Two teams do the same work. One runs flat out — utilisation at 95%, every dashboard tile green, nobody idle. The other runs at 80%, looks a little slack, and quietly finishes the same work faster. The busier team is the slower team, and almost no one in it can see why.
This is the measurement trap. The numbers that feel most rigorous — utilisation, local efficiency, cost-per-unit — all reward one thing: keeping every part busy. And a busy part, past a point, is a slow part.
Why "keep everyone busy" feels so right
Idle time looks like waste. A person waiting, a machine stopped, a team at 70% — every instinct and most cost models say fill it. Utilisation is easy to measure, easy to compare, and feels like diligence. So we manage to it: load every resource up, drive occupancy toward 100%, and report the greens.
The problem is that utilisation is an input, not an outcome. It tells you how busy a resource is — not how quickly work reaches the customer. And those two things pull in opposite directions far sooner than anyone expects.
The maths nobody puts on the dashboard
Here is the part that changes how you read every efficiency metric. The relationship between how busy a resource is and how long work waits for it is not a straight line — it is a curve that bends upward and then shoots for the sky. MIT's operations-management notes put it bluntly: "The relationship between waiting time and capacity utilization is strongly non-linear" — in a simple queue, the wait scales with ρ/(1−ρ), so as utilisation nears 100%, the queue heads toward infinity [3].
In plain terms: going from 50% to 80% busy roughly quadruples the wait; 80% to 90% doubles it again; and the last few points before 100% are catastrophic. Drag the curve and feel it.
The Busyness Curve
The Busier a Resource Gets, the Slower the Whole System Runs
Drag utilisation up and watch the wait explode as it nears 100%
At 75% utilisation with normal variation, the function is room to flow: busy enough — and still fast. Relative waiting time is about 3.0 times a steady, half-busy baseline. There is enough slack to absorb the normal ups and downs of demand, so work keeps moving. Utilisation looks "incomplete" on a dashboard, but lead time is short and predictable — the number your customer actually feels.
Illustrative model (wait ∝ ρ/(1−ρ)). Spikier work makes the same utilisation far more costly.
Room to flow
Busy enough — and still fast
Relative wait
3.0×
There is enough slack to absorb the normal ups and downs of demand, so work keeps moving. Utilisation looks "incomplete" on a dashboard, but lead time is short and predictable — the number your customer actually feels.
Hold here. Resist the urge to "tighten up" a system that is already delivering fast, predictable throughput.
Two things move the curve. The first is utilisation — how full you run. The second is variability: the spikier and less predictable the work, the more waiting time you get at any level of busyness. A steady, predictable operation can run hot safely. A lumpy one cannot — which is exactly why "just push utilisation up" fails in the messy real world.
It's not just theory — a hospital runs on this
The clearest proof is somewhere lives depend on it. Across 143 NHS Trusts, where beds ran at an average occupancy of 93.2%, researchers found that each additional 1% of bed occupancy was associated with a 9.5 percentage-point fall in the chance of meeting the four-hour A&E target — and the relationship was non-linear, deteriorating faster the fuller the hospital got [2].
Each 1-point rise in hospital bed occupancy cut the chance of hitting the four-hour A&E target by 9.5 points, in a non-linear relationship — the system slows sharply the busier its beds are kept.
A bed sitting empty looks like waste on a utilisation report. It is also the slack that lets the next emergency flow straight through instead of queuing. Maximise the local metric and you lose the outcome the patient — and the system — actually feels.
The dashboard that lies by telling the truth
None of these numbers are fake. That is what makes the trap so hard to see: every local tile can be truthfully green while the whole is quietly failing. Flip the lens below from the local view to the end-to-end view, and watch what each "win" was hiding.
Operations dashboard
Everything looks on target
Team utilisation
96%
Fully productive
Cost per case
−12%
Efficiency up
Agent occupancy
92%
No idle time
Cases closed / person
+18%
Throughput up
End-to-end lead time
Hidden in local view
—
This is Goodhart's law on an operations board: when a measure becomes the target, it stops being a good measure. Push utilisation, cost-per-case and closures-per-person hard enough and people will optimise exactly those — by letting work pile up in the gaps the dashboard never shows.
The same trap at economy scale
If this were only about queues, it would still be worth fixing. But the pattern — parts that each look optimal adding up to a whole that is not — runs much deeper. Economists call it misallocation. Studying manufacturing plants in China and India, Hsieh and Klenow found that simply moving capital and labour to where they would be most productive — to the efficiency the US already achieves — would raise total manufacturing productivity by 30–50% in China and 40–60% in India [1]. Every firm can be locally rational while the whole economy leaves a third to a half of its output unclaimed.
Whether it is a queue, a factory or an economy, the lesson is the same: optimise the parts in isolation and you can starve the whole.
Can you take slack too far?
Yes — and it is worth saying plainly, because "leave slack" is easy to misread as "relax." Running everything at 60% is its own waste; slack has a cost, and unused capacity you never need is just expense. The goal is not minimum utilisation. It is utilisation matched to variability, with the buffer concentrated where demand is spiky and the cost of delay is highest.
And busyness is only the first place flow leaks. Once you stop overloading the parts, the next tax is the handoffs between them — the seam tax, where value is lost in the gaps no single team owns. The measurement trap is what hides both: a dashboard full of greens, and a customer still waiting.
How to re-measure around flow
Escaping the trap is not a tooling project; it is a change to what you put on the board and reward. These four moves re-instrument an operation around the number the customer actually feels.
Four moves from busyness to flow
Re-instrument the operation around the number the customer feels
Utilisation is an input — it tells you how busy a resource is, not how quickly work reaches the customer. Put lead time (start to done) and flow efficiency (the share of that time the work is actually moving, not waiting) on the same board, and the trade-off becomes visible.
Why it works
In most operations the work spends far more time waiting than being worked on. A dashboard that only shows busyness is blind to exactly the time the customer feels most.
In practice
For one end-to-end flow, measure elapsed lead time and estimate how much of it is active work versus waiting. Review that ratio, not just output counts.
The busiest-looking operation is rarely the fastest. The measurement trap rewards motion; what you want is flow. So the question to carry out of here is not "is everyone busy?" — it is "is the work moving?" They are not the same question, and only one of them is the business your customer experiences.
Sources
- National Bureau of Economic Research. Misallocation and Manufacturing TFP in China and India (Working Paper 13290). 2007.Using micro-data on manufacturing establishments, Hsieh and Klenow measure sizable gaps in the marginal products of labour and capital across plants within narrowly-defined industries. 'When capital and labor are hypothetically reallocated to equalize marginal products to the extent observed in the U.S., we calculate manufacturing TFP gains of 30-50% in China and 40-60% in India.' Each firm can look locally sensible while the whole economy leaves a third to a half of its productivity on the table — misallocation of the parts against the whole.View source
- London School of Economics (LSE Research Online). Spillover effects of inpatient bed capacity on Accident & Emergency performance in England. 2020.Longitudinal analysis of 143 NHS Trusts (June 2016–October 2019), where average bed occupancy was 93.2%. The study finds 'a 1% increase in bed occupancy was associated with a 9.5 percentage point decrease in the Trusts' probability of meeting the waiting target', and reports a non-linear relationship in which performance deteriorates sharply as occupancy climbs — a real-world case of high utilisation throttling end-to-end flow.View source
- MIT OpenCourseWare (15.760 Operations Management). Capacity and Queueing — Class 3 lecture notes. 2004.MIT's operations-management teaching notes state the core result plainly: 'The relationship between waiting time and capacity utilization is strongly non-linear.' In a single-server queue the expected wait scales with ρ/(1−ρ), so as utilisation (ρ) approaches 100%, queue length and waiting time rise toward infinity — and variability amplifies the effect.View source