A business unit deploys AI across several functions. Customer operations reduce handling time. Finance improves forecasting. Procurement identifies lower-cost alternatives. Project teams automate reporting. Sales improves lead qualification. Service operations detect emerging problems earlier.

Each team may be able to demonstrate a legitimate improvement while the expected enterprise benefit remains elusive. Costs may have shifted rather than disappeared, faster throughput in one function may create a backlog elsewhere, and a local productivity gain may increase demand on a downstream team that cannot absorb it. None of this requires the AI intervention itself to have failed. The local improvement can be real while the performance of the wider system remains unchanged or deteriorates.

Improving a component and improving the system are not the same thing.

W. Edwards Deming and Eliyahu Goldratt approached that problem from different traditions. Deming emphasized organizations as systems of interdependent components working toward an aim and warned against improving parts in ways that weakened the whole. Goldratt focused management attention on the condition limiting system performance and on improvement that changes global rather than merely local results.

Neither was writing about AI. Their work remains useful because AI can make improvement inside individual activities much faster and cheaper while leaving the relationships among those activities, and the performance of the overall system, unresolved.

1. Deming: the organization is the system, not the collection of parts

Deming's systems thinking begins with interdependence and aim.

He described a system as a network of interdependent components working together toward an aim. The management implication is important: individual departments or processes cannot be judged independently of what their performance does to the wider system.

The aim matters because a local result cannot be judged solely by whether the component improved. Management must also ask whether that improvement contributes to what the wider system exists to accomplish.

Organizations nevertheless have good reasons to divide responsibility. Budgets are allocated. Functions have targets. Managers are accountable for defined areas. Improvement programs naturally look for opportunities inside the boundaries of a team, process, or business unit.

That decomposition is necessary. The problem appears when managerial decomposition becomes managerial independence.

A procurement organization can reduce unit cost while increasing operating instability elsewhere. A service function can reduce average handling time while creating more repeat contacts. A business unit can improve its own economics while shifting cost, risk, or workload into another part of the enterprise.

Deming's argument is not that individual components should not improve. It is that improvement has to be understood in relation to the aim and performance of the wider system.

Viewed under AI conditions, this becomes more consequential because the ability to improve individual activities can become substantially cheaper and faster.

2. AI makes local improvement easier to produce

Many AI opportunities are first discovered locally.

A function sees expensive manual work and automates part of it. A team sees a forecasting problem and applies better analytics. A process contains repetitive judgment and gains an AI assistant. A manager sees an information bottleneck and introduces a tool that reduces the effort required to produce an answer.

That is a rational way for adoption to begin. Organizations do not need to redesign themselves before every experiment. Local experimentation allows useful capabilities to emerge, provides evidence, and helps people understand where the technology can genuinely improve work.

The management problem appears as successful interventions accumulate.

Once AI improves one component, the surrounding system has to absorb the change.

If software development accelerates, testing, architecture, security review, deployment, adoption, or business readiness may become limiting conditions. If sales qualification improves materially, delivery capacity or commercial governance may become more important.

A local productivity measure can therefore improve while overall system performance changes very little. In some circumstances, the wider result can deteriorate.

Viewed through Deming's systems lens, this is not paradoxical. The intervention improved a component of an interdependent system. Management still has to understand what that change does to the system and its aim.

3. Goldratt: improve where the system is constrained

Goldratt approaches the problem from a more concentrated question:

What currently limits the performance of the system?

Theory of Constraints starts from the proposition that the performance of a system is limited by its constraint, the factor that currently prevents the system from achieving more of its goal.

Its focusing process directs management to identify the constraint, decide how to exploit it effectively, subordinate other decisions to that choice, elevate the constraint where appropriate, and then look again because improvement may change where the system is constrained next.

That logic is highly relevant to AI investment.

Imagine an organization where management believes analytical capacity is limiting performance. AI improves it substantially. Reports are produced faster, risks are identified earlier, recommendations become available sooner, and more alternatives can be evaluated.

If decisions still wait weeks for the appropriate authority, analytical capacity may not have been the system constraint. Or, if it was, relieving it has exposed the constraint that now limits performance.

That constraint might be management attention, regulatory approval, specialist capacity, integration capability, customer readiness, physical infrastructure, or the ability of another process to absorb additional throughput.

Goldratt's contribution changes the management question from:

Where can AI improve performance?

to:

Where will improvement change the performance of the system?

Those questions can produce very different investment priorities.

A substantial productivity gain away from the system constraint may contribute less to overall system performance than a smaller improvement at the point currently limiting throughput or goal attainment.

That is an AI-era interpretation of Goldratt's logic, not a claim he made about AI. The underlying principle is that local improvement should be judged in relation to the constraint and the performance of the system as a whole.

4. Local success can move the constraint

Successful improvement changes the system.

An organization identifies a limiting condition, improves it, and then discovers that something else now governs performance. Theory of Constraints explicitly anticipates this dynamic. Once a constraint has been relieved, management returns to the question of what now limits the system rather than allowing yesterday's constraint to determine today's improvement priorities.

AI can make that movement happen quickly.

Consider a customer-service operation in which triage is slow and specialist cases are difficult to identify. An AI-enabled triage process classifies and routes cases much more quickly.

The improvement is real.

But suppose specialist resolution capacity does not change. Cases now reach specialists faster, the queue in front of them grows, and end-to-end resolution time barely improves.

Management could continue optimizing triage because that is where the technology has already demonstrated success. The system perspective directs attention elsewhere.

The intervention has changed the operating conditions. Management now needs to identify what limits performance next.

This discipline matters because technology investment naturally gravitates toward places where technical gains are visible and repeatable. System performance may require a different intervention entirely: more specialist capacity, a redesigned handoff, changed authority, altered demand, revised commercial terms, or less work entering the system.

AI-enabled improvement can expose a constraint, relieve it, or change where the system becomes constrained next.

The management task is to keep following the system rather than the technology.

5. What Deming and Goldratt reveal under AI conditions

Deming and Goldratt operate at different levels of the management problem.

Deming directs attention to the aim and interdependence of the system and the danger of suboptimizing its components. His broader System of Profound Knowledge extends beyond systems thinking into variation, knowledge, and psychology. Goldratt provides a focusing logic for identifying the condition currently limiting system performance and organizing improvement around it.

Together they reveal an important problem for AI-enabled organizations.

AI makes local improvement increasingly available. That is good news, but successful local improvements may aggregate into enterprise improvement without there being any reason to assume that they will.

Local changes alter workload, timing, demand, dependencies, capacity, incentives, and the location of constraints elsewhere in the system. They can shift cost rather than eliminate it. They can improve a measure attached to one function while weakening an outcome that matters more to the whole.

The management requirement is therefore larger than identifying good AI use cases. It is to understand what those use cases do to the system.

6. The AIMR extension: from local improvement to enterprise coherence

AIMR describes one requirement for doing this as Management-System Coherence.

AI capability can proliferate across the enterprise faster than the surrounding management system can adapt to its effects. Coherence asks whether local changes combine in ways that support the performance and aims of the wider system.

That requires management to examine not only the capability introduced but also the capacity it releases or consumes, the throughput it changes, the dependencies and handoffs it affects, the constraints it moves, and the enterprise outcomes to which the improvement is supposed to contribute.

This also changes how AI value should be interpreted.

A capability improvement shows that something became possible or better. An efficiency gain shows that resources were used differently. A performance improvement shows that a relevant operating result changed. The next management question is whether that change contributed to the performance of the wider system and to the enterprise outcomes it was intended to support.

Local improvement does not by itself establish enterprise contribution or value realization.

The organization still has to trace whether the improvement changed system performance, how that change propagated through the enterprise, and whether the resulting contribution supported outcomes the organization actually values.

That is where Deming and Goldratt remain useful. Deming reminds management that the aim belongs to the system rather than to its individual components. Goldratt reminds management that improvement has different consequences depending on where the system is currently constrained.

AI gives organizations extraordinary new ways to improve the parts. Management still has to make those improvements add up.

The question is not whether AI made a component better. It is whether that improvement changed the performance of the system the enterprise actually depends on.

Sources and further reading

• W. Edwards Deming, Out of the Crisis.

• W. Edwards Deming, The New Economics for Industry, Government, Education.

• The W. Edwards Deming Institute, “Managing the Organization as a System” and “Appreciation for a System.”

• Eliyahu M. Goldratt and Jeff Cox, The Goal: A Process of Ongoing Improvement.

• Eliyahu M. Goldratt, What Is This Thing Called Theory of Constraints?

• Goldratt material on the Five Focusing Steps and the process of ongoing improvement.