What Happens When Intelligence Is Commoditised?

For a long time, producing sophisticated knowledge work has required scarce expertise, considerable time and institutions capable of bringing both together. AI is changing the cost and availability of some of that work. Analysis, software, writing and strategic options can become easier to produce without the organisations using them becoming better at achieving their goals.

That gap is what interests me. We can have more capability and still make little progress. The Theory of Constraints offers a useful way to examine why: improving a part of a system does not necessarily improve the whole. We have to understand what is limiting it, and whether the new capability changes that limitation.

I use commoditised intelligence here to describe increasingly accessible, less expensive cognitive capability. It does not mean that every form of expertise is interchangeable, that AI is reliable at every task, or that the cost of using it disappears. The question is what happens as more of the work we once found difficult becomes easier to produce.

A faster machine, the same factory

Imagine a furniture factory with five stages. Every item passes through cutting, shaping, assembly, finishing and packaging. For this simplified example, assume there is enough demand and material, the stages operate reliably, and each capacity is expressed in equivalent furniture units per day.

Stage	Capacity per day
Cutting	100
Shaping	80
Assembly	40
Finishing	70
Packaging	100

The factory can complete 40 units a day. Assembly is the bottleneck. Now suppose a technological breakthrough increases cutting capacity from 100 to 1,000 units at almost no additional cost. Cutting productivity has risen tenfold, but the factory still completes 40 units. Nothing has changed at the stage that limits its output.

If the factory starts cutting more wood simply because it can, unfinished work accumulates. Money is tied up in material waiting for assembly. It needs space, handling and attention. The breakthrough can make one department look more productive while making the factory harder and more expensive to run.

This is the distinction the Theory of Constraints asks us to notice. In an interdependent system, the capacity of the limiting stage can matter more than impressive gains elsewhere. A restaurant with room for 100 customers but a kitchen capable of preparing only 30 meals an hour faces a similar problem. More tables will not, by themselves, help it serve more meals.

Now replace the faster cutting machine with AI. A company can produce more analysis, code or proposals, but those outputs still have to pass through the rest of its system. Someone has to make decisions. Customers have to find the work useful. Products need to be deployed and supported. The question is whether cognitive production was the constraint, or simply the most visible source of effort.

The Theory of Constraints, in brief

Eliyahu Goldratt popularised this way of thinking through The Goal, his business novel co-authored with Jeff Cox. Its challenge to management is straightforward: optimising each department independently can undermine the performance of the organisation. Sales, engineering, operations and finance are connected. A target that makes sense within one function may create work another cannot absorb.

Goldratt’s five focusing steps provide a sequence for addressing that problem:

Identify the constraint. Find what currently limits progress towards the system’s goal.
Exploit the constraint. Make better use of its existing capacity before adding more.
Subordinate everything else. Organise the rest of the system to support that decision.
Elevate the constraint. Increase its capacity if it still limits performance.
Repeat. When the constraint changes, identify the next limitation rather than preserving the old rules through inertia.

The order matters. Buying more capacity before examining how existing capacity is used can be expensive. In the factory, assembly might lose time waiting for the right parts or correcting avoidable defects. Protecting it from those interruptions could improve output before another machine is needed.

Subordination is often the uncomfortable step. It may require the cutting department to produce less than it could, so the factory can work better. A manager rewarded for machine utilisation has a reason to resist that change. The measure of success has to move from keeping a resource busy to helping the system achieve its goal.

Goldratt’s financial measures distinguish throughput, inventory or investment, and operating expense. Broadly, they ask us to look at money generated through sales, money tied up in the system, and the cost of operating it. His finance and measurements programme develops this relationship between local decisions and the economics of the whole. Producing something is not the same as selling it, and keeping every department active is not the same as improving the business.

When I apply this lens beyond a commercial operation, the goal needs to be defined afresh. A hospital, a public institution and a person’s life cannot simply inherit a factory’s measure of throughput. What counts as progress is part of the question, not a detail to settle after optimisation begins.

When software becomes easier to build

Consider a software company whose ideas have always competed for a limited pool of developers. Every feature carries an opportunity cost, and product managers have to choose carefully. Now suppose AI lets the team produce usable software ten times faster. This is an illustrative scenario, not a claim about a typical productivity gain.

Initially, the benefits could be substantial. Prototypes become cheaper, experiments that were previously unaffordable become possible, and useful improvements reach customers sooner. But the team may then discover constraints that were less visible while engineering capacity was scarce. Customers have limited attention. Sales relationships take time. Leadership still has to decide which opportunities deserve investment.

Each new feature also creates obligations: testing, security, support, integration and maintenance. A working prototype is only part of a deployed product. The company can become capable of building much more than it can usefully put into the world, while the responsibilities created by that software continue to grow.

If it keeps measuring lines of code generated, features shipped and development cycles shortened, it may celebrate production gains while overlooking the constraint that now matters. Perhaps it cannot attract customers. Perhaps its users cannot absorb another change. Perhaps nobody can explain why the next feature deserves to exist.

In that situation, increasing output further resembles cutting more wood while assembly remains overwhelmed. The gain in engineering capacity is real. Its contribution to the business depends on what happens after the code is written.

Time saved is not the whole result

AI adoption often begins with a task that takes too long: preparing reports, summarising meetings, analysing spreadsheets or drafting communications. Saving time can be useful. But an hour no longer spent on a task is not automatically an hour converted into a better outcome.

Suppose a management team reduces weekly report preparation from ten hours to one. If preparing reports was delaying consequential decisions, the change may help the organisation move. If the reports were already arriving faster than anyone could act on them, faster preparation leaves the limiting problem untouched. The team might even produce more reports and add to the queue of things requiring attention.

Imagine another company producing 20 strategic proposals a year but able to commit to and execute only three. AI allows it to produce 200 proposals. The supply of possible decisions has increased; its capacity to choose and act may not have changed at all. More options can now compete for the attention needed to make any one of them real.

This does not make automation outside a bottleneck worthless. Lower operating expenses can improve the economics even when throughput stays the same. Work can become less exhausting, and released capacity can be directed towards a constraint elsewhere. The mistake is treating a local time saving as sufficient evidence of system-wide improvement.

We need to follow the benefit through. What happened to the saved time or money? Did it improve a decision, relieve another limitation or make the work more sustainable? Did the system’s outcome change? These are harder questions than how many minutes a tool removed from a task, but they help establish what the improvement actually achieved.

Where value moves

When a capability becomes easier to obtain, its scarcity value can diminish. Other difficulties become more visible. Publishing offers a familiar analogy: reducing the cost of putting information into the world does not give readers more time to attend to it. Discovery, reputation and trust become important parts of the problem. A technically competent photograph is also only one part of work whose value may depend on access, perspective or emotional resonance.

AI may create similar shifts in cognitive work. If a competent analysis is easier to produce, advantage may lie elsewhere: obtaining trustworthy evidence, knowing which analysis matters, or persuading people to act on it. If code is easier to generate, demand and integration may become more consequential constraints in building a business. This is a way of interpreting the shift, not a prediction that every profession will follow the same path.

It is tempting to declare that the next scarce resource must be judgement, creativity, trust or taste. The Theory of Constraints asks for a more specific answer. A pharmaceutical company may be limited by clinical validation. A startup may be limited by customer acquisition. A public institution may lack legitimacy for a proposed change. An individual may lack energy, attention or the willingness to commit.

There is no universal replacement bottleneck. Constraints belong to particular systems, pursuing particular goals, in particular circumstances. AI changes the relative capacities of their parts. Understanding those imbalances is more useful than choosing a fashionable human quality and assuming it will always be scarce.

Some constraints are rules we have made

A bottleneck need not be a machine or an understaffed team. An organisation might have the expertise and resources to test an idea but require seven layers of approval before doing so. A leadership team might have compelling evidence against its strategy and still refuse to reconsider it. A professional might know enough to begin independent work but remain attached to an identity that makes leaving employment difficult.

These examples involve policy, incentives and belief rather than a simple shortage of capacity. Adding intelligence may help make alternatives visible, but the rules governing action remain. A system can be well informed and still unable, or unwilling, to respond.

Goldratt’s work extends into examining assumptions and the changes needed to act on them. The useful question is sometimes why we use a resource in the way we do, rather than how to acquire more of it. An approval rule may once have protected the organisation from a genuine risk. We need to understand whether it still does, what changing it would expose, and who has the authority to decide.

AI could help people identify inconsistencies, explore consequences or challenge assumptions. It cannot be assumed that surfacing an unnecessary rule will cause the rule to disappear. Someone may benefit from it. Someone may carry the risk of changing it. Knowing what should change and being able to bring about that change are different capabilities.

Coherence as a working thesis

I suspect coherence will become a consequential constraint in many AI-rich organisations. By coherence, I mean a workable relationship between a system’s purpose, decisions, activities and the reality it encounters. That does not require everyone to agree about everything. It does require differences and trade-offs to be understood well enough for people to act together.

Consider a company using agents across its functions. One is directed towards customer acquisition, another towards profitability, another towards satisfaction and another towards reducing support costs. Each may perform its assigned task well. Without a shared understanding of the goal and its limits, their actions can pull against one another. More capable components do not guarantee a more capable organisation.

This is familiar from functional silos and conflicting incentives. AI may amplify it by making action cheaper and faster. Initiatives begin, commitments are made and outputs demand integration before the organisation has resolved what should take priority. Its ability to generate work can exceed its ability to coordinate it.

A different organisation might identify the outcome it seeks and the constraint that limits it, then organise people and agents around that understanding. It could produce fewer reports and start fewer projects while achieving more. That possibility is what I mean by directing abundant intelligence towards coherence. It remains a hypothesis to examine in actual work, rather than an explanation for every organisation’s difficulties.

Coherence also cannot simply be purchased as another software subscription. Tools may help, but the work involves how decisions are made, conflicts are resolved and consequences are shared. A system can have excellent information and still lack a way to act on disagreement.

Capability and the life we want

The same lens becomes uncomfortable when we apply it to ourselves. Modern professional life rewards capability. We acquire knowledge, develop skills and build identities around what we can do. When machines make more of that work easier, an understandable response is to produce more: learn another tool, start another project, create another piece of content.

But what is the system meant to achieve? Imagine someone seeking financial independence, meaningful relationships, good health and time for interesting work. They have access to tools that help them research, write and build faster than before. Maximising intellectual output would not, by itself, give them the life they want.

If income is the constraint, those tools might help them test a business. If health is the constraint, another business plan may do little. If commitment is difficult, generating more options may deepen the difficulty. The same capability takes on different value depending on what currently prevents progress.

This reframes productivity as a means. The question becomes what is preventing the life we want, and how the resources available might help. It also makes purpose harder to leave implicit. A tool can help us pursue a goal without helping us examine whether we still want it.

Human agency matters here because choices have consequences we live with. We participate in defining purposes, accepting responsibility and deciding which trade-offs we are prepared to make. This is not a claim that humans will always have a monopoly on judgement or creativity. It is a distinction between having cognitive capability available and exercising purpose through a commitment. An excellent recommendation does not remove that commitment from the picture.

Begin with the outcome

Efficiency concerns how economically an activity is performed. Effectiveness concerns whether it advances the intended purpose. AI can improve the former without improving the latter. That suggests an approach to adoption that begins with an outcome and traces what currently prevents it, before compiling an inventory of tasks to automate.

Take two companies struggling to grow sales. In one, qualified prospects wait too long for customised proposals. If preparation capacity is the limiting factor, AI-assisted drafting might allow the team to serve more of those prospects. It would still need to check accuracy and determine whether faster proposals actually convert into worthwhile business.

In the other company, demand is weak. Faster proposal preparation might reduce costs while leaving that problem untouched. Researching unmet needs, testing a different proposition or examining why prospects decline could be more useful applications of the same technology. We cannot determine the value of the tool from the task alone.

The questions I would want to ask are connected: what outcome are we trying to improve, what prevents more of it, and why does that limitation exist? Can we use what we already have differently? Would AI help with that change? If we succeeded, what would limit us next?

This approach does not require ignoring smaller improvements. It helps establish what each improvement is for, what evidence would show that it helped, and what else needs to change for the benefit to reach the system’s outcome.

Abundance can create a queue

Open an AI interface with a problem and it can quickly produce a strategy, a project plan, a technical architecture and several alternative approaches. Each answer offers another possibility to examine. The intellectual work can be useful, and the feeling of progress can be convincing. Yet it is possible to spend hours improving the account of what we might do without making a decision that changes anything outside the conversation.

Those plans resemble inventory in a loose sense: work waiting to become an outcome. They are not inventory in Goldratt’s financial definition, but the analogy draws attention to the queue. If more possibilities arrive faster than we can evaluate or act on them, producing another set may be less useful than working through what is already there.

The limiting step might be a conversation with a customer, a prototype placed in someone’s hands, or a decision that closes off an attractive alternative. These actions bring consequences and evidence we cannot fully control. Staying with another round of analysis can postpone that encounter while still feeling like responsible preparation.

I do not think the answer is to abandon analysis. It is to keep it connected to a next action and the uncertainty that action can address. Intelligence becomes useful when it helps us move between understanding and reality: form a hypothesis, intervene, observe what happens and reconsider. A better plan is worth something when it changes how we do that work.

The goal may need to change too

A factory is a useful teaching example because its stages and capacities can be made explicit. Organisations operate under less settled conditions. Their purposes may be contested, participants have different interests, and the environment changes in response to their own actions. Several interacting limitations may matter at once, and the apparent bottleneck may be a symptom of something we have not understood.

The Theory of Constraints is a lens for investigating these conditions, not proof that every complicated situation reduces to one measurable bottleneck. We need to test our account of the system. We also need to remain open to the possibility that its goal no longer makes sense.

A company can become better at manufacturing a product whose market is disappearing. Relieving production constraints would improve the operation without addressing its future. An individual can become more efficient at sustaining a way of life they no longer want. In both cases, improving execution can distract from questioning the commitment behind it.

This is where the constraint lens meets the argument in We Built Factories When We Should Have Tended Gardens. Adaptive organisations need to notice what is changing, connect evidence to decisions and revise their work accordingly. They still need discipline and accountability. They also need room to discover that an earlier definition of progress was wrong.

People and agents could support shared cycles of observation, interpretation and experimentation. The purpose would be to help the organisation understand what limits it now and respond as that understanding changes. Whether software can meaningfully improve those cycles is one of the questions behind Duyên. It cannot be assumed that a new operating system will overcome the incentives and relationships in which it is used.

What is actually stopping us?

More accessible cognitive capability should expand what individuals and small teams can attempt. It can lower the cost of experimentation and make previously inaccessible work possible. Those are opportunities worth exploring. But abundance does not remove the need to understand what happens between capability and consequence.

The Theory of Constraints gives us permission to stop trying to improve everything at once. We can begin with a purpose, examine the system and focus on what currently limits progress. We can make better use of existing capacity, organise supporting work around it and add capacity where it helps. When conditions change, we can reconsider the diagnosis rather than keep defending an improvement whose usefulness has passed.

That is part of the inquiry behind Nya. I am interested in what abundant intelligence enables, but also in the decisions, arrangements and commitments needed to make those possibilities matter. Coherence is one working thesis. The constraints will have to be discovered in the particular people and organisations doing the work.

When intelligence becomes easier to obtain, another question becomes harder to avoid: what is worth doing, and what is actually stopping us?

The factory, restaurant and software-company examples are simplified illustrations, not case studies or claims of measured AI productivity. The five focusing steps follow Goldratt’s published formulation. Applying the constraint lens to AI, human agency and coherence is the developing argument of this essay.