ATTENTION, PEOPLE WHO HAVE FIXED A BACKLOG BY SCHEDULING EVERYONE HARDER!
PEOPLE WHO SEE A LINE OF WORK AND THINK, “EXCELLENT, THE WORK IS NOW IN ONE CONVENIENT PLACE!” PEOPLE WHO HAVE CALLED A SYSTEM EFFICIENT BECAUSE NOBODY HAS A FREE MINUTE TO DRINK WATER!
Are you tired of not watching a queue metastasise across the building?
Has another perfectly ordinary week become a festival of urgent messages, because the schedule was packed, the work kept arriving, and somebody proposed the revolutionary remedy of packing the schedule further?
Then congratulations.
You qualify for the Five Number One Rules of Operations Before the Queue Eats the Building, an all-purpose flow-preservation kit containing not one, not two, but FIVE NUMBER ONE RULES.
“Wait,” you’re yelling at the dashboard, “they can’t all be Number One!”
That’s exactly what Big Numbering wants, while it sells a single utilisation target as an all-inclusive vacation from reality.
Operations failures are specialists.
Spare capacity won’t identify the bottleneck.
A precise bottleneck measurement won’t explain bursty arrivals.
And a buffer won’t prove that a staffing ratio copied from somewhere else belongs here.
Five different routes by which ordinary work becomes an indoor weather event.
So here they are.
Five rules.
Five Number Ones.
No substitutions. No “we’re all busy.” No stapling another task onto the clipboard and calling it throughput.
RULE #1: LEAVE CAPACITY FOR VARIABILITY
Introducing MAXIMUM-BUSYNESS FOREVER™, the operations package that converts every idle moment into a moral failure and every lunch break into a suspicious absence from the dashboard.
Don’t run a queueing resource above roughly 80–85% utilisation when wait time matters.
Not because an empty-looking interval is wasteful. Because queueing delay is nonlinear, and that changes everything about how the last few points behave.
Mean wait scales with ρ/(1−ρ). Watch what that does.
At 50% utilisation, the ratio is 1.
At 80%, it’s 4.
At 90%, it’s 9.
At 95%, it’s 19.
The work didn’t double between 90 and 95. The waiting did.
As utilisation approaches 1, the denominator shrinks toward zero and the wait climbs toward a vertical asymptote. There is no gentle warning stretch before that happens, which is exactly why a packed schedule feels fine right up until it doesn’t.
A system at moderate utilisation absorbs a late job, a long call, a burst. A system pressed toward 100% has nowhere to put the surprise except behind work that’s already waiting.
And it doesn’t take a crisis to trigger it. Only the ordinary variation the schedule declared impossible.
Kingman’s approximation names the rest of the mechanism: utilisation is one multiplier, arrival variation is another, and service-time variation is a third.
Utilisation isn’t the whole machine. Which is why a calm average can sit on top of a feverish queue.
Capacity is not unused labour. It is where variation goes instead of into the queue.
The calendar may be full. The system may be full. Those are not the same achievement.
RULE #1: MEASURE THE BOTTLENECK BEFORE ADDING WORK
From the makers of “EVERY DEPARTMENT IS THE BOTTLENECK” comes MORE-WORK-IN-A-BOX™ — simply add arrivals and wait for the universe to confess!
Before adding work, measure the flow.
Little’s Law: the average number of items in a system equals the arrival rate times the average time each spends there. Ten customers an hour, each staying half an hour, means five customers in the store at any moment.
Simple. And unusually well-behaved — it needs no assumption about Poisson arrivals, exponential service times, or queue discipline.
What it does need is stable long-run averages.
Which is the trap. A queue still filling, or still draining, isn’t a steady-state photograph. Its momentary average is not a certificate of normal operations, however calm the number looks.
So measure long enough to distinguish a stable system from a live evacuation of work into the hallway.
Count work in process, arrival rate, elapsed time. Then ask which resource or handoff actually controls departures.
Because adding arrivals to a constrained system increases the amount in flight, the time each item takes, or both. It does not persuade the constraint to grow a second pair of hands.
The same targeting logic runs through inventory. ABC analysis typically puts around 20% of items at 70–80% of value, and about half of them at 5–10%.
Conventional breakpoints, not an eternal formula. What they buy is attention aimed where a stock-policy error costs something.
A backlog is evidence about a constraint, not an invitation to feed it.
More work isn’t a cure for slow flow. It’s slow flow with a larger mailing address.
RULE #1: SEPARATE ARRIVAL VARIANCE FROM AVERAGE VOLUME
BEHOLD AVERAGE-IS-EVERYTHING™ — the charting system that takes a storm, averages it with Tuesday, and reports that both days were pleasantly damp.
An average arrival rate says how much work turns up over a period.
It says nothing about whether that work arrives evenly, in clumps, or in a stampede immediately before the resource needs a break.
Kingman puts arrival variation and service-time variation right beside utilisation, as separate multipliers, because both create waiting independently.
Which means two systems with identical mean volume can behave completely differently. The one with burstier arrivals or more variable handling times produces more delay — from the same total work.
Don’t call that a demand surprise afterwards. It was visible the whole time, in a statistic nobody was keeping.
So measure the arrival pattern separately from its mean, and handling-time variation separately from its average. Only then decide whether the resource needs more capacity, smoother arrivals, more predictable work, or a model that actually represents the burst.
Call centres make this painfully billable. Erlang C computes the chance an arriving call has to wait, and a worked example — 200 calls an hour, three-minute handle time, an 80%-in-20-seconds target — lands on around 14 raw agents, then 20 once you apply 30% shrinkage for holidays, training and everything else that isn’t answering calls.
Those are model outputs under stated assumptions. Not a staffing tattoo.
And the assumption worth knowing is the last one: Erlang C assumes callers never abandon the queue.
Real callers hang up. Where abandonment matters, Erlang C systematically overstates the agents needed for a given service level. Erlang A adds abandonment; a simulation can carry more of the local behaviour still.
The average tells the system how much water arrives. Variability tells it whether the pipe floods.
Never let a mean convince you the surge didn’t happen. The queue has receipts.
RULE #1: HOLD INVENTORY FOR A NAMED FAILURE MODE
NOW AVAILABLE: EXTRA-STUFF-AROUND™, the storage philosophy in which every shelf holds a mysterious quantity of just in case, maintained by folklore, adhesive labels, and one person’s increasingly distant memory.
Name the risk before holding the inventory.
Safety stock covers variation in demand during lead time — computed as a service-level factor times the standard deviation of demand over that lead time. Roughly 1.65 standard deviations for a 95% cycle service level, about 2.33 for 99%.
That’s a buffer with a job description. Not a compliment paid to the warehouse.
And it comes with assumptions. The formula treats demand during lead time as roughly normal, which fits high-volume steady items and fails badly for intermittent or lumpy demand — where long runs of zero sit next to a spike, and a normal distribution quietly understates the risk.
When demand and lead time both vary, use the combined-variance calculation rather than demand variation alone.
Then make the buffer operational with a reorder point: average daily demand across the lead time, plus the safety stock. In a two-bin system, the reserve bin is sized to exactly that.
You’ll also meet the 50% rule — hold half of average lead-time demand as buffer. That’s a field shortcut, not the statistical formula, and it’s fine as long as everybody knows which one they’re using.
Keep EOQ in its own box, because it answers a different question entirely. Order quantity balances ordering cost against holding cost, assuming constant known demand, fixed lead time, no discounts, and no stockouts.
Safety stock protects against a disruption. EOQ optimises a cost tradeoff. Neither substitutes for the other, and confusing them produces a warehouse that is both expensive and unprotected.
Inventory is not a confidence blanket. It is a buffer against a named way the plan can fail.
If nobody can say what the extra units are for, that isn’t safety stock. It’s a hostage situation with pallets.
RULE #1: RECALIBRATE THE RATIO AGAINST LOCAL DATA
AND NOW, COPY-PASTE-STAFFING™: a ratio from somewhere else, laminated for confidence, with all the local arrivals, case mix, shrinkage and abandonment carefully removed for your convenience.
First, identify what kind of number you’re copying. They are not all the same species.
A call-centre occupancy target of 80–85% is accumulated industry convention — not an Erlang C output. Rates above 90% get flagged for burnout and attrition risk, because agents have too little recovery between calls.
The 80/20 service level — 80% of calls answered in 20 seconds — is likewise a durable telecom convention. Not a controlled finding that 20 seconds is the universal limit of human patience.
Other numbers carry legal force rather than model authority. California’s Title 22 sets minimum nurse-to-patient ratios: 1:2 in intensive care, 1:1 in the operating room, 1:5 in medical-surgical units. Oregon’s statute closely mirrors those, with fines up to $5,000 per violation.
Those are legislated minimums. Whether they’re locally adequate remains a clinical acuity question, and the statute doesn’t claim otherwise.
And even a model-derived threshold can travel badly. A discrete-event simulation calibrated to a 200-bed hospital found emergency-admission failure near zero below 85% occupancy, around 1% at 90%, and about 19% at 100%.
That last jump is the finding. But its validation restricts it to a similar bed configuration and case mix — which is what makes it useful, and what makes carrying it into a different hospital numerical cosplay rather than benchmarking.
Local data are necessary. Familiar sample-size folklore doesn’t make them sufficient.
The n ≥ 30 shortcut has no derivation making a normal approximation reliable regardless of population shape; heavy skew can require hundreds of observations. And estimating a proportion at 95% confidence with a 5-point margin needs around 385 — a calculation for those specific inputs, not a permission slip for every staffing question.
Zero observed failures deserve the same restraint. Zero incidents in thirty trials still leaves an upper bound near 10%.
That may mean the rate is small. It may only mean the sample hasn’t earned a conclusion yet.
A ratio becomes local knowledge only after local variation, scope, and sample strength have had their say.
Borrow the question. Don’t borrow the answer with the labels still on it.
BUT WAIT, THERE’S MORE!
“What if the queue is a hospital, a call centre, a workshop, a restaurant, or a software request system?”
Same rules. Leave room for variation. Measure stable flow and find the constraint. Separate averages from variability. Buffer a named risk. Recalibrate every borrowed threshold.
“What if there’s already a backlog?”
Same rules — and don’t mistake a growing queue for steady-state evidence. Measure the flow, then determine whether arrivals exceed capacity, whether variation is driving the delay, or whether the schedule is simply too tight.
“What if the number comes from a law, a consultant, a famous operations book, or an extremely persuasive spreadsheet?”
Same rules. Work out whether it’s a model output, a convention, or a statute — because legal force, operational usefulness and statistical fit are three unrelated properties that arrive in identical fonts.
“What if we have no time to gather perfect data?”
Same rules. Start measuring the inputs that govern the immediate decision. A rule of thumb can guide action while the record grows. It can’t turn a small skewed sample into a general proof.
The details change.
The architecture doesn’t.
THE FIVE, WITHOUT THE SIRENS
Leave capacity below the point where ordinary variation turns into nonlinear waiting.
Measure work in process, arrival rate and elapsed time before adding work — and find the constraint.
Treat average volume, arrival variation and service-time variation as three separate inputs.
Hold safety stock against a named failure, and trigger replenishment at a defined reorder point.
Recalibrate every borrowed threshold against local assumptions, scope, and enough of the right data.
ACT NOW, BEFORE THE WAITING ROOM BECOMES A LANDMARK
Tonight, pick one queue that has become part of the furniture.
Write down what’s in it, how fast work arrives, and how long work spends inside. Then find out whether the system is stable, filling, or draining — because that determines whether any of those numbers mean anything yet.
Mark the resource or handoff that controls departures. That’s your candidate constraint, and it’s frequently not the department everybody complains about.
Now look at its schedule. If it’s aimed at 80–85%, ask what variation is supposed to fit in the remaining space. If it’s aimed higher, say out loud where the late work goes.
“Somewhere” is not a queueing model, although it has appeared in several brochures.
For one stocked item, name the specific failure you’re buffering: demand variation, lead-time variation, or both. Write the reorder point. If the buffer is “half, because that’s what we do,” write down that it’s the shortcut — then decide whether this item deserves the real calculation.
Finally, take one copied ratio off the wall and ask what kind of number it is. Model output, convention, statute, or scoped simulation result?
Then gather the local arrivals, shrinkage, abandonment or case mix it needs — and check whether your sample actually supports the conclusion, rather than asking n ≥ 30 to perform magic in a heavy-tailed hat.
For the low, low price of measuring before scheduling harder, the complete Five Number One system is yours.
No subscription. No dashboard wallpaper. No platinum certificate of 100% UTILISATION mailed in a commemorative tube.
And if you act now, we’ll include the thing every overloaded system needs, at no additional charge:
room to absorb reality.
Operators are no longer standing by.
The operator is whoever the queue reaches first.