The Last Honest Order

IThe whip

In the late 1950s Jay Forrester, an engineer who had helped build some of the first digital computers at MIT, took up a puzzle brought to him by managers at a household appliance maker. Their plants hired and laid off in waves, while the stores they supplied sold at a fairly steady pace. Forrester’s answer, set out in Industrial Dynamics in 1961, was that nobody in the chain was doing anything wrong. Each link ordered from the next one up on its own reading of what it had just been asked for, plus a margin of safety, and then waited. A small rise at the till looked like a trend to the distributor. To the factory it looked like a boom.

His group turned the idea into a board game, the beer game, still played in business schools: four players run a retailer, a wholesaler, a distributor and a brewery. Customer demand steps up once and then holds. Table after table produces wild swings, and the brewery ends up buried in beer. When John Sterman studied the players in 1989, most of them blamed the customers. Then, in the 1990s, a large diaper maker noticed that its orders swung hard even though babies use diapers at about the same rate every week. Hau Lee and his colleagues at Stanford named the pattern the bullwhip effect in 1997 and gave it four causes: forecasting from orders, ordering in batches, price promotions, and the rationing game, where buyers facing a shortage ask for more than they need so that their share comes out bigger.

Here is the principle in one line. The farther a link stands from the use, the more it sees orders instead of needs, and an order is a need with a guess attached.

An order is a need with a guess attached. The guess is the part that climbs.

Figure 01

The same small wobble, one link at a time

IIThe one link that did not guess

The part people miss is that the whole chain had one anchor, and it was you. When you bought bread on a normal Saturday you were not forecasting bread. You were mostly buying what your house would eat that week. The basket was the closest thing the chain had to the use itself. That is why the cure the trade settled on, when it worked, was to send the till data straight up the chain, so the factory could see what people bought rather than what the distributor asked for. Shared sales data, suppliers managing the store’s stock, forecasts built together. Each of these is a version of one sentence: pass up the need, not the order.

The March 2020 aisle is what happens when the anchor slips. For a few weeks shoppers stopped buying what they used and started buying what they feared. The basket became an order with a guess in it, and the whip gained a new first link. It ran for months. But households guess slowly: one trip, one cart, one week at a time.

IIIWhat the delegate changes

AI delegates are taking over the basket. Not at once and not for everything, but the direction is plain. The pantry reorders itself. The delegate notices the soap is low. The delegate reads a headline about a port strike and tops up. The household order stops being a record of use and becomes a forecast, made by a machine that is very good at forecasting.

That changes two inputs of Forrester’s law at once, and they push in opposite directions. The delay falls: a delegate sees consumption the hour it happens, and shorter delays damp the whip. But the anchor goes: if the delegate sends an order, the bottom of the chain now guesses like each link above it, and it guesses fast, in step with millions of other delegates reading the same signal. The rationing game that took a frightened person a week takes a program a second. The 2020 aisle could be emptied before anyone has left the house.

So the consequence is not that the whip gets worse, or better. It turns on one design choice inside your delegate: does it pass up an order, or does it pass up the use? A delegate that says “send twelve rolls” is one more guessing link. A delegate that says “this house goes through a roll every two days and has eight on hand” gives the chain the thing it has been missing since Forrester drew it: the need itself, clean, from millions of homes. The same machine, with opposite effects on the shelf.

IVWhat I expect

1. Stores will throttle machine orders before human ones

   When a shortage bites, the first rules will land on orders placed by delegates: a ceiling per household, a delay, a queue. The rationing game runs fastest where the buyer is a program, so that is where sellers will put the brake.

2. Short stock will be shared out by use, not by request

   In a shortage, sellers will allocate by what each household has actually used. Showing your consumption becomes the price of priority, and the delegates that can show it get served first.

3. The use becomes worth more than the order

   Makers and grocers will pay, in price or in priority, for honest consumption data from delegates, because it is the one input that kills the whip. The argument over who holds that data will be louder than the argument over who places the order.

4. The first run on a store made by machines

   Within a few years a local shortage will be traced to delegates reacting to one signal at the same moment. It will be reported as a glitch. It will be the bullwhip with its first link moved into software.

5. Calm becomes a feature you can buy

   Delegates will be sold on not panicking: holding a steady stock and ignoring headlines that do not touch your use. The best one, in a scare, will be the one that orders least.

VWhere I have stood in this

I spent part of my career on this exact problem. E2open, the company I co-founded, built software for supply chains, a business built on the fact that a demand signal distorts as it climbs a chain of decisions. The lesson the trade learned slowly was that a cleverer forecast at each link helps less than letting each link see the use at the bottom. So when I think about my own delegate, the question I care about is not how clever its orders are. It is whether it reports what my house uses, or adds one more guess to the stack. I wrote about the older cousin of this law, the cobweb that swings markets when supply answers an old price, in Planted at Yesterday’s Price. That piece was about the lag. This one is about the anchor.

VIThe claim ledger

Already true

The mechanism is record, not argument: Forrester’s chains, the beer game’s tables, the diaper orders that swung while the babies did not, the four causes named in 1997. So is the cure the trade reached for, which was to share the till and not just the order. And March 2020 showed what happens when households stop reporting use and start guessing.

What has to happen

Delegates have to place a meaningful share of household orders, and the people who build them have to choose between sending orders and sending use. Sellers have to be willing to take use data and act on it. If delegates only ever carry out what a person typed, the anchor holds and little here changes.

Where I am probably wrong

Delegates may be built so cautious, and stores may cap them so quickly, that machine panics get no room to run, and my fourth prediction never happens. People may refuse to share use data at any price, which leaves delegates sending orders and the whip running at the speed of software. If, five years from now, shortages look just like 2020 and nobody can tell whether the buyer was a person or a program, I had this wrong.

VIIThe paper aisle

It is a Saturday morning again, and the paper aisle is full. Not because people got calmer. Because what your house sends up the chain is no longer a guess. For a hundred years the shopper’s basket was the last honest order. The next one will be honest only if your AI tells the truth about what you use.

The whip starts with the first guess. Make sure it is not yours.