What Becomes Scarce Next?
How boards create enterprise and shareholder value when technology moves the constraint
Idea in Brief
The problem. Companies invest heavily in AI and measure success by local gains: more code, more leads, more analysis. Yet enterprise performance often barely moves.
Why it happens. When technology expands one capability, the binding constraint moves elsewhere, to testing, qualification, governance, or customer demand. Investment that keeps flowing to the original success lands on the wrong side of the new constraint.
The way forward. Boards should anchor on one objective, sustainable net value creation, and treat technology as something that changes relative scarcity. The recurring question isn't how much AI to buy, but what becomes scarce next.
Consider a software organization that gives its 100 engineers AI coding tools and raises coding productivity by 50%. Useful software output rises just 10%. The AI worked. The organization failed to capture the gain because the bottleneck had moved to testing, integration, and review.
That pattern is not a story about software, or even about AI. It happens whenever a technology sharply expands one productive capability: relative scarcity shifts, and economic value moves toward whatever is now scarce. Artificial intelligence is today’s technological shock. It should not become our management theory.
The shock is real. Machine intelligence is becoming more capable, more widely available, and more deeply embedded in the infrastructure where work already happens. The 2026 Stanford AI Index finds the leading model providers clustered in a narrow band on major evaluations, with competition shifting toward cost and reliability. Enterprise platforms are beginning to treat intelligence as a resource to allocate rather than simply acquire. Microsoft’s hybrid intelligence architecture, for example, routes each task between local and cloud models according to what the task requires.
In the first phase of generative AI, enterprises tried to acquire as much intelligence as they could. The next phase is about allocation: where intelligence belongs, how much is enough, which decisions should become autonomous, where human judgment still earns its cost, and which surrounding capabilities become strategically important. Leaders cannot make those choices well until they settle a prior question: What is the enterprise trying to optimize?
The Objective Is Value, Not Technology
Businesses do not exist to maximize technology, automation, intelligence, or autonomy. These are capabilities. They matter only insofar as they serve a larger objective: sustainable net value creation.
That objective has three parts worth keeping separate. Value creation is the benefit an enterprise contributes to customers and society through its products, services, decisions, and innovations. Value capture is the share of that benefit the enterprise retains through revenue, margins, and cash flow, enough to reward capital and keep investing. Value destruction is the harm created or exposed along the way: failed investments, cyber losses, operational breakdowns, unsafe decisions, regulatory failures, lost trust, and poorly governed automation.
Together, they give boards a better objective than maximizing any particular technology:
Create sustainable value for society, capture enough of it to sustain and grow the enterprise, and minimize avoidable value destruction.
Shareholder value sits naturally inside this formulation. Over any meaningful horizon, shareholder returns depend on the company continuing to create something others value and retaining a viable share of it. Technology matters because it changes what can be accomplished economically. It changes the feasible set; it does not change the objective.
Technology Changes Scarcity
The economics here are not new. Theories of constraints, organizational economics, innovation, and control point to the same principle: when one productive capability expands materially, the relative importance of other capabilities changes with it.
The mechanism runs as a cycle (Exhibit 1). A capability expands, and the constraint it addressed relaxes. Another resource becomes relatively scarce, and its economic importance rises. Management reallocates capital and redesigns the system. Eventually the newly scarce resource expands, and the constraint moves again. I call this cycle Constraint Migration.
Exhibit 1. Constraint Migration Cycle
When technology relaxes one constraint, value shifts toward whatever is newly scarce, and management must reallocate before the cycle turns again.
The claim is not that another bottleneck always exists; that would be trivial and impossible to disprove. Many local bottlenecks matter little strategically, some disappear quickly, and a new technology can expand several capabilities at once. A successor constraint matters strategically only when it materially limits the enterprise’s ability to create, capture, or protect value and cannot be expanded economically before the damage is done.
That shifts the management question from “Where is the bottleneck?” to a more useful one: What has become economically consequential because the previous constraint was relaxed? At bottom, that is a capital-allocation question.
When Successful Technology Moves the Constraint
Return to the software organization. Before adopting AI development tools, its 100 engineers can specify, write, review, test, and release about 100 units of useful work a week. Coding capacity is a real limitation.
AI then raises coding productivity by 50%, so the team can theoretically produce 150 units. But testing, integration, architecture review, security, and release can absorb only 110. Review queues grow. Integration testing that took two days now takes six. Product managers cannot specify enough valuable work to use the new coding capacity. Useful output rises from 100 units to 110.
Put a price on it. If each released unit is worth $10,000, management expected roughly $500,000 a week in new value and received about $100,000. The other $400,000 of theoretical capacity is not merely idle. The excess code piles up as work in process, which carries its own costs: rework, merge conflicts, stale code, coordination overhead, and technical debt.
Exhibit 2. The AI worked; the system absorbed a fifth of it
Measure | Before AI | Management’s expectation | Actual result |
Coding capacity (units/week) | 100 | 150 | 150 |
Useful output released (units/week) | 100 | 150 | 110 |
Incremental value ($/week) | Baseline | $500,000 | $100,000 |
Binding constraint | Coding | Coding, assumed | Testing, integration, review |
Hypothetical organization of 100 engineers; each released unit assumed to be worth $10,000.
The problem is sharper than diminishing returns. Once coding capacity exceeds what the rest of the system can absorb, the next dollar spent on coding returns close to nothing.
Management can test this directly. Add more coding agents, and output stays near 110. Instead, expand testing or integration capacity; if output rises materially, the constraint has demonstrably moved. The right conclusion is not that the AI investment failed. Its success changed the system's economics.
The same mechanism shows up well beyond software. Suppose AI lets a sales organization generate three times as many promising leads, while account executives can follow up on only 20% more. Lead generation has improved, but the constraint has moved to qualification, selling capacity, customer attention, or implementation. More lead-generation intelligence adds almost nothing until you address the next constraint. The more successful an intervention is at expanding one capability, the more important it becomes to look somewhere else.
Constraint Migration Leaves Evidence
A theory that executives cannot observe is of little use. A newly consequential constraint leaves signals. Queues build in front of a process, waiting times rise, and service levels slip. Exception volumes increase. A resource runs near practical capacity, customers start waiting, and the marginal value of one more unit climbs. Over time, bargaining power and margins may shift toward whoever controls it.
There is also a simple experimental test. Keep adding capacity to the resource management believes is still the constraint. If system output barely responds, the enterprise is investing on the wrong side of the constraint.
This discipline matters because organizations instinctively keep funding whatever produced the first success. If coding AI raised productivity, the reflex is to buy more coding AI. Once the constraint moves, the highest-return investment may sit somewhere else.
Sufficient Intelligence, Not Maximum Intelligence
None of this means machine intelligence is abundant everywhere. Frontier capability still matters, and the Stanford data shows why the argument needs care: even as the leading providers cluster, the gap between the strongest closed and open models widened to 3.3% in 2025, up from 0.5% in August 2024. Intelligence has not become a commodity.
What has changed is that useful intelligence is now available at many levels of capability, cost, privacy, and latency. That turns it into an allocation problem. Conventional software, search, or a database query may be sufficient for one decision. Another may call for a smaller model working over proprietary data. Ambiguous work may need stronger reasoning, and a novel, high-consequence decision may justify frontier capability plus additional verification and human judgment.
The economically mature question is: What is the least expensive intelligence sufficient for the decision at hand? This is ordinary resource economics. A hospital does not send every patient to its most specialized physician, and a company does not send every decision to the CEO. Machine intelligence should be allocated the same way.
The Mature Enterprise Will Not Maximize Autonomy
My years in manufacturing make me skeptical of a related assumption: that technological maturity means steady progress toward full autonomy.
A sophisticated plant does not work that way. Some processes run almost autonomously because volumes are high, variation is controlled, and the economics justify the investment.
Others run automatically within set limits under supervision. Some automate the normal cases and escalate exceptions to experienced technicians. Others stay largely manual because variability, low frequency, capital cost, or the consequences of failure make deeper automation unattractive. A mature plant runs a portfolio of operating modes. That is not inconsistency; it is good system design.The mature enterprise will look much the same
Exhibit 3. The mature enterprise runs a portfolio of operating modes
Operating mode | Who decides | Where it fits |
Fully autonomous | Machine | High volume, low variation, consequences that are contained or reversible |
Autonomous within limits | Machine, inside defined boundaries | Routine decisions with clear thresholds |
Management by exception | Machine, with people handling exceptions | Mostly routine work with occasional unusual cases |
Decision support | Person, informed by machine recommendations | Judgment-heavy decisions with material consequences |
Predominantly human | Person | Novel or ambiguous decisions where judgment creates more value than automation |
Autonomy, in other words, is an allocation decision made to create value, not an end state. The best enterprise is not the most autonomous one. It is the one that creates the most sustainable net value from the right combination of machine capability, human judgment, and control.
Governance Can Increase Productive Capacity
Seen through the lens of net value, governance looks different too. It is usually cast as innovation’s counterweight: technology creates possibilities, and governance limits them. Poorly designed governance creates needless friction. But too little governance can become the constraint itself.
Imagine an AI system that can adjust prices, commit funds, negotiate customer concessions, or change production schedules. Management may still confine it to making recommendations, because the enterprise lacks reliable ways to set authority boundaries, observe actions, verify outcomes, reverse mistakes, or intervene when consequences exceed acceptable limits. Building those controls lets the organization delegate more authority in a way that is more economically useful. In that case, governance has expanded what the technology can productively do.
The goal is neither maximum nor minimum governance, but sufficient control for the consequence at stake. Control theory provides the logic: as a system's range and speed of possible actions expand, its control capacity must keep pace to keep outcomes within acceptable bounds. That does not mean a person approves every action; human approval everywhere becomes its own constraint. Under the right conditions, good governance is not a tax on innovation. It is productive capacity.
Cybersecurity Belongs in the Same Equation
Cybersecurity becomes easier to place once the objective includes value destruction. If a new autonomous capability creates $100 million in benefit but exposes the enterprise and its stakeholders to $150 million in expected cyber, operational, or reputational loss, it destroys net value.
Security reduces that expected loss, but its role is not only defensive. Stronger identity, containment, access control, observability, policy enforcement, and recovery make it possible to authorize activities that management would otherwise judge too dangerous. Microsoft’s recent Windows security work is one concrete example: agents receive their own local or Entra-backed identities, so their activity can be distinguished from human activity, and policy-driven containers limit what they can reach. The specific technologies will change; the principle will not.
Autonomous systems also raise a subtler problem: individually permitted actions can combine into an impermissible outcome. An agent may legitimately read information, legitimately make a change, and legitimately execute a transaction, while the sequence produces a result no one intended to authorize. Governance therefore cannot stop at individual transactions. It increasingly has to watch trajectories and accumulated consequences, and know when to intervene.
Security preserves value by reducing expected loss, and it creates value by enabling safer delegation of authority. It belongs inside the production system, not beside it.
Strategy Follows the Constraint
When an important capability becomes easy to obtain from many sources, the value it creates does not vanish, but where that value lands can change. Innovation economics has long shown that returns often flow to whoever owns the scarce complementary assets needed to turn a technology into useful outcomes. For AI, those complements might include proprietary context and data, customer relationships, distribution, workflow position, physical infrastructure, organizational learning, decision authority, or the ability to verify and execute consequential decisions.
This offers a sharper way to think about the much-discussed AI “control points.” A control point isn't strategic just because someone labels it as such. It becomes strategic when it governs a scarce complement that materially limits value creation, capture, or protection. If reasoning becomes easier to obtain, proprietary context may matter more. If autonomous execution grows faster than the organization’s capacity to control it, authority and verification may matter more.
The same logic applies to the firm's boundaries. Companies may bring a capability in-house when controlling it becomes strategically decisive, and rely on partners or markets once it becomes standardized. AI does not predict a universal move toward integration or toward specialization; the direction depends on where scarcity moves.
Uncertainty complicates every one of these choices. Companies can lock themselves into vendors, architectures, workflows, labor structures, or governance regimes before they know where the lasting constraint will sit. Investment theory offers the corrective: when uncertainty is high, flexibility has economic value. The danger is not commitment, which strategy eventually requires. It is premature irreversibility. Staged investment, modular architecture, and the ability to switch suppliers can look slightly inefficient in a static business case while preserving valuable options.
Seven Questions for Boards
The board need not become a technology architecture committee. It does need to oversee how a major shift in productive capability affects value, capital allocation, authority, risk, and enterprise design. Seven questions frame that oversight:
1. What sustainable value are we trying to create, for whom, and how will we capture enough of it to reward capital and keep investing?
2. What actually constrains that value creation today, and will the new capability materially relax that constraint?
3. If the technology succeeds, what becomes scarce next, and is that successor constraint economically consequential?
4. What combination of machine autonomy, human judgment, and deterministic control produces the best system outcome, rather than the most local automation?
5. What value could be destroyed through cyber loss, poor decisions, operational failure, safety problems, regulation, or lost trust, and what controls are economically justified?
6. Which scarce capabilities should we own, which should we obtain through partners or markets, and where should we preserve flexibility while uncertainty remains high?
7. What observable evidence would show that our assumptions about the constraint, the economics, or the required controls are wrong?
At heart, these are not questions about AI. They are management questions.
A Useful Theory Must Be Capable of Being Wrong
Constraint Migration becomes dangerous if every outcome can be read as confirming it, so its limits matter. In some domains, superior technology may stay scarce long enough to dominate the economics. Some interventions simply make an activity cheaper without creating any significant successor constraint. Some autonomous systems may handle enough variability that human supervision stops being a material constraint. Vertically integrated architectures may beat modular ones when the gains from integration outweigh the value of flexibility.
We should not explain these cases away. The proposition is deliberately narrower: when technology materially expands one productive capability, management should expect relative scarcity to change and should look for evidence that something else has become more consequential to system value. Sometimes the evidence will show that nothing strategically important has moved, and that is a legitimate finding. A management theory’s job is not to explain every outcome after the fact. It is to improve decisions before the outcome is known.
AI Is the Shock, Not the Theory
AI may prove to be the most consequential technological transition of our working lives. But durable management principles cannot put AI at their center. Technological revolutions repeatedly expand productive capabilities. That expansion changes relative scarcity, which changes the value of complementary resources. Capital and attention respond, control systems evolve, and firm boundaries shift, until the newly scarce capability expands and the cycle begins again.
AI makes the cycle unusually visible because intelligence has historically been one of the scarcest and most expensive resources inside any organization, and its supply is now growing at extraordinary speed. That does not abolish scarcity. It moves it.
The lasting advantage, then, does not come from owning whatever is scarce today. It comes from recognizing what has become economically consequential next, understanding why, and reallocating capital, capability, authority, and control around it before competitors do.
Technology changes what is possible. Economics reveals what is scarce. Management decides where to act. Governance sets the boundaries for consequential action. Strategy decides what to own, orchestrate, source, or keep reversible.
The question that survives every technology cycle is simple: What becomes scarce next?
Appendix: Intellectual Foundations
This argument synthesizes ideas, not a claim that its underlying economics are new. Constraint Migration does not replace the theories below; it applies them to how technology shifts economically consequential constraints.
Field | What it explains | Foundational contributors |
Theory of constraints and operations management | Why a few binding constraints govern system performance, and why improving a nonconstraint yields little | Eliyahu Goldratt |
Information and managerial attention | How abundance of one resource makes another scarce | Herbert Simon |
Cybernetics and control theory | Why control capacity must match the range of behavior a system can produce | W. Ross Ashby |
Transaction-cost economics and the theory of the firm | Why firms own some capabilities and source others through markets or partners | Ronald Coase, Oliver Williamson |
Innovation economics and complementary assets | Why innovators often fail to capture the returns, which flow to owners of hard-to-replicate complements | David Teece |
Real options and investment under uncertainty | Why staged commitment and the option to expand, defer, switch, or abandon have value | Stewart Myers; Avinash Dixit and Robert Pindyck |
Path dependence and organizational capability | How past investments and routines constrain future choices | Paul David, W. Brian Arthur; Richard Nelson and Sidney Winter |
Decision theory under uncertainty | How to choose well when outcomes are uncertain and information incomplete | Leonard Savage, Howard Raiffa |
Comments