Eliyahu Goldratt, although trained as a physicist, had a profound impact on industrial management with books such as The Goal (1984) and The Haystack Syndrome (1990). His contribution is based on the idea that the performance of a system is always limited by its bottleneck. No matter what gains are made elsewhere, this bottleneck alone determines the performance of the system.

At this point, you’re probably wondering what the “old” world of factories is doing on a blog that talks a lot about tech and knowledge workers, but it’s precisely because, as an article in The New Yorker pointed out:

“Peter Drucker noted that during the twentieth century, the productivity of manual workers in the manufacturing sector increased by a factor of fifty as we got smarter about the best way to build products. He argued that the knowledge sector, by contrast, had hardly begun a similar process of self-examination and improvement, existing at the end of the twentieth century where manufacturing had been a hundred years earlier.” (Slack Is the Right Tool for the Wrong Way to Work).

It was from this reflection that the theory of constraints, which I have mentioned many times here, was born. It proposes a fairly simple approach: identify the bottleneck, focus efforts on it, make the most of it, then move on to the next one, because once the problem is solved, the bottleneck moves. But the key takeaway is that performance is not measured by each individual involved in an activity, but by the whole. As Goldratt writes: “Local optimum is not global optimum”.

In short:

• The performance of a system is determined by its bottleneck, not by the efficiency of its individual components. Improving an unconstrained resource does not improve overall performance.
• The Theory of Constraints proposes a sequential approach: identify the constraint, focus efforts on resolving it, then start again with the next one, because every constraint that is removed reveals another.
• In knowledge organizations, bottlenecks are intangible (processes, people, coordination) but their impact is equivalent to those observed in industrial environments, often masked by an illusion of productivity gains.
• Local optimization, if not synchronized with the rest of the system, generates imbalances, inventories, delays, and costs without overall benefit, a phenomenon observable in management, algorithms, or strategy.
• Automation or AI projects only produce real gains if they are part of a systemic approach, otherwise, they amplify inefficiencies by creating unnecessary surplus instead of improving the overall flow.

The “law of the local maximum”

Although the theory of constraints is a discipline in its own right within management theory, the latter has never specifically formalized any of Goldratt’s laws on bottlenecks or the opposition between global and local performance, nor any law of the global maximum. I will therefore do so myself: achieving maximum performance at one point in a chain does not necessarily improve overall optimum performance and may even accentuate imbalances.

Local improvement only makes sense if it spreads throughout the chain without interruption. Otherwise, it only generates a surplus that accumulates in the form of inventory, pending tasks, or additional delays, without contributing anything to the business in terms of productivity and therefore without benefiting the business and its customers, whether internal or external.

Talking about local maximum or global optimum is not just playing with words: it highlights two radically different approaches: one that improves things on the surface and the other that has a real impact on performance.

The system, always the system

You know my fondness for systems thinking and my aversion to any initiative that focuses only on symptoms and not causes, as well as those that forget that the limit to performance does not lie with a local contributor but with the system as a whole (The Problem Isn’t the Employee, It’s the System and You do not rise to the level of your goals. You fall to the level of your systems. (James Clear))

Rather than seeking to optimize all components of a system, Goldratt proposes focusing attention on the bottleneck that sets the overall pace. This choice is based on a simple observation: unconstrained resources can always be adjusted, while the constrained resource determines the effective capacity of the system.

A common management mistake is to require every resource, machine, or individual to operate at maximum capacity at all times. This intention seems rational and, in fact, seemed obvious to me for a long time: the more a resource is used, the more it contributes. However, in reality, this often has the opposite effect.

In a factory, a machine that runs continuously when the downstream process is unable to keep up only accumulates intermediate stocks, which generate financial, logistical, and organizational costs without ever speeding up delivery to the customer or increasing revenue or margins. In a service business, the logic is the same: overloading an employee’s workload creates queues and delays in validation, which ultimately slow down the entire team’s work.

In other words, the full utilization of an unconstrained resource never improves the overall optimum. It even exacerbates inefficiencies by producing surplus, variability, and additional costs. This is why the theory of constraints does not aim for maximum utilization of each resource, but rather for improvement of the overall flow by acting on the true constraint.

When I first read The Goal, I was struck by the metaphor used to illustrate this idea: that of a group hike. Indeed, the speed of the group is not determined by the fastest walker, but by the slowest, and as long as the latter does not speed up, the group progresses at their pace. Trying to make the fast walkers go even faster is pointless, except to create a growing gap between the two ends of the line, and will not make the whole group arrive any faster, since those walking behind the slowest cannot overtake them, as in a workflow. In the end, the slowest will not arrive any faster, while the fast will wait for them, standing still, with nothing to do.

This perfectly sums up the logic of the bottleneck: only constraints determine overall performance.

From industrial chains to intangible flows

Goldratt’s thinking originally arose in an industrial context: machines, workshops, production lines. In this world, bottlenecks are visible, measurable, and tangible: a machine breaks down, the pace slows, inventory piles up, and overall efficiency depends directly on the capacity of the weakest link.

But in knowledge work, the logic is the same, even if the flows are intangible and therefore more difficult to perceive. Here, the bottlenecks are not machines, but individuals, committees, and processes. A report waiting to be approved, a manager overwhelmed with meetings, an endless compliance procedure: each of these blockages plays exactly the same role as an industrial bottleneck (Just because work is invisible, it doesn’t mean that it can’t be improved). Goldratt later wrote about the specific case of project management, but we’ll talk about that another time.

In any case, the illusion of productivity is even stronger in this sector, because intangible activities leave no visible traces. An employee can fill their days, producing more and more documents or data, without anything really moving forward (A few ideas to ensure you are more productive than busy and Is creating documents really work?). The surplus translates into cognitive overload, a pile-up of “unfinished” tasks to do, and deadlines that are slowly but surely getting longer (Digital Infobesity: When Collaboration Tools Degrade Productivity, QWL and Amplify Mental Workload). Just like a machine that accumulates work in progress without speeding up the flow, a knowledge worker can give the impression of being intensely productive when in fact they are only feeding an invisible bottleneck.

Collaboration introduces an additional dimension. The bottleneck is not always an isolated individual, but the way teams work together. Validation that goes through an endless chain of managers, a meeting that delays a decision, or digital tools that multiply versions of the same document and double or triple entries are all invisible constraints. Everyone may give the impression of being productive within their own scope, but the whole process gets bogged down. And I’m not even talking about the misuse of collaboration tools that are supposed to make us more efficient but only serve to overwhelm our capacities (Eliyahu Goldratt’s fictional interview on infobesity and bottlenecks in knowledge work). In this case, the illusion of individual performance masks collective inefficiency: the overall flow deteriorates not because of a lack of effort, but because of a lack of synchronization or coordination.

Thus, what Goldratt observed on a production line translates perfectly to the knowledge economy. The bottleneck is simply less visible, but its effects on overall performance remain the same: every organization is limited by the most saturated human or decision-making link.

In other words: when work imposes dependencies/interdependencies between people or departments, the productivity of the whole is equal to the productivity of the least productive. The time lost by others has no impact as long as they remain less productive than the least productive, but the time lost by the least productive is lost for the whole, and this is irreversible. Nothing will ever make it up.

Taking this logic to its logical conclusion also means that instead of “pushing” a non-bottleneck person to their limits, you can, depending on the number of tasks behind schedule on the bottleneck resource, tell them to go for a coffee or take a vacation because making them work brings you absolutely nothing. This calls into question certain approaches to productivity in business: increasing the ratio of quantity produced to hourly cost may never translate into the company’s books, and for good reason.

Resonances in other disciplines

This distinction between local maximum and global optimum extends far beyond industrial management.

In algorithms, researchers distinguish between so-called greedy approaches, which choose the most advantageous option locally at each stage (Greedy algorithm). This seemingly rational approach is in fact highly inefficient: a greedy algorithm can become locked into a partial way of thinking that prevents it from seeing much better solutions on a global scale. The classic example is that of the traveling salesman, which recently came to mind during an AI certification course to understand the difference between combinatorial analysis and probabilistic models: choosing the nearest city at each stage results in a route that multiplies detours and ultimately costs more (Traveling Salesman Problem). It’s like a hiker who only looks ten meters ahead: each fork in the road seems optimal at the time, but the entire route becomes incoherent.

In organizational management, the logic of silos illustrates the same dynamic. Finance seeks to reduce costs, production seeks to speed up turnaround times, and human resources seeks to maximize internal satisfaction. Each achieves its own goal, but these strategies can contradict and cancel each other out: cost reductions can slow down innovation, speed gains can degrade quality, and the proliferation of internal projects can disrupt production. The overall result deteriorates, as in an orchestra where each musician plays louder to “make their instrument heard better”: each gets the best out of their instrument, but the ensemble becomes cacophonous.

In strategy, Herbert Simon showed that the rationality of decision-makers is limited by the information available to them (Bounded Rationality). Faced with complexity, they settle for a solution deemed “sufficient” in their immediate scope, which he called satisficing. But this “good enough” solution at the local level traps the organization in a sub-optimal logic. A typical example is that of businesses that prioritize fast wins in one division to meet quarterly targets, at the expense of heavy investments that would have improved overall performance in the long term. It’s the equivalent of a chess player grabbing the first piece he can take: the move seems immediately satisfactory, but it compromises the entire game.

A reminder in the age of AI

It is very useful to keep these concepts in mind at a time when businesses are multiplying AI projects without seeing the promised productivity gains materialize (MIT report: 95% of generative AI pilots at companies are failing). Automation technologies and artificial intelligence enable spectacular gains on specific tasks: generating text, producing code, analyzing data, but these gains, taken in isolation, say nothing about the overall performance of the organization (AI in the workplace: going beyond augmentation to actually transform).

If bottlenecks persist elsewhere (validation, integration, user adoption), overall optimization does not progress and, in some cases, even regresses: local overproduction generates an accumulation of information or deliverables that slows down decision-making.

The illusion of productivity then becomes a factor of inefficiency and a generator of additional financial and cognitive costs.

Bottom Line

What Goldratt formulated in an industrial context is still highly relevant today. Whether it’s a production line, a siloed organization, an intangible process, or a business deploying artificial intelligence, the logic remains the same: performance is always determined at the bottleneck. The obsession with optimizing every resource, every task, or every service only multiplies local maximums that are of no value to the end customer and therefore to the business.

True rationality consists in accepting that progress is never uniform, but sequential: as Goldratt suggests, it is a matter of identifying the constraint, addressing it, and then starting over, because at that point the next constraint in the workflow will appear. It is this path, although long and repetitive, that leads to the global optimum. The law of the local maximum reminds us that when it comes to performance, illusion is easy, but that only a systemic approach avoids mistaking costs for gains.

To answer your questions

Image credit: Image generated by artificial intelligence via ChatGPT (OpenAI)

Related articles:

Eliyahu Goldratt’s fictional interview on infobesity and bottlenecks in knowledge work How can you avoid becoming a bottleneck in the workplace? Just because work is invisible, it doesn’t mean that it can’t be improved Thinking of work as a flow: appealing, but is it realistic? OPEX, AI, automation: what will happen to knowledge workers? Work about work: when the reality of work consists of making things that don’t work work