Operations Management: Where Work Waits

You join the queue at a pharmacy. The pharmacist never stops moving, yet your three-minute prescription takes forty minutes. Nobody is lazy, nobody is slow, and the shop is not even especially busy by the end of the day. Still you waited.

A business graduate who took operations management looks at that room and sees something else: a process, with work flowing in, work waiting, work being done and work flowing out. The waiting is not a mood. It follows laws that can be written down, and the same laws govern a hospital emergency room, a car factory, a bakery and a website. This piece is about the handful of those laws that every business school teaches, and why they surprise people.

A process is a chain, and the slowest link sets the pace

Draw any operation as boxes in a row: take the order, cook, pack, deliver. Work enters on the left and leaves on the right. The rate at which finished work leaves is the throughput. The box that can do the least per hour is the bottleneck, and it caps throughput for the whole line, whatever the other boxes can do.

That is why a graduate asks "where is the bottleneck?" before "how do we speed up?". Speeding up anything else buys nothing except a bigger pile of unfinished work in front of the slow step. The idea was made famous as a novel by Eliyahu Goldratt in The Goal (1984) and as the Theory of Constraints. Try it: four stations in series, and a pile that grows wherever one station feeds a slower one.

With the starting settings the line makes 40 per hour. Doubling Cut changes nothing but adds work to the pile in front of Weld. Lifting Weld to 50 lifts output only to 45, because Pack is next in line. Every improvement moves the bottleneck, which is why the hunt never finishes.

Little's law: stuff in the system equals flow times time

Count the people inside a café. Now count how many come in per hour, and how long each stays. A tidy law connects the three. Little's law says that the average number in a system equals the average arrival rate times the average time each spends there. John Little proved it in 1961; Philip Morse had published the form a few years earlier and challenged readers to find a case where it fails.

Its power is that it holds without knowing anything about the shape of the randomness, the order of service or the layout. A café with 30 arrivals an hour and a 15-minute stay holds 30 × 0.25 = 7.5 people on average. A hospital with 40 admissions a day and a 5-day average stay holds 200 patients. A software team with 10 tasks finishing per week and 30 in progress has an average cycle time of 3 weeks, and the quickest way to cut it is to start fewer things. Play with it, and watch a random simulation agree.

Little's law is accounting, not physics. It cannot make the line shorter by itself; it tells you which two levers to pull. To shorten the time people spend in a system you must either move work through faster or hold less of it at once.

Queues: why a busy server means a long wait

Here is the result that surprises almost everyone. If customers arrive at random and a server takes a random time, the average wait does not grow in step with how busy the server is. It rises slowly and then explodes. For the simplest model, one server with random arrivals and random service times (called M/M/1), the average wait before service is the service time multiplied by the utilisation divided by one minus the utilisation.

Take a clerk who needs 5 minutes per customer. At 50% busy the wait is 5 minutes. At 80%, 20 minutes. At 90%, 45. At 95%, 95 minutes. At 99%, nearly 8 hours. Going from 90% to 95% busy, an apparently small change, doubles the wait. A manager who "uses the staff fully" has bought a queue.

The second half of the story is variability. The wait comes from irregularity, not from the average load. Kingman's approximation, from the 1960s, splits it into three factors: how busy (ρ/(1−ρ)), how variable (the average of the squared variability of arrivals and of service), and the service time. If arrivals come on a perfect schedule and service takes exactly the same time every time, the wait is zero even at 99%. A dentist who books appointments is using the first factor to tame the second. Slide to see it, then simulate.

The practical rules a graduate carries: never plan a shared resource at 100%; keep slack where randomness is high (emergency rooms, call centres); and attack variability itself, with appointments, standard work and smaller batches, because that lowers every queue at once. The same mathematics sets the buffer in a network router.

Inventory: how much to order and how much spare to hold

Stock is a bet in two directions. Hold too little and you run out; hold too much and cash sits on a shelf. Two textbook ideas cover most of it. The first answers "how big should each order be?". Every order has a fixed cost (paperwork, shipping) and every unit in stock has a holding cost (space, capital, spoilage). Big orders mean fewer orders but more holding; small orders the opposite. The balance is the economic order quantity, written down by Ford W. Harris in 1913 and popularised by R. H. Wilson in the 1930s.

A shop that sells 12,000 units a year, pays $50 per order and $2 per unit per year to hold stock should order √(2 × 12,000 × 50 ÷ 2) ≈ 775 at a time, for about $1,549 a year in total. Ordering 20% less costs only about 2.5% more: the bottom of the curve is flat, so a rough answer is nearly as good as an exact one. The real lesson is the square root: quadruple the demand and the best order only doubles.

The second idea is safety stock, the spare units that protect you against demand being higher than forecast while you wait for a delivery. If weekly demand wobbles by a standard deviation σ and the lead time is L weeks, the wobble over the lead time is σ√L, and the cushion is a multiplier z (from the normal curve) times that. 95% protection needs z ≈ 1.65, 99% needs 2.33, 99.9% needs 3.09: every extra nine costs more than the last.

The bullwhip effect: small wobbles, huge swings

Jay Forrester of MIT described in Industrial Dynamics (1961) how orders swing more the further you go from the customer. Hau Lee, V. Padmanabhan and Seungjin Whang gave it a name and four causes in 1997: each company forecasts from the orders it sees, rations when short, orders in batches, and reacts to price promotions. Each step distorts a little and the distortions multiply.

You can feel it in the beer game, developed at MIT's Sloan School in the 1960s: four players run a retailer, a wholesaler, a distributor and a brewery. Customers want 4 cases a week, then suddenly 8. Nothing else changes, yet the players typically order wildly, run out, panic, over-order, and end with mountains of beer. John Sterman's 1989 paper in Management Science showed that the cause is human: players underweight the orders already in the pipeline. Below is a computer version with a sensible ordering rule, so any swing you see comes from structure, not from bad players.

Defaults give a doubling of real demand and a factory peak of about seven times the new weekly level. Sharing real demand with everyone cuts the amplification sharply, but not to zero: even a perfect forecast must build stock once for a bigger business. The cures that firms use are exactly the levers here: share sales data, shorten delays, order smaller and more often, and avoid promotions that make buying lumpy.

The 2021 chip shortage is the real-world case that made supply chains a dinner-table topic. In spring 2020 carmakers cut chip orders as sales collapsed; chip makers shifted capacity to electronics, whose demand rose; then car demand came back faster than expected. A string of accidents made it worse: a February 2021 winter storm shut Samsung, Infineon and NXP plants in Austin, Texas (The Verge, 17 February 2021), and a March 2021 fire at a Renesas plant in Japan threatened a supplier of about 30% of the world's car microcontrollers (Reuters, 31 March 2021). A consulting firm's estimate, reported by CNBC on 23 September 2021, put the hit to 2021 automotive revenue at $210 billion, and the shortage had mostly subsided by 2023. Analysts still argue about the lesson: critics blame lean, low-stock supply chains; defenders note that the chip makers' capacity, not carmakers' habits, was the limit, and that few firms could see who made the chips inside their suppliers' parts. Both are partly right: low inventory buys efficiency and sells resilience.

Lean and the Toyota Production System: less waiting, not more rushing

Taiichi Ohno and Eiji Toyoda built the Toyota Production System between 1948 and 1975. Its two pillars are just-in-time (make only what is needed, when it is needed, in the amount needed) and jidoka, "automation with a human touch": any worker or machine can stop the line when something is wrong, signalled by an andon board. Kanban cards pull work from upstream only when downstream has used up its stock. The word lean became popular after Womack, Jones and Roos's 1991 book The Machine That Changed the World. Toyota names waste (muda) in several kinds, with overproduction the worst, because it hides every other problem under a pile of stock.

Little's law explains why this works. If you cap the number of jobs allowed on the line, as a kanban system does, then cycle time is jobs divided by output, and output barely falls when the cap is cut, until the cap gets very small. For a balanced line of four stations with random work times, output with w jobs allowed on the line is the single-station rate times w ÷ (w + 3). Find the knee.

Going from 40 jobs on the line to 12 gives up under a quarter of the output (55.8/h down to 48/h), yet cuts the time any job spends from 43 minutes to 15. The lean slogan "less work in progress" is a free lunch up to a point, and the model shows where that point is. Lean also comes with critiques: it can make a chain fragile to shocks, as above, and in services it can turn into pressure on workers. Toyota's own version rests on stopping the line and fixing the cause, not on squeezing people.

Quality and Six Sigma in plain words

Every process produces results that vary. Quality means keeping the variation inside what the customer will accept: the specification limits. The question is how far, measured in standard deviations (σ), the limit is from where the process normally sits. A distant limit means few defects. Six Sigma began at Motorola in 1986 with engineer Bill Smith and became famous when Jack Welch made it central to General Electric in 1995. Its management routine is DMAIC: define, measure, analyse, improve, control.

The famous number comes from the normal curve. A "six sigma" process, with the limit six standard deviations away, would produce about 2 defects per billion. Yet the usual quoted figure is 3.4 defects per million, because the convention assumes the process drifts 1.5σ toward a limit over time, which leaves a 4.5σ gap. The 1.5σ shift is itself criticised by statisticians as arbitrary (Donald Wheeler called it "goofy"). Reports of success are mostly company-reported, and one 2010 report put the share of projects that fail above 60%; the tools are sound, the hype is not.

Set the drift to 1.5 and the sigma level to 6 to get 3.4. Then notice what the readout says about chains: a 99% step sounds excellent, but twenty of them in a row give only 82% of items right the first time. Real quality problems hide in the product of many small ones, which is why the graduate counts defects per step, not per product.

What a graduate just sees

• Every system has a bottleneck. Improving anything else is an expensive way to build a pile.
• Inventory = flow × time. To cut the time things take, hold less at once.
• A resource at 95% busy is a queue machine. Slack is not waste; it is how you buy speed.
• Variability costs more than the average does. Anything that evens out arrivals or service shortens every queue.
• The best order size is a square root, and the cost curve is flat near the best, so a rough number is nearly as good.
• Information distorts as it moves upstream. Passing real demand along beats forecasting from orders.
• Low stock buys efficiency and sells resilience. Choose it on purpose.
• A defect rate looks small until it is multiplied across all the steps.

Where it shows up next

• Construction Management (34/70): a building is a process with a critical path and a resource peak; the idea of one limiting link is the same.
• Strategy & Competition (34/54) and Marketing Management (34/57): service levels, delivery promises and availability are operations choices that customers feel.
• Operating Systems (34/26) and Computer Networks (34/27): schedulers and routers are queues, and the 95%-busy explosion is why a loaded server or link slows suddenly.
• Software Engineering (34/31): kanban boards and work-in-progress limits are Little's law applied to code.
• In the world: hospitals, airlines, warehouses, restaurants, call centres, and every company that has ever said "we ran out".

If you want to go deeper

• 📘 Textbook. Gérard Cachon and Christian Terwiesch, Matching Supply with Demand: An Introduction to Operations Management (McGraw Hill). The standard business-school text, strong on process flow, queues and inventory.
• 📘 Textbook. Wallace Hopp and Mark Spearman, Factory Physics (Waveland Press). The rigorous source for Little's law, variability and WIP caps, with the maths shown.
• 📘 Gentler. Eliyahu Goldratt, The Goal (1984). A novel about a plant manager who learns bottlenecks the hard way. No maths.
• 🌐 Free course. MIT OpenCourseWare 15.763J Manufacturing System and Supply Chain Design (Spring 2005, Simchi-Levi and Graves): lecture notes on capacity, inventory and supply chains.