When every buyer orders twice.

Picture two pieces of paper from a small retailer, stapled together. The first is the week's sales slip: 100 units of one product sold to customers. The second is the purchase order the retailer sent its supplier the same afternoon: 150 units. Between them, in pencil, someone has written two lines of arithmetic. Thirty more to rebuild the shelf after a busy fortnight. Twenty more because the supplier now quotes three weeks for delivery instead of two. Nobody lied. The sales slip is accurate, the order is accurate, and the pencil note explains the difference. But the supplier receives only the second piece of paper, and to the supplier, 150 looks like demand.

That gap is the subject of this essay. The sales slip records end-customer consumption: what people actually bought, at the boundary where goods leave the business. The purchase order records a replenishment order: a decision taken by a buyer in response to sales, stock on the shelf, delivery times and the incentives in front of them. Both are real records of real events. They answer different questions, and a supplier that reads the second as if it were the first is forecasting from a decision, not from demand.

My argument is that orders are actions, not observations, and that the rules people use to place them can turn a small change in consumption into a large swing in production further up the chain. When each participant treats the next participant's order as a direct reading of what customers want, the error compounds. Better forecasting helps, but only when the organization also understands the feedback rule that produces the order stream. A more accurate model of a distorted signal is still a model of the distortion.

Two records at two boundaries

Most of the trouble comes from treating the two records as interchangeable because they sit side by side in the same database. They do not measure the same thing. A point-of-sale record is an event: a customer took a product at a price at a time. A purchase order is a judgment: a buyer decided how much to ask for, given what was sold, what was left, what was already on its way and what the buyer expected to happen next.

For a single retailer looking at its own business, the difference is obvious. The trouble starts one step up. A distributor sees the retailer's orders, not its sales, and places its own orders with the manufacturer on that basis. The manufacturer sees the distributor's orders and plans production from them. Each step adds its own judgment to the one below it, and each judgment is invisible to the next step, which sees only a number.

A data architect can make every one of those numbers perfectly accurate. The order table can be complete, deduplicated and reconciled to the invoice. None of that tells the supplier how much of the 150 was consumption and how much was the buyer's response to its own circumstances. That information lives in the pencil note, and the pencil note is usually not stored anywhere.

How a small bump grows

Figure 2 is a constructed example, not observed sales from any company. A retailer sells 100 units a week and aims to keep 200 on the shelf. Its supplier delivers every order exactly two weeks after it is placed, with no limit on capacity. In week 1, sales rise once, to 120, and then return to 100 for good. The retailer starts with 200 on hand and two earlier orders of 100 each on the way.

The retailer could use one of two ordering rules. The first counts the pipeline: order this week's sales, plus enough to bring what is on hand and what is already on order up to 300. The second ignores the pipeline: order this week's sales, plus whatever the shelf is short of 200. The second rule sounds sensible. It is how a busy manager might reason when looking at a low shelf.

Follow the rule that ignores the pipeline. In week 1 the shelf ends at 180, so the retailer orders 140: the 120 it sold plus 20 to refill. In week 2 the shelf is still at 180, because nothing ordered after the bump has arrived yet. So the retailer orders another 20 on top of sales, 120 in all, for a shortfall it has already ordered once. In week 3 the first extra shipment lands and the shelf overshoots to 220, so the order drops to 80. In week 4 the second extra shipment lands, the shelf reaches 240, and the order drops to 60. A single 20-unit rise in consumption has produced orders that swing from 140 to 60, a range of 80, and a shelf 40 units heavier than the retailer wanted. Run the rule further and it keeps swinging.

Now switch to the rule that counts the pipeline. Week 1 is the same: 140. But in week 2 the retailer sees that the 140 is already on its way and orders only 80. By week 3 it is ordering 100 again, and by week 4 the shelf is back at 200 with orders at 100. The rule still overreacts for a week, because it replaces sales and refills at the same time, but the swing dies out instead of repeating. The only difference between the two retailers is whether the stock already ordered is visible to the person deciding.

The supplier, meanwhile, sees neither rule. It sees an order stream: 140, 120, 80, 60. If it plans production by forecasting from that stream, it is forecasting the retailer's ordering rule, and passing an amplified version of it to its own suppliers.

What the experiments showed

The mechanism is old and well studied. In a 1989 paper in Management Science, John Sterman described an experiment in which people managed a simulated inventory distribution system with multiple actors, feedbacks, nonlinearities and time delays. The decisions, taken together, produced aggregate behavior that diverged systematically from the best available. Sterman modeled the subjects' choices with a simple anchoring and adjustment rule for managing a stock, found that it explained their behavior well, and traced the poor results to what he called misperceptions of feedback: the subjects did not take enough account of how their own decisions fed back into the system they were managing.¹

Peter Senge brought the same kind of simulation to a management audience in The Fifth Discipline, where he uses it to make a point about blame. When the shelves empty and then overflow, each participant tends to blame the one next to them: the customers were fickle, the supplier was slow. Senge's argument is that the structure of the system, its delays and its decision rules, produces the behavior, and that people inside it cannot see the structure from where they stand.³ That is a useful habit of mind for an operating leader. It is not, by itself, a measurement. A classroom simulation shows that a mechanism exists. It does not say how large the effect is in any particular company.

Incentives bend the signal too

Delays and ordering rules are not the only cause. In 1997, Hau Lee, V. Padmanabhan and Seungjin Whang studied what they called the bullwhip effect: the information carried by orders tends to be distorted, the variance of orders may be larger than that of sales, and the distortion tends to grow as it moves upstream.² Their companion article for managers opens with Procter & Gamble's experience with Pampers diapers. Retail sales of the product were not especially variable, yet the orders P&G received from distributors swung more, and P&G's own orders to its suppliers, such as 3M, swung more still.⁴

The authors named four sources of the distortion. The first is how each stage turns the orders it receives into a forecast and adds safety stock, which is the mechanism in Figure 2. The second is order batching: buyers who order once a month, or by the truckload, send a lumpy signal even when consumption is smooth. The third is price: promotions and discounts lead buyers to stock up when the price is low and stop ordering when it is not. The fourth is rationing. When a supplier allocates scarce product in proportion to orders, buyers learn to ask for more than they need, and the order book fills with requests that will be cancelled once supply returns.²

That fourth source is where the title of this essay comes from. Under an allocation rule that rewards large requests, every buyer has a reason to order twice. The rule makes the shortage look larger than it is, and the supplier, reading the order book as demand, may add capacity for customers who do not exist. The remedies the authors proposed follow from the causes: make downstream sales data available upstream, reduce the batching that makes orders lumpy, stabilize wholesale prices, and allocate scarce product based on past sales rather than current orders.⁴ These are mechanisms with evidence behind them. They do not say that every industry shows the same magnitude, and I do not claim it.

Some lumpiness is rational

The strongest critique of the argument so far is that a smooth order line is not the goal. A buyer facing a large fixed cost for each delivery should order in batches. A buyer facing uncertain supply should hold more stock, and may reasonably order more when lead times lengthen, as the pencil note in Figure 1 did. Seasonal products should be ordered ahead of the season, and service commitments may justify a buffer that looks extravagant in an ordinary week.

None of that is distortion. It is the order stream doing its job. The point is not to make orders look like sales. It is to make sure that the people reading the orders know which part is consumption and which part is a decision, so that they respond to the right one. The objective is economically sensible inventory and reliable service, and sometimes that requires exactly the lumpy order a naive analyst would call noise.

Keep the pencil note

The practical remedy is to store the pencil note. For each order, keep a decision record alongside it: the sales in the period, the stock on hand, the stock already in transit, the lead time the buyer expected, and the reason for any adjustment above or below sales. With that record, an analyst can separate consumption from replenishment, see when a buyer has ordered again for a gap already filled, and notice when lead-time assumptions have changed across many buyers at once.

For a supplier, the most valuable version is the customer's own sales data, shared upstream, so that production can be planned from consumption rather than from orders alone. That is the first remedy Lee and his coauthors proposed, and it remains the most direct.⁴ But information alone cannot fix an incentive. If the allocation rule still rewards inflated requests, buyers will still inflate them, and the shared data will show consumption while the order book shows something else. The rule that turns requests into allocations has to change too.

Inside a single company the same discipline applies to internal orders: the requests that regional teams send to a central warehouse, or that business units send to a shared service for capacity. Each is a decision by someone with their own shelf to protect. A forecast built from those requests inherits every one of their pencil notes, unread.

Objections, considered

The real problem is unreliable suppliers. Sometimes it is. Long or uncertain delays can start the process and make it worse, and the remedy for a slow supplier may be a faster one. But the argument here is about how the ordering rule responds to a delay, not a denial that delays and physical shortages are real. In Figure 2 the supplier is perfectly reliable, and the rule that ignores the pipeline still swings.

Customers will not share their sales data. Many will not. Then use what can be observed or agreed: contractual commitments, actual fill rates, cancellation rates after a shortage, and a bounded range of scenarios for consumption. What a supplier should not do is pretend to a visibility it does not have, by treating its order book as if it were the customers' sales.

More inventory is sometimes the right answer. Yes. A buffer bought deliberately, sized to a known risk, is good management. The problem is a different buffer: one ordered again and again because the stock already on its way was out of view. The first is a decision. The second is the system ordering twice.

The order book records what buyers decided, not just what their customers consumed.

Before you plan from the order book

• Does the forecast ingest sales or orders? If orders, say whose decision rule it is forecasting.
• Is pipeline stock visible to the person ordering? Check that what is already on its way appears where the decision is made.
• Is the reason for an adjustment recorded? Rebuild, longer lead time, promotion or expected shortage should each be a field, not a memory.
• Does an allocation rule reward inflated requests? If scarce supply follows the size of the order, expect every buyer to order twice.
• Which swings are deliberate? Separate batches, seasonal buys and chosen buffers from orders that repeat a gap already filled.

Back to the two papers stapled together. The sales slip still says 100 and the purchase order still says 150, and both are still correct. The difference between them is no longer a mystery: thirty to rebuild a shelf, twenty for a longer wait. The supplier who receives that order with its pencil note attached can plan for a customer base that bought 100. The supplier who receives the order alone will plan for 150, and next week, when the shipments land and the order falls, will wonder where the demand went.