Bottleneck analysis that survives a change in the mix
Nov 9, 2026, 9 min read

Highest utilisation, longest queue, and largest shadow price identify three different resources, and only the third is defined by the plan you are trying to improve. Annual and monthly averages hide constraints that bind for eight weeks a year, and mix changes move the binding resource outright: in the worked example below, the resource at 85 percent under one mix falls to 58 percent under another while a resource previously at 55 percent goes to 91. Kingman's 1961 queueing approximation explains why a resource sitting at 90 percent already behaves like a constraint even though its load calculation says there is room. Recompute weekly rather than monthly, and count weeks above 85 percent instead of ranking annual averages.
The capital request is for a second filler at 3.4 million, and the justification is one line: the filler runs at 92 percent. It has been in the plan for two years. On-time-in-full has not moved in either of them.
Pull the load by week rather than by year and a different picture appears. The filler is above 90 percent in eleven weeks and below 70 in twenty-six. The weeks where service actually fails are mostly not the eleven. In those weeks the resource that runs out is the packing hall, which sits at 61 percent for the year and never appears in a capital case because nobody ranks resources by anything other than the annual average.
This is the ordinary failure mode of bottleneck analysis. The method is sound, the arithmetic is easy, and the answer depends almost entirely on choices made before any arithmetic happens.
Three definitions, three answers
Ask what a bottleneck is and you get one of three things, which coincide only in simple plants.
The resource with the highest utilisation. Load divided by available time, ranked. It is easy to compute from routings and a plan, and it is the definition built into most capacity reports. It answers the question of which resource is fullest on average.
The resource with the most work waiting in front of it. Measured from work in progress or from queue time. It answers the question of where material actually piles up, which frequently differs from the utilisation ranking because arrival variability matters as much as load.
The resource whose extra hour changes the plan the most. This is the shadow price on the capacity constraint in a capacitated plan: the improvement in the objective per additional hour made available. It is the only one of the three defined by reference to what you are trying to achieve, and it is the number a capital request needs, because it converts directly into an annual value for the hour you would be buying.
The three diverge in a specific and predictable way. A resource can be at 95 percent and have a shadow price of zero, because demand for what it makes is already satisfied and the extra hour would produce something nobody wants. A resource at 70 percent can carry a large shadow price if it gates a product that is short. Most plants only ever run the utilisation ranking.
The average is what hides it
Aggregation over time is the single largest source of wrong answers here, and it is invisible because the arithmetic at each step is correct.
Take a resource with 720 hours available a month and a monthly load of 612 hours, which reports as 85 percent. Now split the month into four weeks and suppose the load distributes as 214, 198, 105 and 95, which is an ordinary pattern in a plant that runs to a month-end push. Available time is 180 hours a week. The first two weeks come out at 119 and 110 percent, and the last two at 58 and 53. The month says there is room. Two of the four weeks are infeasible, and the work that does not fit in them either runs late or displaces something else that then runs late.
The same effect operates across the year. A resource whose annual utilisation is 61 percent can be over 100 percent for eight consecutive weeks in the season, which is where the service misses come from, and the annual ranking puts it near the bottom of the list.
Aggregation across product families does the same in another dimension, concealing that two families compete for the same resource in the same weeks.
The rule that follows is straightforward. Compute load at the finest bucket where you actually control sequence, which in most plants is a week and sometimes a shift, and stop reading monthly utilisation as a feasibility statement.
Eighty-five percent already behaves like a constraint
A load calculation treats capacity as a wall: below 100 percent the work fits, above it the work does not. Real production queues do not behave that way, and the gap explains most of the cases where a resource that reports slack is the one holding the plant back.
Kingman's 1961 approximation gives waiting time at a single server as a variability term multiplied by a utilisation term multiplied by the processing time, where the utilisation term is u divided by one minus u. Hopp and Spearman set the same relationship out for manufacturing in Factory Physics (1996). The utilisation term is what matters here. At 80 percent it is 4. At 90 percent it is 9. At 95 percent it is 19. Queue time grows without bound as utilisation approaches one, and the growth is steep well before it gets there.
A resource at 90 percent is therefore carrying roughly nine times the queue time it would carry at 50 percent, for the same variability. A utilisation report has no column for that queue. Its effects arrive as lead time, work in progress, and an inability to absorb a schedule change, and Little's Law (Little, 1961) converts the work in progress figure into the days of lead time it represents.
The second definition above, the resource with the most work waiting, is therefore finding something real that the first misses. Two resources at identical utilisation behave differently if one receives smooth arrivals and the other a batch every third day. Where you have work in progress data, rank by queue time as well as by load and look at what appears high on one list only.
The bottleneck moves with the mix
Utilisation is a function of what you happen to be making, and different products consume the resources in different proportions. That means the identity of the binding resource is a property of the mix rather than a property of the plant.
A worked case with two resources and two products. Mixing has 160 hours a week available, packing has 200. Product X consumes 2.0 mixing hours and 1.0 packing hour per thousand cases. Product Y consumes 0.8 mixing hours and 2.5 packing hours per thousand cases.
Run 60 thousand cases of X and 20 of Y. Mixing takes 120 plus 16, which is 136 hours, or 85 percent. Packing takes 60 plus 50, which is 110 hours, or 55 percent. Mixing is the constraint and everyone in the plant knows it.
Now run the seasonal mix, 20 thousand cases of X and 65 of Y. Mixing takes 40 plus 52, which is 92 hours, or 58 percent. Packing takes 20 plus 162, which is 182 hours, or 91 percent. Packing is the constraint, mixing has become comfortable, and nothing about the plant changed.
Average the year and mixing comes out around 75 percent against packing at 68, so the annual report names mixing. Mixing is the resource with slack in precisely the months when the plant cannot deliver. Two years of improvement spending goes into the resource that is free when it matters.
Goldratt's five focusing steps end with an instruction to go back to the first step once the constraint is broken, and the reason that instruction exists is this one. The constraint is a moving object, and an analysis run once is a photograph of a mix that has since changed.
The version of the calculation that handles this is not complicated. Compute load by resource for each week of the plan under the mix that plan implies, then count how many weeks each resource exceeds a threshold. Eighty-five percent is a reasonable threshold given the queueing behaviour above. The resource with the most weeks over the line is a better candidate than the resource with the highest mean.
What changes once you know
Identifying the constraint is worth something only if the plan and the measures change in response.
Schedule the constraint first and derive everything else from it. The constraint's sequence sets the rhythm for the plant, upstream resources release material to feed it rather than to hit their own output targets, and downstream resources are sized to absorb whatever it produces.
Protect it with time rather than with quantity. A buffer of work in front of the constraint measured in hours of its own running time is what stops a starve, and buffers scattered across every station are working capital that protects nothing in particular.
Stop measuring non-constraints on local efficiency. A resource upstream of the constraint run at 95 percent to hit an efficiency target converts material into work in progress that queues in front of the constraint, adds lead time, and produces no additional output. The plant records an improvement and delivers less. This is the most common way a bottleneck analysis gets undone after it is completed, because the measurement system was left pointing the other way.
And re-run it. Quarterly is usually enough in a plant with a stable portfolio, and after any launch or delisting that moves more than a few percent of volume.
The constraint that is not in the routing
Load calculations only see resources that appear in routings, and a meaningful share of real constraints do not.
Dock slots and yard space. A single qualified operator or a single set of change parts. A lab test with a five day turnaround sitting between production and release. A supplier who has put you on allocation. A regulatory hold period. A tank that has to be empty before the next grade can run.
Each produces the same symptom: the routing-based load calculation reports the plant as feasible and the plant misses anyway. Represent them as resources with a capacity even when the representation is coarse. A pseudo-resource called release testing with a capacity of forty lots a week is a rough model and is far better than the current one, which assumes infinite capacity.
Where the constraint is a storage volume rather than a rate, capacity interacts with sequence in a way a load calculation cannot express, and the analysis has to move into a scheduling model.
Where this stops
Everything above rests on routing standards, and routing standards decay. Rates set for a previous pack format, cycle times that were optimistic when they were written, and changeover matrices that have never been updated all bias load in one consistent direction. Reconcile standard hours earned against clock hours for one full month before trusting any ranking. Where the two disagree by more than a few percent, fix that first.
Shadow prices come with a specific trap. When two resources bind at the same time the dual solution is not unique, and a genuinely binding constraint can report a shadow price of zero depending on which vertex the solver stopped at. The check is to add an hour to the resource, re-solve, and read the change in the objective directly instead of trusting the reported dual.
Load alone will not settle the question in a plant with high variability. A resource at 80 percent receiving lumpy arrivals can be the practical constraint while one at 95 percent with smooth arrivals is not, and no amount of refinement to the load calculation captures that. Where variability is the issue, the queue-based view is the one to follow.
Some constraints are people. A crew that can run either of two lines but not both at once has a structure no capacity number describes, and cross-training changes its shape rather than its size.
The honest limit is the one underneath all of it. A bottleneck analysis tells you which resource binds under the mix you assumed. It inherits every weakness of the demand plan that produced that mix, and a constraint identified against a forecast that is wrong about the seasonal shift will be the wrong constraint. Run the analysis against two or three plausible mixes rather than one, and pay attention to whether the answer holds across them.
Take last year's actual output, recompute load by resource in weekly buckets rather than monthly, and count how many weeks each resource went over 85 percent. If that ranking disagrees with the one your capacity report produces, the disagreement is where to start.
More on this in manufacturing on the DemandLabs platform.




