The Decision Factory

The machine on one page

Papers 01 to 03 argued for three components: a priced objective, a ledger that prices both directions of error, and forecasts that carry their uncertainty instead of discarding it. This paper shows where those components sit, what each one does, and, as importantly, what each one is not allowed to do.

The machine is called a factory for a reason. A factory has stations. Each station takes a defined input, does one job, and hands a defined output to the next. Nobody at station three reaches back to rework station one. When something goes wrong, you can find out where.

Unattended does not mean unsupervised. It means nobody assembles the inputs by hand, nobody reconciles two copies of the truth, and nobody re-keys a number from one screen into another. The stages run on a clock. People watch them for drift and change them deliberately. What disappears is the monthly ritual of building the number.

The example we follow through every stage

One product: Solstice SPF 50, a 200 ml sunscreen, sold by the case of twelve to retailers in one region. The season is the sixteen weeks from Memorial Day to Labor Day. The plant commits the season’s production in April, so there is one large decision and very little room to correct it later. Everyone has bought sunscreen. Almost nobody has thought about how hard it is to sell the right amount of it.

The ledger is the one from Paper 02, carried forward unchanged: $10 for each case carried that was not needed (capital, storage, and the markdown or year of shelf time an unsold case of sunscreen faces) and $85 for each case of demand not served ($40 of forgone contribution plus $45 of customer damage). The critical ratio is 85 ÷ 95, or 89.5 percent, exactly as in Paper 03.

Marketing has $60,000 and two campaigns to choose from. That choice, and how much sunscreen to make alongside it, is the decision this paper follows.

Stage 1: One data platform

Every decision system starts as a data system, and most of them die as one. The first stage has a single job: publish one governed version of what happened, at one grain, that every later stage reads and none of them edits.

Takes in

Raw feeds from the systems that already exist.

Hands on

One table, with flags, that the models read and nobody overwrites.

How it changes as data arrives

The rule that matters most is the dullest one: stockout weeks are flagged, never deleted and never left alone. Last summer, Solstice ran out of stock during the heat wave in weeks seven and eight.

Hand the model the sales column and it learns that the hottest fortnight of the year looks like an ordinary one. It will then size next summer’s heat-wave response 960 cases too small, and do it confidently. Delete the two weeks and the model forgets heat waves happen. Flag them, and the model knows demand was at least what sold, which is the truth.

Stage 2: Learned models

The second stage holds three models. Each answers one question, and the discipline is keeping each one to its own question.

1. The demand model: what will customers want, given the things we do not control?
2. The response models: what do the things we do control change about that?
3. The economic ledger: what does it cost when we are wrong, in each direction?

The line between the first two is the one most organizations blur, and it follows directly from the definition that runs through this series. A decision is anything you control. Weather is not a decision. A campaign is.

2a. The demand model

As the summer reveals itself, the model updates weekly. By July the weeks already played are no longer uncertain, and the band over the weeks remaining narrows. Nobody adjusts this by hand. Weather is data, not an opinion.

2b. The response models

Marketing has two options for the same $60,000.

Campaign B does not start from nothing. A new campaign borrows its estimate from its nearest relatives, and it borrows their spread too. That is why its band is so wide that it includes zero: the honest answer, before it has run, is that it might do nothing. Each season it runs, its own evidence takes over from the borrowed evidence and the band closes in. Move the slider to watch it.

Two things happen as the slider moves. The band narrows quickly at first and slowly after: the first season of real evidence is worth more than the next four together. And the ledger cost of width falls with it, from $40,549 to under $27,000. Nothing about the campaign changed. Only what the organization knows about it.

2c. The economic ledger

The ledger is the slowest-moving component in the machine, and it should be. It changes when the business changes, through a signed revision, not when a forecast misses.

When it does change, everything downstream moves with it. Suppose finance reviews last season’s write-offs and finds that an unsold case actually costs $14, not $10, because markdowns ran deeper than assumed. The critical ratio drops from 89.5 to 85.9 percent, and the engine’s commitment for Campaign A falls from 13,750 cases to 13,475. Not one forecast changed. Two hundred seventy-five cases moved because the price of being wrong moved, which is exactly how it should work.

Stage 3: The decision engine

The engine is where the three models meet a decision. It generates candidates, runs each one against thousands of possible seasons drawn from the models, prices every season with the ledger, applies the constraints, and ranks the candidates in dollars.

Here is a slice of what it produces for Solstice. Value is the season’s contribution, less the ledger cost of error and the $60,000 of campaign spend.

Read the ranking against the way the question gets settled today. Campaign B promises 200 more cases than A. At $40 a case, a point-estimate meeting picks B and credits it with $8,000 more. The engine picks A, by $5,715. The swing between the two answers is $13,715, and all of it is the cost of B’s width: it needs 1,175 more cases of production to protect the same customers, and it still runs short more often.

The meeting asks which campaign sells more. The engine asks which campaign costs less to be wrong about.

That answer is not permanent. It depends entirely on how wide B’s band is, and the band is something the organization can change.

Stage 4: Human judgment

This is the one stage with a person in it, and it is the reason the other four exist. The machine is built so that a decision owner spends their time on the decisions where their knowledge changes the answer, not on assembling the number they will be asked to defend.

The system asks for three fields, not a justification. Here is the first override of the Solstice season.

The person knows something the model does not, and the record says what. If the stores open as planned, the override is worth $2,273 against the engine’s recommendation. If they do not, it costs money, and the record will show that too. Either way, it is evidence. The value of a recorded override is not that it was right this time. It is that the next hundred of them can be scored.

Stage 5: The learning loop

The last stage closes the line. Actuals and overrides return, and each component updates on its own clock. A factory that does not do this is not a factory. It is a forecast with extra steps, and it starts decaying the day it ships.

The last row is the one most systems skip. The 30 stores opened, and demand ran about 540 cases above what the model expected. The override was right. Across the catalog, “distribution change” overrides were right seven times in nine over two seasons. A reason code that keeps being right is a data feed nobody has built yet. So the store-opening calendar becomes an input to the demand model, and next year nobody has to remember to override it.

An override that keeps being right is not a judgment call. It is a missing input.

The Lab, beside the line

Every organization runs what-ifs. Most run them in a spreadsheet that someone built for the occasion, with its own demand assumption, its own cost assumption, and no connection to the numbers the business will actually commit to. The answer is only as good as whoever built it last.

The Lab is the same idea, with one rule that changes everything: it uses the same data, the same models and the same ledger as the line. A what-if in the Lab and a recommendation on the line disagree only because the question differs, never because the assumptions do.

The commercial director asks the obvious question about Campaign B: what happens if we commit to it and the summer is cool? The Lab runs both campaigns, at their engine commitments, across every summer the demand model can produce, and then across three weather scenarios.

The director asked about the weather. The Lab’s answer is that the weather is the wrong worry. A beats B in every kind of summer, by between $2,000 and $11,000 on average, and the gap is the campaign’s own uncertainty, not the climate. B’s 1-in-10 bad season is roughly $52,000 worse than A’s. Its 1-in-10 good season is roughly $36,000 better.

So the useful question is not whether to fear a cool summer. It is whether B’s width is worth paying to reduce. That is a question only the Lab can price, and it is the subject of the next section.

What a new campaign is worth

Campaign B costs $5,715 this season compared with A. That is the price of running it. What it buys is a season of B’s own evidence, which narrows its uncertainty from ±1,800 cases to ±805, measured as one standard deviation. Exhibit 5 already showed that below 1,342, the ranking flips.

But the band does not just narrow. It moves, to wherever B’s real effect turns out to be. Nobody knows that in advance, so the Lab prices the full range of what the season could reveal.

The asymmetry is the whole argument. If B turns out weak, next season goes back to A and loses nothing. If B turns out strong, it wins by tens of thousands of dollars a season, every season. The downside is capped at the price of the test. The upside is not. That is what an option looks like, and organizations that only ever run the campaign with the best point estimate never buy one.

This changes the decision, and the change is recorded the same way as any other. The decision owner overrides the engine’s ranking with a reason code of Deliberate test, a note pointing to the Lab’s run, and the new value: Campaign B. Buying information is a legitimate reason to depart from the ranking. It just has to say so.

One season through the machine

Here is the season as the machine saw it, at three moments.

Count the meetings in that table. There are none where anyone builds a number. There is a director’s question in week zero, answered with a price. There are two override records, each three fields long. Everything else ran on a clock.

What to ask on Monday

Most organizations already own pieces of this machine. The question is which stations are missing and which ones are doing another station’s job. Five questions are usually enough to find out.

1. What happens to stockout weeks in our history? If they are deleted, or left in as if they were demand, every model downstream is learning the shelf instead of the customer.
2. Is marketing’s effect a model with a band, or an adjustment someone makes? If it is an adjustment, the organization pays for every campaign twice: once to run it, and again to forget what it did.
3. Where does the price of being wrong live, and who signed it? If the answer is a service-level target, the engine has nothing to rank against.
4. When someone overrides the recommendation, what gets written down? If the answer is nothing, the organization’s best knowledge leaves the building every time someone changes jobs.
5. Do our what-ifs run on the same assumptions as our commitments? If each analysis builds its own, the organization does not have a Lab. It has a collection of opinions with formulas in them.

None of these require a platform purchase, and the machine runs on the data and tools most organizations already have. What it does require is people in roles that rarely exist today, which is where the next paper picks up.