Who Does What Now

The question that gets answered too late

Programs of this kind rarely fail on the mathematics. They fail on the organization, and they fail in a specific and repeatable way.

A decision factory does not slot into the existing structure. It changes what a substantial part of the planning organization is for. When that change is left implicit, which it usually is until well after the first deployment, the old work quietly reconstitutes itself alongside the new system. Both then run in parallel, at greater total cost than before, until somebody senior concludes that the investment underdelivered.

If a planner is still judged on producing a defensible number by hand, they will keep the spreadsheet, and the spreadsheet will win.

It will win because it is the artifact their performance is measured against, because it is the artifact they can defend in a meeting, and because they are behaving entirely rationally given the incentives in front of them. The old job has to be genuinely abolished and the new one genuinely defined, staffed and measured. This paper sets out what those roles are.

From debating the number to arguing the economics

Watch a monthly planning meeting in most organizations and you will see a negotiation over a single figure. Sales says fourteen thousand, operations says eleven, finance wants it lower still, and the plan settles on whatever the most senior person in the room can live with. Nobody prices what it costs to be wrong in either direction, so nobody can say whether the compromise was a good one, and next month the same argument starts again.

A decision factory ends that meeting, not by silencing the people in it but by removing the thing they were arguing about. The forecast arrives as a distribution, produced without anyone touching it, and a distribution cannot be negotiated up or down. What is left to argue about is the part a machine genuinely cannot supply: what a lost sale is worth here, how much downside the business is willing to carry, and what is about to happen that the history does not contain.

Nobody wins an argument by asserting a number any more. They win it by being right about the economics.

That difference is the whole point. An argument about economics can be settled by evidence, so the person who was right gains standing and the organization learns. An argument about a point forecast is settled by seniority, so nobody learns anything at all. Moving the conversation from the first kind to the second is what returns planners to the work they were hired for, which is judgment.

What there is left to argue about

“Arguing the economics” is not a slogan. In practice it means the room disagrees about four things, each of which a person owns and the evidence can eventually settle.

1. The price of each error. What a lost sale costs against what a surplus unit costs, in this category, at this point in the season. Sales will argue a stockout loses the customer as well as the margin; finance will argue it loses the margin only. Repeat-purchase data on customers who met an empty shelf will say who is right.
2. The appetite for risk. Which tail of the distribution the business is willing to carry. A launch the company is betting its year on and a mature line it is harvesting deserve different answers, and that is a leadership choice, not a statistical one.
3. What the history does not contain. A competitor leaving the market, a supplier fire, a regulation taking effect next quarter. The model cannot see what has not yet happened; a planner who has heard about it can, and should say so in a form the system can use.
4. The constraint that binds. Which capacity, policy or approval step is actually capping the result this cycle, and what relaxing it would be worth.

One disagreement, worked through, shows the difference. A planning team must commit launch inventory for a new product. The demand model expects twelve thousand units, with a one-in-ten chance of sixteen thousand or more. A unit left over costs twelve dollars to mark down. The only real disagreement is what a lost sale costs.

In the old meeting this would have been a fight about whether the launch will sell thirteen thousand units or fifteen, which no one in the room could know. Here it is a fight about fifty dollars of value per lost sale, which the business can measure, and once it is measured the order follows mechanically. Nobody had to win the forecast. Somebody had to be right about the customer, and next season everyone will know who was.

The machine draws the curve. People choose the point.

A great deal of what planning teams do by hand is now routine work for a machine. Estimating demand is a statistical task, and it belongs to the data rather than to a meeting. Computing the orders that best serve a stated objective is an optimization task, which a machine performs in seconds across thousands of products. Neither should consume a planner’s month. What no machine can do is decide what the objective should be when two good things pull against each other.

That is why the most useful thing an optimizer gives a leadership team is not a single answer but a frontier: the set of plans for which no alternative is better on every measure that matters. Exhibit 2 shows one for a simulated portfolio of forty products, trading expected profit against profit in a bad season.

Three things in the exhibit deserve attention. First, today’s plan sits well inside the curve. A rule as innocent as “forecast plus five percent everywhere” leaves $5.3 million of bad-season protection on the table for no gain in expected profit, and plan B beats it on both measures at once. No amount of meeting time would find this, because the meeting cannot see the curve.

Second, moving along the curve is a genuine trade, and the curve states its price. Going from A to B gives up $0.4 million of expected profit for $1.4 million of protection in a bad season. Going on from B to C gives up a further $0.9 million for only $0.6 million more. The returns to caution diminish, and now everyone can see by how much.

Third, which point is right is not a mathematical question. A company betting its year on a launch, with a balance sheet that can absorb a poor season, might rationally choose A. One coming off a weak year with lender covenants to protect might rationally choose C. That is a strategic choice informed by knowledge of the business, its customers and its position, and it belongs to leadership. It should also be written down, because it is the risk appetite everything else is then measured against.

Accountability does not go away. It gets sharper.

The most common objection to all of this is that once a machine produces the plan, nobody answers for it. The objection assumes today’s process is accountable, and mostly it is not. A consensus number has many authors and no owner. When it misses, the miss belongs to everyone, which is to say to nobody, and the review that follows debates the forecast rather than the decision.

In a decision factory, every number the engine relies on was put there by a named person, and every one of them is checked against what actually happened.

1. Finance answers for the prices. If the ledger says a stockout costs forty dollars and the evidence says ninety, that is a finding against the ledger, and its owner recalibrates it.
2. Leadership answers for the risk appetite. The point chosen on the frontier was a written choice. A bad outcome inside the stated appetite is a cost the business agreed to bear; one outside it is a governance failure with a name attached.
3. Planners answer for their overrides. Each override is recorded with its reason and scored against the outcome. Within a year the record shows whose judgment adds value and whose subtracts it, which is more than most organizations know about their planners today.
4. The model steward answers for calibration. If the ninety-percent ranges contain the outcome only seventy percent of the time, that is measurable, and it is theirs to fix.

“The model said so” is never an acceptable answer. The machine executes; people own every input that tells it what good looks like.

Where human judgment stays decisive

None of this makes judgment less important. It makes it rarer and more consequential. The engine is reliably better than any person at the routine majority: thousands of decisions where the history is a fair guide to the future and the economics are already priced. It is reliably worse at three things that matter more. It cannot recognize that the world has changed before the data shows it. It cannot decide what the business values when two objectives conflict. And it cannot choose which risks are worth carrying. Those are the decisions that move results, and in the old process they received whatever time was left after the number had been assembled and defended.

What each role does, before and after

Titles differ between organizations. The eight functions below do not, and every deployment we have seen has needed all of them regardless of what they were called. Three of them are, in practice, new.

Where the same effort gets redeployed

Exhibit 3 describes changes in kind. Exhibit 4 describes the change in distribution, which is what a leadership team actually has to plan for.

Read the bottom four rows carefully, because they are the ones that get cut. Owning the cost model, running the Lab, system unlock and governing the system together account for a meaningful share of effort in the target state, and close to none today. They look like overhead in a business case, because they are not attached to a transaction and produce no visible output in any given month.

System unlock is worth pausing on, because it is the row most easily mistaken for something that already happens. It does not. It is a standing habit of asking, every cycle, what actually capped throughput this time: a cost inefficiency nobody had priced, a policy nobody remembers agreeing to, an approval step that adds delay and no information, a constraint the ledger now makes visible for the first time. Before the ledger exists there is no reliable way to find this, so the habit has nowhere to attach. Once it exists, finding and removing that one binding constraint is a higher-return use of a planner's afternoon than reviewing a hundred routine recommendations the model already has right.

Organizations that treat this as a cost program fund the automation and skip these four. The consequence is not a loud failure. An unstewarded ledger goes stale within a year, an unstaffed Lab means nothing is tested before commitment, drifting response models quietly degrade every recommendation the engine produces, and with nobody hunting the constraint the organization keeps optimizing everything except the one thing actually holding it back. The system does not break. It becomes ignorable, and the spreadsheets return.

The honest answer to the headcount question

It will be asked in the first meeting, and the credibility of everything else depends on answering it straight.

This changes what the headcount does far more than how many there are, at least over the first two years. The assembly work genuinely disappears and it is a large fraction of current effort. Against that, three functions appear that are not staffed today, and the people capable of filling them are largely the same people currently doing the assembly, which is convenient but not automatic: the transition requires deliberate reskilling rather than reassignment.

Over a longer horizon, planning organizations that complete this transition do tend to run leaner, for the straightforward reason that decision volume can grow without proportional headcount. But an organization that sells the program on near-term headcount reduction will make two mistakes at once. It will cut before the new functions are staffed, which guarantees the decay described above, and it will lose the experienced judgment that the architecture specifically depends on for the minority of decisions that were never automatable.

Where the resistance actually sits

Sponsors routinely brace for resistance from the wrong place. Front-line planners are usually the easiest constituency in the program. They are the people currently spending three weeks of every month on reconciliation, they know precisely how little of it adds value, and the offer of getting that time back is genuinely attractive.

The difficulty is concentrated one level up, among people whose authority derives from controlling the number. In an unpriced process, influence flows to whoever can most credibly assert what demand will be, and that influence is real, hard-won and about to be transferred to an explicit calculation. Nobody experiences that as a process improvement.

Three things help, and none of them is communication.

1. Give the constituency a better role rather than a smaller one. The regional lead who owns local parameters and challenges the model with recorded reason codes has more durable influence than one who wins arguments, because their knowledge now compounds into the system instead of evaporating.
2. Make the first evidence come from a peer, not from the center. A backtest run in a receptive unit and presented by its own leadership is worth several times the same analysis delivered by a corporate team.
3. Retire the shadow process explicitly, on a date, in writing. Ambiguity here is read, correctly, as an invitation to wait the program out.

Sequencing the people change

The organizational transition has its own critical path, and it runs ahead of the technical one rather than behind it.

Name the ledger owner before the first ledger is built, because an artifact without an owner becomes a study. Staff the model steward before the first model is deployed, because a model without a steward drifts silently. Stand up the Lab, even as a single analyst with a laptop, before the first large commitment is put to it, because the first time a commercial director asks the Lab a question is the moment the capability either earns its place or does not.

And change the measurement before you change the process. A planner told to trust the system while still being appraised on manual forecast accuracy has received two instructions, and will follow the one attached to their appraisal. This single misalignment accounts for more failed deployments in our experience than every technical cause combined.