the enterprise experiencing it.
Consider electricity. For a hyperscaler waiting years for additional power, scarce interconnection capacity is a constraint. It delays deployment and destroys potential value. For the owner of difficult-to-replicate generation capacity, scarcity may increase the value of an asset. For an infrastructure provider that can help customers overcome the constraint, the same scarcity creates a market opportunity.
The underlying scarcity is the same. The economic position is different.
This is where Teece becomes particularly useful.
A constraint tells you what limits value. Control of a scarce complement helps determine who captures value.
Durable scarcity therefore does not automatically mean “invest here.” If the enterprise depends on a slow-replenishing complement controlled by somebody else, durability may be exactly what makes the situation dangerous.
Management needs to determine its economic position relative to the scarcity.
If it is constrained by it, the response may be to expand capacity, substitute around it, reduce dependence on it, contract for it, or redesign the system.
If it controls the scarce complement, ownership may create bargaining power and strategic advantage.
If it can resolve the constraint for others, the scarcity may create a product or platform opportunity.
These distinctions matter particularly for AI vendors.
The uncomfortable implication for AI software vendors
A paradox hides inside the current AI market.
A vendor can identify a real customer bottleneck, build software that resolves it—and simultaneously destroy the scarcity that made the opportunity attractive.
Code review illustrates the problem. Faster code generation made review more consequential. That attracted technology, capital, and competitors. If software can expand review capacity quickly, the scarcity may replenish quickly as well.
The customer benefits.
The vendor does not necessarily earn durable rents.
Teece's framework predicts why. Resolving the constraint is not enough if the solution is readily imitated and the vendor controls no difficult-to-reproduce complementary assets. Durable returns are more likely when the offering is coupled to specialized capabilities competitors cannot easily replicate: proprietary context or data, deep workflow integration, distribution, trusted enterprise relationships, regulated credentials, switching costs, network effects, operational infrastructure or other co-specialized assets. (DOI)
That leads to an uncomfortable proposition for today's AI software industry:
Many AI vendors may be competing to solve precisely the constraints that commoditize fastest.
The strategic question for a vendor is therefore not merely: Which customer bottleneck can we solve?
It is: Can we maintain an advantaged position after we solve it?
There may be another route. A platform can create advantage not because one bottleneck persists forever, but because the platform can follow the customer's constraint as it moves—across data, workflow, authority, security, verification, execution, and observation.
The durable capability then resides less in one point solution than in the ability to absorb successive constraints.
That is a much higher bar than attaching AI to a software category.
What would make this framework wrong?
Constraint Migration becomes meaningless if every change in relative scarcity is automatically called a strategically important new constraint.
It needs an economic test.
If AI expands a productive capability and the enterprise can convert that additional capacity into its economic objective without another limitation materially restricting the result within the relevant system boundary and decision horizon, then no strategically consequential successor constraint exists to manage.
Take the benefit.
Likewise, if a supposed successor constraint disappears quickly once ordinary software, labor, or capital is applied, management should not treat temporary scarcity as a durable strategic asset.
And if differences in surrounding capabilities do not materially alter the economic results organizations achieve from AI, then the argument that Constraint Migration requires system-level redesign becomes much weaker.
The evidence today points in the other direction. DORA finds that surrounding technical and organizational capabilities materially change the association between AI adoption and performance. PJM demonstrates that some physical capacity adjusts vastly more slowly than digital capability. Rio demonstrates that sophisticated operators can anticipate system effects rather than automate components independently. One NZ demonstrates that decision rights and governance can evolve with machine capability.
None of these proves a universal law.
They establish a sufficiently plausible mechanism to make a different management discipline worth testing.
A different sequence for management
Most enterprise AI strategies still begin with a technology question:
Where can we use AI?
That naturally generates a portfolio of use cases.
A constraint-oriented strategy starts somewhere else:
What economic outcome are we trying to improve, and what prevents us from producing more of it?
The management sequence then becomes:
What economic outcome are we trying to increase, and is there demand for more of it?
What constraint—or set of near-binding constraints—limits that outcome today?
If AI materially relaxes one of them, what becomes limiting: capacity, demand, authority or institutional constraints, management design, or several simultaneously?
Where capacity is constrained, how much usable capacity can be added within our decision horizon, and what prevents it from expanding faster?
Who controls that capacity: decision right or scarce complement?
What is our economic position relative to it: a constrained user, an owner of a valuable complement, or a provider capable of resolving the constraint for others?
What action converts the technological improvement into economic value: replenish, substitute, contract, own, govern, redesign, stimulate demand—or expand the system boundary itself?
This is a different way to allocate capital.
Investing heavily in a bottleneck that software will eliminate next year can destroy value. So can failing to secure a complement that will remain difficult to expand for a decade. Producing more when demand is already constrained can worsen economics. Preserving human review around activities that machines can reliably verify can strand productivity. Removing human authority from decisions whose consequences remain difficult to bound can create risk faster than value.
Governance that looks like overhead locally can become productive infrastructure when it enables greater machine autonomy at the system level.
When intelligence becomes abundant
For most of modern economic history, human intelligence has been expensive and difficult to scale.
Organizations were built around that assumption. Expertise was concentrated. Analysis took time. Information moved through hierarchies. Decision rights were often bundled with positions because the capability to interpret context and exercise judgment could not easily be separated from the person doing the work.
AI changes that assumption.
But it does not repeal economics.
Intelligence remains one productive input in systems that also require infrastructure, demand, capital, context, trust, authority, verification, distribution, and management. Making intelligence cheaper changes the relative scarcity of those other inputs—and therefore changes what management needs to pay attention to.
No permanent list of what becomes scarce after AI exists.
Human verification may become consequential and then partly automate. Compute may become scarce until capacity expands. Electricity may remain constrained much longer because physical infrastructure responds slowly. Authority can be changed rapidly in one institution and painstakingly in another. Content can become nearly free to produce while human attention becomes harder to earn.
The strategic task, then, is not to predict the next fashionable scarcity.
It is to understand the enterprise's changing constraint set.
Three questions capture the discipline:
If intelligence becomes abundant, what in our system cannot keep up? How long will it stay that way? Who controls it?
The first identifies what limits value.
The second distinguishes a temporary bottleneck from a strategically consequential scarcity.
The third determines whether that scarcity is a problem to relieve, a complement to secure, an asset to own, or an opportunity to solve for somebody else.
That is where AI strategy stops being primarily a technology-adoption problem and returns to something management has always had to do: allocate scarce resources, redesign systems, and capture economic value as the underlying economics change.
Comments