THE MACHINERY OF UTILIZATION
A Complete Guide to How Capacity Actually Behaves
Why Running Hot Breaks Everything
What follows is not advice.
It is not a productivity framework. Not a scheduling hack. Not five tips for getting more out of your team. Not a lean implementation playbook.
It is mechanism.
The actual machinery that determines what happens when a system approaches its capacity limit. The physics underneath the spreadsheet. The reason a restaurant that looks 90% full is already broken. The reason a team running at 95% produces less than a team running at 70%. The reason the operator who maximizes utilization destroys the thing they are trying to optimize.
Most operators carry a single assumption about capacity. More utilization equals more output. The assumption is wrong. Not slightly wrong. Structurally wrong. And the structure of the wrongness creates a specific failure mode that appears in every business, every team, every kitchen, every server farm, every organization that mistakes busy for productive.
This document is a description of that structure.
What the operator reading it does next is their business.
PART ONE: THE REFRAME
Utilization Is Not Efficiency
The word “utilization” carries a moral weight in most operator minds. High utilization is good. Low utilization is wasteful. A machine sitting idle is a machine losing money. A person with slack in their schedule is a person not pulling their weight.
This frame is inherited from Frederick Taylor. In 1911, Taylor published The Principles of Scientific Management and installed a single idea into the operating system of every business that followed: the purpose of management is to maximize the productive output of every resource, every minute. Taylor measured workers with stopwatches. He broke tasks into atomic units. He optimized each unit for speed. Output at some factories increased by 400%.
The frame worked for a specific kind of system. Assembly lines. Single-product factories. Environments where variability was low, demand was constant, and the cost of idle capacity was the dominant cost.
It does not work for any system where variability exists.
And variability exists in every real system.
The reframe is simple. Utilization is not efficiency. Utilization is a load parameter. It describes how much of a system’s capacity is currently consumed. Efficiency is a ratio of value produced to resources consumed. These are not the same thing. A system can be highly utilized and deeply inefficient. A system can be moderately utilized and extremely efficient.
The confusion between the two is the source of more operational damage than almost any other single conceptual error.
UTILIZATION VS EFFICIENCY
┌──────────────────────────────┐ ┌──────────────────────────────┐
│ │ │ │
│ UTILIZATION │ │ EFFICIENCY │
│ │ │ │
│ How much capacity is │ │ How much value is │
│ currently consumed │ │ produced per unit │
│ │ │ of resource consumed │
│ │ │ │
│ Input metric │ │ Output metric │
│ │ │ │
│ Can be measured by │ │ Can only be measured │
│ looking at the resource │ │ by looking at the │
│ │ │ customer │
│ │ │ │
│ Higher is not │ │ Higher is better │
│ necessarily better │ │ (by definition) │
│ │ │ │
└──────────────────────────────┘ └──────────────────────────────┘
The operator who conflates utilization with efficiency will always push utilization higher. They will interpret idle capacity as waste. They will staff tighter, schedule fuller, run closer to the edge. And then they will discover, through pain, that the edge is not a line. It is a cliff.
PART TWO: THE HOCKEY STICK
The Nonlinear Cost Curve
Queuing theory is the mathematics of waiting. It was formalized in the early twentieth century by Agner Krarup Erlang, a Danish engineer working for the Copenhagen Telephone Exchange. Erlang needed to know how many phone lines were required to handle a given call volume. His answer produced a formula that governs every system where arrivals compete for limited capacity.
The fundamental model is called M/M/1. One server. Random arrivals. Random service times. The formula for average response time is:
R = S / (1 - ρ)
Where S is the average service time and ρ (rho) is utilization, defined as arrival rate divided by service rate.
The denominator is everything.
At 50% utilization, the denominator is 0.5. Response time is 2S. Twice the service time.
At 80% utilization, the denominator is 0.2. Response time is 5S. Five times the service time.
At 90% utilization, the denominator is 0.1. Response time is 10S. Ten times the service time.
At 95% utilization, the denominator is 0.05. Response time is 20S. Twenty times the service time.
At 99% utilization, the denominator is 0.01. Response time is 100S. One hundred times the service time.
At 100% utilization, the denominator is zero. Response time is infinite.
THE HOCKEY STICK CURVE
Response
Time
│
│ ▲
│ ╱
│ ╱
100S │ · · · · · · · · · · · · · · · · · · ·╱· · ·
│ ╱
│ ╱
│ ╱
│ ╱
│ ╱
20S │ · · · · · · · · · · · · · · · ╱ · · · · · ·
│ ╱
10S │ · · · · · · · · · · · · · ·╱· · · · · · · ·
5S │ · · · · · · · · · · · · ╱ · · · · · · · · ·
2S │ · · · · · · · · · ···╱· · · · · · · · · · ·
1S │════════════════╱═══════════════════════════
│
└────────────────────────────────────────────►
0% 20% 40% 60% 80% 90% 95% 99%
UTILIZATION (ρ)
Below the line: service time (the work itself)
Above the line: waiting time (the queue)
This curve is not a suggestion. It is not a tendency. It is a mathematical law that governs every system with variable arrivals and limited capacity. The shape is always the same. Below about 70% utilization, the curve is gentle. Above 80%, it steepens. Above 90%, it goes vertical. The “hockey stick” is the right-hand portion where a small increase in utilization produces an enormous increase in wait time.
The operator looking at a resource running at 85% utilization and a resource running at 95% utilization sees a 10 percentage point difference. The queuing math sees a difference of 3x to 20x in wait time. The dashboard says “ten percent more loaded.” The physics says “functionally broken.”
Why the Curve Exists
The curve is not arbitrary. It emerges from a specific mechanism.
When utilization is low, arrivals find the server free most of the time. They start service immediately. No queue forms.
When utilization is moderate, arrivals occasionally find the server busy. A small queue forms. But the queue drains quickly because there is enough spare capacity between arrivals to catch up.
When utilization is high, arrivals frequently find the server busy. The queue grows. And because arrivals are variable, they sometimes cluster. A cluster of arrivals at high utilization creates a burst that the server cannot absorb. The queue grows faster than it drains. Each new arrival waits behind all the previous ones. The wait compounds.
At near-100% utilization, there is no spare capacity to absorb any burst at all. Every cluster of arrivals creates a queue that never fully drains before the next cluster arrives. The queue grows without bound.
This is not a failure of the server. The server is working as fast as it ever worked. The failure is structural. There is no margin between the arrival rate and the service rate, so any variability in either one produces a queue that cannot recover.
WHY QUEUES FORM
┌──────────────────────────────────────────────────────┐
│ │
│ LOW UTILIZATION (50%) │
│ │
│ Arrivals: · · · · · · · · │
│ Server: ████ ████ ████ ████ ████ ████ │
│ Queue: (empty) (empty) (empty) │
│ │
│ Gaps between arrivals > service time │
│ Server idles between jobs. Queue never forms. │
│ │
└──────────────────────────────────────────────────────┘
┌──────────────────────────────────────────────────────┐
│ │
│ HIGH UTILIZATION (90%) │
│ │
│ Arrivals: · · · ·· · · ·· · · · ·· · · · │
│ Server: ████████████████████████████████████ │
│ Queue: _ __ ____ ______ ____ __ _ │
│ │
│ Gaps between arrivals < service time (often) │
│ Bursts create queues that take a long time │
│ to drain because spare capacity is thin. │
│ │
└──────────────────────────────────────────────────────┘
The mechanism is the same whether the “server” is a kitchen line, a software engineer, a delivery driver, a hiring pipeline, or a customer support queue. Arrivals are variable. Service is variable. When utilization is high, variability kills.
PART THREE: THE KINGMAN EQUATION
Variability Is the Multiplier
In 1961, John Kingman published an approximation formula for the average wait time in a queue with general (non-random) arrival and service distributions. The formula is known as the VUT equation because it decomposes wait time into three independent factors.
W ≈ (ρ / (1 - ρ)) × ((Ca² + Cs²) / 2) × S
V is the variability factor: (Ca² + Cs²) / 2, where Ca² is the squared coefficient of variation of inter-arrival times and Cs² is the squared coefficient of variation of service times.
U is the utilization factor: ρ / (1 - ρ), the hockey stick curve.
T is the average service time: S.
Wait time is the product of all three.
The insight that most operators miss is that variability and utilization are multiplicative, not additive. High variability at moderate utilization produces a moderate queue. Moderate variability at high utilization produces a large queue. High variability at high utilization produces a catastrophic queue. The interaction is the thing that matters.
THE VUT EQUATION
┌──────────────────────────────────────────────────────┐
│ │
│ WAIT TIME │
│ │
│ = │
│ │
│ ┌──────────┐ ┌──────────┐ ┌──────────┐ │
│ │ │ │ │ │ │ │
│ │ V │ × │ U │ × │ T │ │
│ │ │ │ │ │ │ │
│ │Variabil- │ │ Utiliz- │ │ Service │ │
│ │ity │ │ ation │ │ Time │ │
│ │ │ │ │ │ │ │
│ │ Ca²+Cs² │ │ ρ │ │ │ │
│ │ ──────── │ │ ─────── │ │ S │ │
│ │ 2 │ │ 1 - ρ │ │ │ │
│ │ │ │ │ │ │ │
│ └──────────┘ └──────────┘ └──────────┘ │
│ │
│ Reducible Reducible Reducible │
│ by leveling by adding by improving │
│ demand or capacity the process │
│ standardizing │
│ process │
│ │
└──────────────────────────────────────────────────────┘
This matters for the operator because it reveals three independent levers for reducing queue pain. Reduce variability (level the demand, standardize the process). Reduce utilization (add capacity). Reduce service time (improve the process).
Most operators reach for the third lever. Make the process faster. This is the hardest lever and the one with the smallest effect. The first two levers are easier and often more powerful. Reducing variability at 85% utilization can produce the same improvement in wait time as cutting service time in half. Adding 15% capacity at 90% utilization can cut wait time by 70%.
The VUT equation is the operator’s diagnostic tool for any system that has a queue. If customers are waiting, orders are stacking, tickets are aging, or projects are delayed, the answer is somewhere in those three factors. The default managerial instinct to “make people work faster” is almost always the wrong lever.
PART FOUR: THE BOTTLENECK PARADOX
Utilization Only Matters at One Point
Eliyahu Goldratt published The Goal in 1984. The book, written as a novel about a factory manager, introduced the Theory of Constraints. The core insight is deceptively simple.
Every system has a constraint. One resource that limits the throughput of the entire system. The throughput of the system equals the throughput of the constraint. No more. No less.
From this follows a rule that violates every instinct the utilization-maximizing operator has.
Utilization of a non-constraint resource is meaningless.
Running a non-bottleneck station at 100% utilization does not increase system throughput. It increases work-in-progress inventory. The extra output piles up in front of the bottleneck, which cannot process it any faster. The pile-up creates storage costs, coordination overhead, quality degradation from aging inventory, and confusion about what is actually needed.
THE CONSTRAINT PARADOX
STATION A STATION B STATION C
(non-bottleneck) (BOTTLENECK) (non-bottleneck)
Capacity: 100/hr Capacity: 60/hr Capacity: 80/hr
┌──────────────┐ ┌──────────────┐ ┌──────────────┐
│ │ │ │ │ │
│ Running at │ │ Running at │ │ Running at │
│ 100% │──►│ 100% │──►│ 75% │
│ │ │ │ │ │
│ Output: │ │ Output: │ │ Output: │
│ 100/hr │ │ 60/hr │ │ 60/hr │
│ │ │ │ │ │
└──────────────┘ └──────────────┘ └──────────────┘
│
▼
┌──────────────────┐
│ │
│ 40 units/hr │
│ pile up here │
│ as WIP │
│ │
│ Cost: storage, │
│ handling, │
│ obsolescence │
│ │
└──────────────────┘
SYSTEM THROUGHPUT: 60/hr
(determined by the bottleneck, regardless of
how hard the other stations work)
The operator who measures every station’s utilization and tries to maximize all of them is creating a system that is maximally busy and not maximally productive. The busy-ness is real. The productivity is limited by the constraint.
Goldratt’s five focusing steps are: identify the constraint, exploit the constraint (squeeze every unit of throughput from it), subordinate everything else to the constraint (run non-bottlenecks at the pace the bottleneck can absorb, not at their own maximum), elevate the constraint (add capacity to it), and repeat. The word “subordinate” is the one that operators resist. It means deliberately under-utilizing non-bottleneck resources. It means accepting idle time at stations that are not the constraint.
This feels like waste. It is not waste. It is the only configuration in which the system produces maximum throughput without generating the secondary costs of excess work-in-progress.
The operator who understands this asks a different question. Not “are my resources busy?” but “is the constraint fed and flowing?” The first question maximizes utilization. The second maximizes throughput. They are not the same question.
PART FIVE: THE PERISHABILITY TRAP
When Unused Capacity Dies
There is an opposite problem.
Some capacity is perishable. If it is not used by a specific moment, it is gone. Not stored. Not banked. Gone.
An airline seat on a flight that has departed. A hotel room on a night that has passed. A restaurant table during the dinner shift that stayed empty. A ghost kitchen’s labor hour during the lunch rush when no orders came. A consultant’s Tuesday afternoon that had no client call.
These are not like factory inventory, which can sit on a shelf. They expire. The revenue they could have generated is permanently lost the moment the window closes.
The airline industry discovered this problem first. American Airlines, under the leadership of Robert Crandall in the 1980s, built the first large-scale yield management system. The system’s job was to maximize revenue per seat by dynamically adjusting prices based on demand forecasts, booking curves, and cancellation probabilities. American Airlines credited the system with $500 million per year in incremental revenue. Delta built a similar system and credited it with $300 million per year.
The core mechanism is overbooking. Airlines sell more tickets than seats, using statistical models (Poisson distributions, binomial arrival probabilities, Monte Carlo simulations) to predict how many ticket holders will not show up. The models are calibrated to maximize expected revenue while keeping the probability of having to deny boarding below a threshold.
THE PERISHABILITY SPECTRUM
◄──────────────────────────────────────────────────►
DURABLE PERISHABLE
CAPACITY CAPACITY
┌───────────┐ ┌───────────┐ ┌───────────┐ ┌───────────┐
│ │ │ │ │ │ │ │
│ Factory │ │ Software │ │ Service │ │ Airline │
│ output │ │ engineer │ │ slot │ │ seat │
│ │ │ time │ │ │ │ │
│ Can be │ │ Lost but │ │ Lost at │ │ Lost at │
│ stored │ │ fungible │ │ shift end │ │ departure │
│ as │ │ across │ │ │ │ │
│ inventory │ │ projects │ │ │ │ │
│ │ │ │ │ │ │ │
└───────────┘ └───────────┘ └───────────┘ └───────────┘
Cost of Cost of Cost of Cost of
under-use: under-use: under-use: under-use:
LOW MODERATE HIGH VERY HIGH
(stored value) (opportunity) (lost revenue) (lost revenue)
The perishability of the capacity changes the utilization calculus completely. For a durable capacity resource, running at 70% and holding 30% in reserve is often optimal because the reserve absorbs variability. For a perishable capacity resource, every percentage point of unused capacity is a permanent revenue loss. The operator of perishable capacity faces a different optimization problem: maximize utilization without exceeding capacity, using price as the lever.
This is yield management. It is not about posting a price and hoping demand matches. It is about continuously adjusting the price to fill as close to 100% as possible while maintaining a small buffer against overbooking costs. The buffer is not slack. It is insurance.
The ghost kitchen operator lives in this territory. Kitchen labor during the lunch shift is perishable. If three cooks are scheduled and only enough orders arrive to keep two of them busy, the third cook’s wages are a permanent loss. The capacity cannot be stored. But if only two cooks are scheduled and a surge hits, orders back up, delivery times spike, platform ratings drop, and the long-term cost exceeds the short-term savings.
The perishability trap is that both under-utilization and over-utilization are costly, but the costs are asymmetric and context-dependent. The operator who treats all capacity the same, applying the same utilization target everywhere, will be wrong in at least one direction.
PART SIX: THE SLACK FUNCTION
Excess Capacity Is Not Waste
In 1963, Richard Cyert and James March published A Behavioral Theory of the Firm. Among its contributions was the formal concept of organizational slack: the surplus resources that a firm holds beyond what is strictly required for current operations.
Slack appears as excess cash on the balance sheet. As unfilled positions in the budget. As time in the schedule that is not allocated to any project. As inventory beyond current demand. As capacity beyond current load.
The utilization-maximizing operator sees all of these as waste. Cyert and March saw them as a structural feature.
Slack serves three functions that cannot be served by any other mechanism.
Buffer. Slack absorbs variability without disrupting operations. When demand spikes, the slack capacity handles the spike. When a key person is sick, the slack in the schedule absorbs the absence. When a customer pays late, the slack in the cash position absorbs the delay. Without slack, every perturbation propagates through the system and creates a crisis.
Search. Slack provides resources for exploration. Innovation, experimentation, new-market testing, process improvement. None of these can happen when every resource is consumed by current operations. The operator running at 100% utilization has no capacity for anything that is not already scheduled. This means no adaptation. No learning. No response to changing conditions.
Conflict absorption. Slack reduces internal competition for resources. When every dollar and every hour is allocated, every new priority requires displacing an existing one. This creates political conflict, negotiation overhead, and the constant friction of reallocation. Slack reduces this friction by providing uncommitted resources that can be directed to emerging priorities without taking from existing ones.
THE THREE FUNCTIONS OF SLACK
┌───────────────────────────────────────────────────────┐
│ │
│ ORGANIZATIONAL SLACK │
│ (Surplus beyond current requirements) │
│ │
└───────────────────────────────────────────────────────┘
│
┌───────────────┼───────────────┐
│ │ │
▼ ▼ ▼
┌─────────────────┐ ┌─────────────────┐ ┌─────────────────┐
│ │ │ │ │ │
│ BUFFER │ │ SEARCH │ │ CONFLICT │
│ │ │ │ │ ABSORPTION │
│ Absorbs │ │ Funds │ │ │
│ variability │ │ exploration │ │ Reduces │
│ without │ │ and │ │ internal │
│ disruption │ │ adaptation │ │ competition │
│ │ │ │ │ for resources │
│ Without it: │ │ Without it: │ │ │
│ every shock │ │ no innovation │ │ Without it: │
│ becomes a │ │ no learning │ │ every new │
│ crisis │ │ no response │ │ priority is │
│ │ │ to change │ │ a fight │
│ │ │ │ │ │
└─────────────────┘ └─────────────────┘ └─────────────────┘
The research on organizational slack and innovation shows an inverted-U relationship. Too little slack and the firm cannot experiment. Too much slack and the firm becomes complacent, funding projects that should have been killed. The optimal amount of slack depends on the firm’s environment. In stable environments, less slack is needed. In volatile environments, more slack is the difference between adaptation and extinction.
The operator who eliminates all slack in the name of efficiency is optimizing for today at the expense of tomorrow. The system runs leaner. It also runs more fragile. The first unexpected event reveals the fragility.
PART SEVEN: THE INVISIBLE QUEUE
Knowledge Work Hides Its Queues
Donald Reinertsen, in The Principles of Product Development Flow (2009), made an observation that should have restructured how every knowledge-work organization thinks about capacity.
Queues in manufacturing are visible. You can see the pile of work-in-progress sitting on the factory floor. You can count it. You can measure it. You can photograph it and show it to management and everyone agrees the pile is too big.
Queues in knowledge work are invisible.
They live in backlogs. In Jira boards. In email inboxes. In “waiting for review” statuses. In the gap between when a task is ready and when someone picks it up. In the list of features that have been specified but not started. In the list of bugs that have been reported but not triaged.
The pile is just as real as the factory floor pile. It imposes just as much cost. Delay cost. Carrying cost. Obsolescence cost (the feature that sat in the backlog for six months and is no longer relevant when it finally gets built). Coordination cost (the more items in the queue, the more effort required to prioritize and sequence them).
But because the queue is invisible, the operator does not manage it.
VISIBLE VS INVISIBLE QUEUES
MANUFACTURING:
┌───────────┐ ┌───────────┐ ┌───────────┐
│ │ │ ████████ │ │ │
│ Station │────►│ ████████ │────►│ Station │
│ A │ │ ████████ │ │ B │
│ │ │ (visible │ │ │
│ │ │ pile) │ │ │
└───────────┘ └───────────┘ └───────────┘
Everyone can see the pile.
Everyone agrees it is a problem.
KNOWLEDGE WORK:
┌───────────┐ ┌───────────┐ ┌───────────┐
│ │ │ │ │ │
│ Specify │────►│ Backlog │────►│ Build │
│ │ │ (47 │ │ │
│ │ │ items) │ │ │
│ │ │ │ │ │
└───────────┘ └───────────┘ └───────────┘
Nobody sees the pile.
The backlog is "normal."
The 47 items each have a cost of delay.
Total delay cost is invisible.
Reinertsen’s central argument is that the cost of queues in product development is dominated by the cost of delay, and the cost of delay is almost never quantified. His finding: approximately 85% of product managers do not know the cost of delay for the items in their queue. They know the cost of adding a developer (capacity cost). They do not know the cost of a feature sitting in the backlog for an extra month (delay cost).
This asymmetry in visibility produces a systematic bias. Organizations over-optimize on capacity cost (keeping utilization high, not hiring, not adding resources) and under-optimize on delay cost (letting queues grow, letting cycle times extend, letting items age in the backlog). The bias exists because the capacity cost is on the P&L and the delay cost is not.
Reinertsen’s formula for the optimal utilization rate in knowledge work balances capacity cost against delay cost. The result, in most realistic scenarios, is that the optimal utilization rate is between 70% and 85%. Not 95%. Not 100%. The operator running knowledge workers at 95% utilization is generating enormous invisible queues whose delay cost exceeds the savings from not hiring the additional person.
PART EIGHT: THE LEVELING PRINCIPLE
Toyota’s Answer
The Toyota Production System recognized the utilization problem decades before queuing theory entered the management mainstream. Toyota’s answer was not to maximize utilization. It was to level it.
Heijunka is the Japanese word for leveling. It means distributing production volume and mix evenly across available time. Instead of building in batches (all of product A on Monday, all of product B on Tuesday), Toyota builds small quantities of each product throughout each day.
The purpose is to eliminate mura (unevenness). Mura is the root cause of both muri (overburden) and muda (waste). When demand is uneven, the system alternates between overloaded and idle. During the overloaded phase, quality drops, errors increase, equipment breaks, people burn out. During the idle phase, resources sit unused.
BATCHED VS LEVELED PRODUCTION
BATCHED (high peak utilization, high idle time):
Utilization
│
100% │ ████████ ████████
│ ████████ ████████
80% │ ████████ ████████
│ ████████ ████████
60% │ ████████ ████████
│ ████████ ████████
40% │ ████████ ████████
20% │ ████████ ████████
│ ████████ ████████
└──────────────────────────────────────►
Mon-Tue Wed-Thu Fri
(overload) (idle) (overload)
LEVELED (stable utilization, minimal queue):
Utilization
│
100% │
│
80% │ ██████████████████████████████████
│ ██████████████████████████████████
60% │ ██████████████████████████████████
│ ██████████████████████████████████
40% │ ██████████████████████████████████
20% │ ██████████████████████████████████
│ ██████████████████████████████████
└──────────────────────────────────────►
Mon Tue Wed Thu Fri
(steady flow, predictable, resilient)
The leveled system runs at lower peak utilization. It also runs at higher average throughput. Because it never enters the hockey-stick region of the queuing curve. Because it never overburdens equipment or people. Because variability is absorbed by the leveling mechanism rather than transmitted through the system as queues.
The mechanism underneath heijunka is exactly the V factor in Kingman’s equation. By leveling demand, Toyota reduces Ca² (the variability of arrivals). This reduces the V factor. Which reduces wait time. Which increases throughput. Which increases quality. The entire chain follows from reducing variability, not from increasing utilization.
Taiichi Ohno, the architect of the Toyota Production System, said it directly: “The slower but consistent tortoise causes less waste and is much more desirable than the speedy hare that races ahead and then stops occasionally to doze.”
The tortoise runs at moderate utilization. Consistently. Without queues. Without crises. Without burnout.
PART NINE: THE TAYLOR TRAP
The Ideology of Maximum Utilization
Frederick Taylor’s legacy is not a set of techniques. It is an ideology. The ideology says: idle resources are waste; management’s job is to eliminate idle time; the goal is 100% utilization of every resource.
This ideology persists because it maps onto a moral intuition. Hard work is good. Laziness is bad. Busy is virtuous. Idle is suspicious. The operator who sees an employee with nothing to do feels, at a visceral level, that something is wrong. The operator who sees every employee heads-down and rushing feels, at a visceral level, that everything is right.
The feeling is wrong.
Taylor’s approach worked in his specific context. Single-product factories with minimal variability, where the dominant cost was labor and the dominant failure mode was workers deliberately slowing down (“soldiering”). In that context, maximizing utilization was approximately correct.
But the context has changed. Modern operations face high variability. Multiple products. Complex interdependencies. Knowledge work where the output is not linear with the input. Service environments where the customer experiences the queue directly.
In these environments, the Taylor ideology produces a specific pathology. The system runs at 95% utilization. Wait times are extreme (hockey stick). Quality degrades (overburden). Innovation stops (no slack). Adaptation fails (no capacity for response). And the operator, seeing all resources busy, concludes that the system is running well. The dashboard confirms the ideology. The customers tell a different story.
THE TAYLOR TRAP
┌──────────────────────────────────────────────────────┐
│ │
│ THE OPERATOR'S VIEW │
│ │
│ "Every resource is busy." │
│ "Utilization is 95%." │
│ "We are running efficiently." │
│ │
└──────────────────────────────────────────────────────┘
│
│ meanwhile
▼
┌──────────────────────────────────────────────────────┐
│ │
│ THE SYSTEM'S REALITY │
│ │
│ Wait times: 20x service time │
│ Queue length: growing │
│ Cycle time: 5x what it could be │
│ Error rate: rising │
│ Employee burnout: accelerating │
│ Innovation: zero │
│ Response to change: none │
│ Customer satisfaction: declining │
│ │
└──────────────────────────────────────────────────────┘
The trap closes because the operator’s response to the symptoms is to push utilization even higher. Customers complaining about wait times? Make people work faster. Quality dropping? Add more inspections (more work, higher utilization). Projects delayed? Cancel slack time and allocate those hours to the backlog. Each response increases utilization, which increases queue times, which increases the symptoms, which triggers the same response.
The loop is self-reinforcing. The exit is to recognize that utilization is the cause, not the solution.
PART TEN: THE TWO FAILURE MODES
The Spectrum Between Waste and Fragility
Every system with capacity sits somewhere on a spectrum. At one end, under-utilization. At the other, over-utilization. Both are costly. The costs are different.
THE UTILIZATION SPECTRUM
◄──────────────────────────────────────────────────────►
UNDER-UTILIZED OVER-UTILIZED
(0% - 50%) (90% - 100%)
┌─────────────────────┐ ┌─────────────────────────┐
│ │ │ │
│ Costs: │ │ Costs: │
│ │ │ │
│ - Carrying cost │ │ - Queue cost │
│ of idle │ │ (wait times) │
│ capacity │ │ │
│ │ │ - Quality │
│ - Revenue lost │ │ degradation │
│ on perishable │ │ │
│ resources │ │ - Burnout │
│ │ │ │
│ - Capital tied │ │ - Zero slack │
│ up in unused │ │ for adaptation │
│ equipment │ │ │
│ │ │ - Fragility │
│ │ │ to any shock │
│ │ │ │
└─────────────────────┘ └─────────────────────────┘
│
▼
┌─────────────────────┐
│ │
│ OPTIMAL ZONE │
│ (context- │
│ dependent) │
│ │
│ Durable capacity: │
│ 70% - 85% │
│ │
│ Perishable: │
│ 85% - 95% │
│ │
│ Bottleneck: │
│ as high as │
│ possible │
│ │
│ Non-bottleneck: │
│ subordinated │
│ to bottleneck │
│ │
└─────────────────────┘
The optimal utilization rate is not a single number. It depends on the type of capacity.
For durable capacity with high variability (a software team, a consulting practice, a support queue), the optimal rate is 70% to 85%. The remaining 15% to 30% is the buffer that keeps the system in the gentle region of the hockey stick curve.
For perishable capacity with high fixed cost (airline seats, hotel rooms, event venues), the optimal rate is 85% to 95%, managed through dynamic pricing. The remaining 5% to 15% is the insurance against overbooking costs.
For the system bottleneck (Goldratt’s constraint), the optimal utilization is as high as possible because every minute of lost throughput at the bottleneck is lost throughput for the entire system. The constraint should never be idle. Everything else subordinates to keep it fed.
For non-bottleneck resources, the optimal utilization is whatever the bottleneck’s pace demands. Running them faster creates work-in-progress that the bottleneck cannot absorb.
The operator who applies one utilization target to all resources is wrong on most of them. The operator who diagnoses each resource’s type, variability profile, and position in the constraint chain can set utilization targets that actually optimize the system.
PART ELEVEN: SYNTHESIS
The Unified Framework
The machinery underneath utilization is one structure repeated at different levels.
At the physics level, queuing theory describes the nonlinear relationship between load and wait time. The hockey stick curve is universal. It applies to every system with variable arrivals and limited capacity.
At the variability level, Kingman’s equation shows that wait time is the product of variability, utilization, and service time. Reducing any of the three reduces the queue. The cheapest lever is usually variability or utilization, not service time.
At the system level, Theory of Constraints shows that only the bottleneck’s utilization determines system throughput. Non-bottleneck utilization is either irrelevant or harmful.
At the organizational level, slack is not waste. It is the mechanism by which the system buffers, adapts, and resolves internal conflict.
At the scheduling level, leveling (heijunka) reduces the variability factor and keeps the system out of the hockey-stick region.
At the ideological level, the Taylor legacy equates busy with productive and creates a self-reinforcing loop that pushes utilization toward the point of system breakdown.
THE FULL STACK
┌──────────────────────────────────────────────────────┐
│ LEVEL 6: IDEOLOGY │
│ The Taylor assumption. Busy = productive. │
│ Must be overridden before any fix takes hold. │
└──────────────────────────────────────────────────────┘
│
▼
┌──────────────────────────────────────────────────────┐
│ LEVEL 5: ORGANIZATIONAL SLACK │
│ Buffer, search, conflict absorption. │
│ The mechanism that makes adaptation possible. │
└──────────────────────────────────────────────────────┘
│
▼
┌──────────────────────────────────────────────────────┐
│ LEVEL 4: CONSTRAINT POSITION │
│ Bottleneck vs non-bottleneck. │
│ Only the constraint's utilization matters. │
└──────────────────────────────────────────────────────┘
│
▼
┌──────────────────────────────────────────────────────┐
│ LEVEL 3: VARIABILITY │
│ Kingman's V factor. Arrival variance x service │
│ variance. The multiplier on the utilization curve. │
└──────────────────────────────────────────────────────┘
│
▼
┌──────────────────────────────────────────────────────┐
│ LEVEL 2: CAPACITY TYPE │
│ Durable vs perishable. Determines the cost │
│ profile of under-utilization. │
└──────────────────────────────────────────────────────┘
│
▼
┌──────────────────────────────────────────────────────┐
│ LEVEL 1: THE PHYSICS │
│ Queuing theory. R = S / (1 - ρ). The hockey stick. │
│ Non-negotiable. Applies to every loaded system. │
└──────────────────────────────────────────────────────┘
Each level constrains the one above it. The operator who tries to solve a scheduling problem (level 5) without understanding the physics (level 1) will schedule for maximum utilization and wonder why everything is late. The operator who understands the physics but not the constraint position (level 4) will add capacity to the wrong station. The operator who understands the constraint but not the ideology (level 6) will fix the system and then watch the organization revert to maximum utilization within six months because the ideology was never addressed.
The fix has to start at the bottom and propagate upward. See the physics. Identify the constraint. Classify the capacity. Measure the variability. Protect the slack. Override the ideology.
The Central Asymmetry
The deepest observation about utilization is an asymmetry in costs.
The cost of under-utilization is linear. Each percentage point of unused capacity costs roughly the same as the last. Ten percent idle costs ten percent of capacity cost.
The cost of over-utilization is exponential. Each percentage point above 80% costs more than the last. The last five percentage points (95% to 100%) cost more in queue time, burnout, and system degradation than the first fifty percentage points of utilization combined.
THE COST ASYMMETRY
Total
System
Cost
│
│ ▲
│ ╱
│ ╲ ╱
│ ╲ ╱
│ ╲ ╱
│ ╲ ┌───────┐ ╱
│ ╲ │ │ ╱
│ ╲ │OPTIMAL│ ╱
│ ╲──│ ZONE │────╱
│ │ │
│ └───────┘
│
└────────────────────────────────────────────►
0% 30% 60% 80% 95% 100%
UTILIZATION
Left curve: cost of idle capacity (linear)
Right curve: cost of queuing + degradation (exponential)
Optimal zone: where total cost is minimized
This asymmetry means that the penalty for being slightly over the optimal utilization rate is much larger than the penalty for being slightly under it. Being five points under costs five points of capacity. Being five points over costs a multiple of queue time, quality loss, and organizational fragility.
The safe direction of error is under-utilization. The dangerous direction is over-utilization. The operator who understands this carries a margin. Not because slack is comfortable. Because the physics of queuing makes the right side of the curve catastrophically expensive.
PART TWELVE: OPERATOR NOTES
Pattern-Level Observations
The following observations are pattern-level. They describe things that repeatedly appear in systems the operator may encounter. They are not prescriptions. They are descriptions of regularities.
The default organizational response to delay is to increase utilization. When projects are late, the instinct is to allocate more hours, cancel slack time, and push people to work harder. This increases utilization, which increases queue time, which increases delay. The response worsens the problem it is trying to solve. The correct response is to reduce work-in-progress, not to increase effort.
Measuring utilization by hours worked is measuring the wrong thing. Hours worked is an input metric. Throughput is an output metric. A team working sixty-hour weeks at 95% utilization with massive queues will produce less throughput than a team working forty-hour weeks at 75% utilization with short queues. The second team finishes more work because each item spends less time waiting.
The most expensive resource in the system should have the highest utilization. Goldratt’s insight applies directly. If the constraint is a $200/hour specialist, their utilization should be maximized because every idle hour costs $200 in lost system throughput. But the $30/hour resources feeding the specialist should run at whatever pace keeps the specialist continuously fed. Not faster. Faster creates inventory the specialist cannot absorb.
Ghost kitchen labor is perishable capacity with high variability. The lunch rush is a two-hour window. Kitchen labor during that window is perishable. Kitchen labor outside that window is either prep work (durable) or idle (perishable). The utilization strategy differs by hour. During peak, maximize throughput on the constraint (usually the line, sometimes the delivery queue). Outside peak, use slack for prep, training, and maintenance. Do not schedule peak-level staffing for non-peak hours. Do not schedule non-peak staffing for peak hours.
Knowledge work has the highest invisible queue costs. Software development, product management, design, and marketing all generate queues that never appear on a factory floor. The backlog is the queue. The “waiting for review” status is the queue. The “blocked by dependencies” state is the queue. Each item in the queue has a cost of delay that is almost never calculated. Reinertsen’s finding that 85% of product managers cannot state the cost of delay for their items is damning. It means the most expensive queues in the organization are the ones that are never managed.
Adding capacity to a non-bottleneck is waste. Hiring a second designer when design is not the constraint will not accelerate product delivery. It will produce more designs sitting in the backlog waiting for engineering. The pile looks like productivity to the design team. It is work-in-progress inventory to the system.
Variability reduction often outperforms capacity addition. Leveling demand (spreading orders across the day instead of letting them cluster), standardizing processes (reducing service time variance), and managing arrival patterns (appointment systems, batch scheduling, pre-ordering) all reduce the V factor in Kingman’s equation. These interventions are often cheaper and more effective than adding capacity.
The 85% rule exists for a reason. Across industries, the empirical optimal utilization for systems with moderate variability converges on approximately 80% to 85%. Manufacturing plants target this range. Federal Reserve monitoring flags capacity utilization above 85% as inflationary. Hospital ERs begin to degrade above 85% occupancy. Software systems are provisioned for 70% to 80% peak utilization. The convergence is not coincidence. It is the hockey stick telling every domain the same thing.
Every crisis reveals utilization. When the system is running at 95% and a shock arrives (a key person quits, a supplier fails, a demand spike hits, a pandemic reshapes the operating environment), there is no capacity to absorb it. The system breaks. Every post-mortem of an operational crisis, if it goes deep enough, will find that the system was running too hot before the crisis hit. The crisis did not cause the failure. It revealed the fragility that high utilization had already created.
On the Operator Profile
The operator reading this has felt the utilization problem even if they have never named it. The team that is always busy and always behind. The kitchen that is always full and always slow. The product backlog that grows faster than the team can drain it. The feeling that adding more work to the schedule is like pouring water into a cup that is already full.
The machinery is the same across all of these. Variable arrivals competing for limited capacity. The hockey stick curve determining the wait time. The bottleneck determining the throughput. The slack (or absence of slack) determining the resilience.
The operator who sees the machinery stops doing the thing that most operators do, which is to push utilization higher when things are slow and wonder why they break when things get busy. The operator who sees the machinery looks at utilization as a dial, not a virtue. Turns it up where the capacity is perishable and the constraint demands it. Turns it down where variability is high and the cost of queuing exceeds the cost of idle capacity.
| This is the same operating principle described in [[THE_MACHINERY_OF_BOTTLENECKS | The Machinery of Bottlenecks]]: the constraint determines the throughput. Utilization is the load placed on the constraint and on everything around it. The relationship between load and performance is not linear. It is a hockey stick. |
| The capacity to hold this view, to accept that idle resources can be the correct configuration, is the same capacity described in [[THE_MACHINERY_OF_SLACK | The Machinery of Slack]]. Slack is not the absence of work. Slack is the presence of capacity for response. The two look identical from the outside. The operator who can distinguish between them makes different decisions from the operator who cannot. |
| The felt pull toward maximizing utilization, toward filling every hour, toward scheduling every resource, toward eliminating every gap, is itself an instance of the mechanism described in [[THE_MACHINERY_OF_CONSTRAINTS | The Machinery of Constraints]]. The ideology of maximum utilization is a constraint on the operator’s thinking. It limits the solution space. It excludes the configurations that would actually optimize the system. Seeing the ideology as a constraint is the first step toward removing it. |
CITATIONS
Queuing Theory and the Hockey Stick Curve
Erlang, A. K. (1909). “The Theory of Probabilities and Telephone Conversations.” Nyt Tidsskrift for Matematik B, 20:33-39.
Kingman, J. F. C. (1961). “The single server queue in heavy traffic.” Mathematical Proceedings of the Cambridge Philosophical Society, 57(4):902-904.
Kingman, J. F. C. (1962). “On queues in heavy traffic.” Journal of the Royal Statistical Society: Series B, 24(2):383-392.
Little, J. D. C. (1961). “A proof for the queuing formula: L = λW.” Operations Research, 9(3):383-387.
Theory of Constraints
Goldratt, E. M. (1984). The Goal: A Process of Ongoing Improvement. North River Press.
Goldratt, E. M. (1990). The Theory of Constraints. North River Press.
Goldratt, E. M. (1997). Critical Chain. North River Press.
Theory of Constraints Institute. “Theory of Constraints.” https://www.tocinstitute.org/theory-of-constraints.html
Product Development Flow and Cost of Delay
Reinertsen, D. G. (2009). The Principles of Product Development Flow: Second Generation Lean Product Development. Celeritas Publishing.
Reinertsen, D. G. (1997). Managing the Design Factory: A Product Developer’s Toolkit. Free Press.
Organizational Slack
Cyert, R. M., & March, J. G. (1963). A Behavioral Theory of the Firm. Prentice-Hall.
Nohria, N., & Gulati, R. (1996). “Is slack good or bad for innovation?” Academy of Management Journal, 39(5):1245-1264.
Tan, J., & Peng, M. W. (2003). “Organizational slack and firm performance during economic transitions: Two studies from an emerging economy.” Strategic Management Journal, 24(13):1249-1263.
Toyota Production System and Leveling
Ohno, T. (1988). Toyota Production System: Beyond Large-Scale Production. Productivity Press.
Liker, J. K. (2004). The Toyota Way: 14 Management Principles from the World’s Greatest Manufacturer. McGraw-Hill.
Womack, J. P., Jones, D. T., & Roos, D. (1990). The Machine That Changed the World. Free Press.
Scientific Management
Taylor, F. W. (1911). The Principles of Scientific Management. Harper & Brothers.
Revenue Management and Yield Management
Talluri, K. T., & van Ryzin, G. J. (2004). The Theory and Practice of Revenue Management. Springer.
Subramanian, J., Stidham, S., & Lautenbacher, C. J. (1999). “Airline yield management with overbooking, cancellations, and no-shows.” Transportation Science, 33(2):147-167.
Smith, B. C., Leimkuhler, J. F., & Darrow, R. M. (1992). “Yield management at American Airlines.” Interfaces, 22(1):8-31.
Energy Proportional Computing
Barroso, L. A., & Holzle, U. (2007). “The case for energy-proportional computing.” IEEE Computer, 40(12):33-37. https://www.barroso.org/publications/ieee_computer07.pdf
Barroso, L. A., Clidaras, J., & Holzle, U. (2013). The Datacenter as a Computer: An Introduction to the Design of Warehouse-Scale Machines. Morgan & Claypool Publishers.
Capacity Utilization Economics
Federal Reserve. Capacity Utilization: Total Industry (TCU). FRED Economic Data. https://fred.stlouisfed.org/series/TCU
Berndt, E. R., & Morrison, C. J. (1981). “Capacity utilization measures: underlying economic theory and an alternative approach.” American Economic Review, 71(2):48-52.
Document compiled from primary source research across queuing theory, operations management, organizational behavior, and production systems literature. Every structural claim traces to a named primary source.
Related Machineries
-
[[THE_MACHINERY_OF_BOTTLENECKS The Machinery of Bottlenecks]]. The constraint determines throughput. Utilization only matters at the constraint. Non-bottleneck utilization is a different variable entirely, governed by subordination rather than maximization. The two machineries are complementary views of the same system. -
[[THE_MACHINERY_OF_SLACK The Machinery of Slack]]. Slack is the inverse of utilization. Where utilization asks “how much capacity is consumed,” slack asks “how much capacity is available for response.” The machinery of slack describes why excess capacity is an asset. This document describes why maximizing utilization eliminates that asset. -
[[THE_MACHINERY_OF_THROUGHPUT The Machinery of Throughput]]. Throughput is the output metric that utilization claims to maximize. The hockey stick curve shows that maximizing utilization does not maximize throughput. It maximizes queue time. The relationship between the two is nonlinear and often inverse above 85%. -
[[THE_MACHINERY_OF_LATENCY The Machinery of Latency]]. Latency is the customer-facing consequence of the queuing curve. Every point on the hockey stick corresponds to a latency experienced by the person waiting. The machinery of latency describes how that wait time compounds into customer loss, reputation damage, and competitive disadvantage. -
[[THE_MACHINERY_OF_VARIANCE The Machinery of Variance]]. Variability is the V factor in Kingman’s equation. The machinery of variance describes why variability exists and how it propagates. This document shows how variability interacts multiplicatively with utilization to produce queue times that neither factor alone would create.