Photo by Frank R on Unsplash

Author’s Note:

The opening example is a composite, paraphrased from several professional conversations over many years. Details have been generalized, and no specific organization, analysis, or business decision is being described.

The concern examined here is the analytical pattern itself.

~Dom

The conversation that removed it probably took about ninety seconds.

We were scoping a consolidation analysis, and someone from operations said that if the work moved, we would lose the handful of people who understood why the exceptions existed; the ones who knew which accounts were coded wrong on purpose and which customers required a step that wasn’t written down anywhere.

Someone asked whether we could put a number on that. The honest answer was no. Not no, that isn’t real. Not no, that doesn’t matter. Just no: we had no source for it, no defensible method, and no time to invent one before the deadline.

So it didn’t go into the model. The meeting moved on to data availability, and by the following week nobody in the room could have told you that the exchange had happened.

---

I know the shape of it well, because I have often been the person who said no. Building the analysis has been part of my job for nearly half of my life. Bounding it is something learned early. And the asymmetry has stayed with me: the concern usually takes about ninety seconds to leave the model, while the consequence of its departure tends to outlive the model by years.

This article is about what happens to a real cost when it fails to become a number.

A Model Is a Set of Decisions About What Counts

A financial model presents itself as a description of the business. It is not. It is a designed artefact, and like every designed artefact it has a schema, an interface, and a scope boundary that someone chose.

The schema is the set of quantities the model is willing to hold. The interface is what the decision-maker sees: scenarios down one axis, outcomes across the other, a number in every cell. The scope boundary is the line between what the model represents and what it does not, and it is the most consequential design decision in the entire artefact.

It is also the only one that does not appear in the interface.

This matters more than is obvious, because representability exists prior to decidability. An option that cannot be expressed in the schema cannot be compared, and a quantity that cannot be compared cannot meaningfully impact the choice. Long before anyone weighs the scenarios, the analysis has already determined the set of things that are eligible to matter. By the time the model reaches a decision-maker, that determination is finished, invisible, and wearing the clothing of arithmetic.

The model does not describe the business. It describes the part of the business that someone decided to make legible.

The Implied Zero

Here is the mechanical fact underneath all of this, and it is the reason the problem is structural rather than a matter of anyone’s character.

Arithmetic can represent uncertainty about a quantity we have chosen to model. It cannot preserve uncertainty about something the schema does not contain. Once a real cost is excluded entirely, its contribution to every modeled comparison is functionally zero, and not because anyone decided the cost was zero, but because the model has no place from which it could exert influence.

I call this the implied zero, and its danger is that it collapses four entirely different situations into one indistinguishable state.

There is the cost we did not measure, because no system captures it. There is the cost we cannot measure, because it resists quantification in principle. There is the cost we chose not to measure, because it was judged immaterial. And there is the cost that genuinely is zero, because the thing does not exist.

Inside the model, however, these are identical. All four are absence, and absence has exactly one value. The decision-maker receives the collapsed result and usually has no way to reconstruct which one they were handed; no marker, no footnote, no residue of the ninety-second conversation that produced it.

Worse, the collapse changes what the omission is. An unanswered question is an open question as long as someone remembers asking it. Once it passes into the model, it stops being a question at all and becomes a property of the output.

The model’s silence is indistinguishable from the model’s verdict, and a verdict is what people act on.

The Exchange Rate Was Always There

Consider a decision that eliminates a hundred livelihoods and raises operating margin by fifty percent. It is easy to imagine a board approving it without much difficulty, and it would be hard to argue they were behaving irrationally.

Ten percent? Many would still approve.

One percent? Some, if the company is large enough, and the story stays quiet.

One one-hundredth of a percent? Now the room gets quiet, and it is worth being clear about why.

The instinctive reading is that we have found a threshold. Somewhere between one percent and a hundredth of one, human cost overtakes economic benefit and the trade becomes wrong. That reading is comfortable, but I think it is mistaken. Nothing about the moral structure of the transaction changed as the numerator shrank. A hundred jobs were set to be exchanged for margin at every rung of that ladder.

The only thing that changed is that at a hundredth of a percent, the exchange became small enough to see.

Magnitude was only supplying the cover. When the gain is large, the trade reads as strategy, but when the gain is trivial, the same trade reads as arithmetic, and the arithmetic is embarrassing. We were not discovering a boundary as the gain was reduced, only losing the ability to look away.

Which means the exchange rate was already there.

Every organization that has ever restructured has one, whether or not anyone has written it down. What the model contributes is not the exchange rate but the denomination: it prices one side of the trade in dollars and leaves the other side as an implied zero, so the ratio never has to be stated, and therefore never has to be defended.

Cheaper Cognition

None of this arrived first with artificial intelligence, and treating it as an AI problem would be the most reliable way to misunderstand it.

Consider the call center that a town competes to attract with abatements, grants, infrastructure, and workforce programs, because several hundred jobs will reshape a local economy. People buy houses. Businesses open around the shift schedule. The municipal budget quietly begins to assume the tax base. Then the incentive period ends, or another geography prices lower, and the operation moves. From the corporation’s side the analysis is two lines and a difference.

From the town’s side the event has considerably more columns, and none of them were in the model.

Move back another generation and the same structure appears in offshoring. If a component costs eighteen dollars to produce domestically and seven elsewhere, an executive has an obvious question to answer, and the analysis supporting the answer is usually correct as far as it goes. The eleven-dollar difference simply does not contain displaced careers, diminished tax bases, lost institutional capability, or the supply-chain dependence that shows up decades later.

It also does not contain the wage that reached a household in the receiving country, or the local economy that grew around the new plant. That is not a caveat I am adding for balance; it is the same defect running the other direction, and it is why the honest conclusion is not that offshoring was immoral. The conclusion is that the cost model was incomplete, and incompleteness has no political direction. It omits what would have supported the decision as readily as what would have complicated it.

Offshoring found cheaper labor. Automation found cheaper machinery. AI is finding cheaper cognition. Four decades, with four different inputs, and one unchanged schema. The thing being substituted keeps moving and the scope boundary never does.

That is what makes this a design problem rather than a technology problem, and it is why each new wave feels unprecedented to the people inside it and familiar to anyone who has read the previous analysis.

I Drew the Boundary

I have set scope on analyses that preceded consequential decisions. I have built the table with four scenarios in it, and I have occasionally known which row would win before I populated the others. Sometimes context makes the answer evident before the question is asked. I have said we can’t get that about things I believed were real, and I said it in good conscience, because it was true.

I want to be careful here, because the easy version of this confession is also the useless one. Bounding a model is not a moral failure. It is, quite literally, the job.

Every quantity you admit into a schema needs a source, an owner, a refresh cadence, and a definition you are prepared to defend in front of someone who does not want to hear it. A model that attempts to contain everything ships sometime around never, and an analysis that arrives after the decision has been made has helped no one. Unbounded scope is not rigor. It is paralysis.

So the boundary is correct; the problem is what we do with the record of having drawn it.

In architecture, I hold that a deliberate deviation from a stated principle is not a failure but a debt, and that the debt has to be booked somewhere: owner, reason, scope, ending condition, and the principle that yielded. I have argued this at length, here and elsewhere, and I believe it.

A decision that is made once and remembered nowhere becomes precedent, and precedent is the most expensive form of forgetting.

A scoping decision, then, is an architectural decision. It has all the same properties: it is deliberate, it is consequential, it constrains everything built on top of it, and it will be inherited by people who never attended the meeting.

And in nearly twenty years of building these things, it has never been a required artefact.

What Analysis Owes

Analysis cannot supply ought; it has no mechanism for it. A model can tell you what the options are, what they cost, what they produce, and which risks can be estimated; it cannot tell you what human consequence is worth, because worth is not a quantity the schema can hold.

Good analysis does not tell leaders what to do. It tells them, as completely as possible, what they are choosing.

But there is a narrower obligation that analysis can absolutely meet, and we have simply not been in the habit of meeting it. Analysis can refuse to launder the omission.

That requires an instrument, but it is a modest one. Alongside the model, a register of what was raised and did not make it in. For each entry: what the concern was, in the words of whoever raised it. Why it isn’t in the model: unmeasured, unmeasurable, or judged immaterial. Who bears it if the scenario is chosen. And what would have to change for it to enter the schema later.

Notice what this does not do. It does not quantify anything, it does not simulate a conscience, and it does not pretend that institutional knowledge has a dollar value we all simply never bothered to compute. It does exactly one thing: it restores the distinction that the arithmetic destroyed. Four states go into the model as one; the register keeps them four on the way out.

A named zero is no longer an implied zero, because a named zero can be argued with.

It also will not make decisions easier or kinder, and I would distrust it if it claimed to. A leader can read the register in full and sign the restructuring anyway, and sometimes that is the correct decision. The alternative to a painful choice is frequently a worse one arriving later with more people attached.

What the register removes is a specific and very common form of shelter: the ability to say afterward that the analysis didn’t raise it.

The Weight of the Pen

There is a reason expensive pens became symbols of authority. A few ounces of resin and gold, and we attach to them contracts, appointments, acquisitions, restructurings, and terminations. A few centimeters of ink that moves millions of dollars and alters thousands of lives.

The pen carries the mythology of the decision. What it carries much less of is the decision itself.

By the time the page reaches the desk, the option space is closed. There are four scenarios because someone chose four. There is a margin column because someone selected the outcome variable. There is no column for the people who knew why the exceptions existed, because ninety seconds in a scoping meeting settled that, and nobody wrote it down.

I do not say this to absolve the person who signs. Someone still has to answer for what the page causes, and the collapse of distributed analysis back into a single act of individual agency is exactly what a signature is for. I say it to deny the same shelter to the person who built the instrument. The analyst’s authority is not the authority to decide. It is something more subtle and, in its own way, harder to inspect: the authority to determine what the decider will be able to see.

Organizations will always count. They must; no one governs an institution of thousands through intuition. The counting is not the failure.

The failure is that when we finish counting, we hand over a page on which the things we could not measure and the things that do not exist look exactly alike. Everything omitted from the calculation is still out there, unpriced and undiminished, waiting where we left it.

A model reports a number for what it measured. It reports one for what it didn’t, too. The whole of our obligation is to make sure someone can tell which is which.