Why this material, and why now

The work that gets more valuable as the models get better

Generative models have made producing analysis and producing documents close to free. That is not a reason to doubt this material. It is the specific reason the material is worth more than it was, because the thing that got cheap was never the thing that was scarce.

For about thirty years, the way a technical person became valuable was legible to everyone involved. You could do things most people in the building could not do. You could get the data out, build the model, run the analysis, and, if you were unusually good, write it up in a way a senior person could follow. Scarcity of production was the moat, and it was a real moat, defended by years of training that most people were unwilling to undertake.

That moat is draining, and faster than the profession has absorbed. A competent model now writes the query, builds a defensible baseline, drafts the deck, and produces a cleanly structured memo with a governing thought at the top and parallel supporting reasons beneath it. It will apply the pyramid correctly if you ask. It knows SCQA. It knows MECE. Much of what has historically separated the analyst who gets promoted from the one who does not is now available to anyone, on request, at a cost that rounds to zero.

It is worth sitting with how uncomfortable that is before reaching for the reassuring answer. The reassuring answer is that judgment, creativity, and human connection will always matter, which is both true and useless, because it names nothing a person could go and learn on Monday.

The question worth asking is narrower and harder. When production is free, what is actually still scarce, and is it teachable?

The answer is the judgment about which analysis could change a decision, and the willingness to own the rule that follows. Neither is a production task. Both get more valuable precisely because production became cheap.

That is the whole argument, and the rest of this page is the case for it. Four reasons follow. Each one supports the claim on its own, which is a deliberate property rather than an accident, and the last section states the conditions under which I would be wrong.

Exhibit 1The three capabilities that got cheap are the three most technical careers were built on

The capability	What it took, until recently	Where it stands now
Writing the query, building the model, running the analysis	Years of training, and the reason you were hired	Produced on request, in minutes, competently
Producing a clean, well-structured written document	A rare skill among technical people, and a visible advantage	Produced on request, to a standard above most human first drafts
Knowing the frameworks: the pyramid, SCQA, MECE, the issue tree	Consulting-firm training, or a book most engineers never read	In the training data. Applied on request, correctly, for free
Naming the decision that is actually pending, and who can commit it	Rarely taught, rarely done	Rarely taught, rarely done, and now the binding constraint
Judging which analysis could change the ranking, and cutting the rest	Indistinguishable from thoroughness when analysis was expensive	The only thing standing between an organization and forty plausible memos
Owning the rule on Tuesday, and absorbing it when the rule is wrong	The definition of authority	The definition of authority

The heavy rule does not separate hard work from easy work. It separates work that can be done without standing in the room from work that cannot.

Source: The Decision Product. The split is the argument.

The case

Four reasons, each of which would carry the argument alone

When a complement gets cheap, the surplus moves to whatever is still binding

This is not a claim about the dignity of human thought. It is the ordinary economics of a bottleneck. When one input to a process collapses in price, the process does not become uniformly cheaper. It becomes constrained somewhere else, and the return accrues to whoever holds the new constraint.

Analysis was expensive, so the organization rationed it, and the rationing did a great deal of quiet work. A team that could run four studies a year had to argue about which four, and that argument, however badly conducted, was a selection mechanism. Remove the cost and you remove the mechanism. The studies multiply, and every one of them is plausible, internally consistent, and competently written.

Nothing in that abundance tells you which one should change what the company does in April. The scarce good was never the study. It was the judgment that decided which four were worth running, and that judgment is now the only thing standing between an organization and an unreadable volume of correct, irrelevant work.

A model is built to answer the brief, and the brief is usually wrong

The first move in this method is to refuse the brief you were given. Not out of contrarianism, but because requests arrive named after deliverables rather than decisions. “We need a churn model.” “Build an agent.” “Can you look into ML.” None of those names a choice, an owner, or a date, and work that begins there produces something accurate that nobody acts on.

A language model is optimized, deliberately and successfully, to be helpful with the request as stated. Ask it for a churn model and you will get a good one, quickly, with a sensible validation strategy and a readable write-up. That is precisely the problem. The tool is a superb executor and a structurally poor challenger of the request, because challenging the request is the one behavior its training discourages.

The consequence is not neutral. Fast, cheap, competent execution of the wrong brief is worse than slow execution of the same wrong brief, because you travel much further down the road before anyone notices there was a fork. The organizations that will suffer most from these tools are not the ones that refuse to adopt them. They are the ones that adopt them enthusiastically without anyone in the room whose job is to ask whose decision this is.

Fluency has stopped being evidence that anybody thought

For as long as anyone has been writing memos, a clean document was a costly signal. Producing twenty structured, well-argued pages required having done the thinking, because there was no other route to the artifact. Readers learned, correctly, to treat polish as a proxy for rigour.

That correlation is now broken, and it broke fast. A reader handed a beautifully structured document can no longer infer from its structure that a person reasoned their way to a position. The proxy is dead, and every senior reader is in the process of working this out.

So the signal migrates. It moves to the things that a system optimizing for plausibility will not volunteer and that a person hedging their reputation will not write down: a named owner, a date, an explicit alternative that was rejected and why, a stated condition under which the author would change their mind, and an honest list of what was not analyzed because it could not have changed the ranking. Those are expensive to write because they are commitments. That is exactly what makes them legible now.

Accountability cannot be delegated to something that cannot bear loss

A decision, in the sense this book uses the word, is a commitment made by someone with authority, under uncertainty, that is still in force after the meeting ends. Authority is not the ability to be right. It is the capacity to absorb the consequence of being wrong.

A model has no such capacity, and this is not a capability gap that a better model closes. It is structural. Nothing that cannot be fired, sued, demoted, or made to answer for a quarter can hold authority, whatever its accuracy. Organizations that have tried to route a real commitment through a system rather than a person discover that the commitment simply becomes nobody’s.

Which means the last mile stays human for reasons that do not expire with the next release. Who owns the rule. Who runs it on Tuesday. What the exception path is. What you will watch, and on what date you revisit. That is the part of this work that a technical person can hold, and it is the part almost nobody is trained to do.

The part nobody enjoys

The people most exposed are the ones who were best at the automated part

The usual telling of this story has the diligent technical person displaced by a machine and the shallow generalist rewarded for talking well. That is not what the mechanism predicts. The exposure falls hardest on the person whose value was production quality: the strongest modeler on the team, the one whose notebooks are immaculate and whose write-ups are the clearest anyone produces. That person invested in exactly the capability that has been commoditized, and they invested more heavily than anyone else.

It is also not an argument for avoiding the tools, and I want to be unambiguous about that, because the sentiment is common among people who would otherwise agree with this page. Refusing to use a model to produce analysis is not a way of protecting the scarce skill. It is a way of spending your week on the part that is now free. The person who comes out ahead uses these systems aggressively for production and supplies the thing they cannot supply: the named decision, the alternatives worth ranking, the evidence that could actually flip the ranking, and a rule somebody owns.

Which brings the strong version of the claim into view. Most of what a technical career competes on is either being commoditized now or was never really yours: the tooling, the modelling, the framework vocabulary, the prose. Domain knowledge and relationships matter, but they belong to the role and the employer as much as to you, and they do not travel. The judgment about what is worth deciding, and the standing to install the answer, is the one holding that appreciates as everything around it gets cheaper. It is not the only thing that differentiates a career. It is the only one whose price is going up.

What would change the ranking

The conditions under which this page is wrong

The method holds that an argument which cannot say what would falsify it is not an argument. This page is subject to its own rule, so here are the three developments that would weaken it, and which of them I actually consider live.

Reason 02 weakens
if models become reliably willing to refuse a brief: to answer “build me a churn model” with “whose decision is this, and what would they do differently,” and to decline to proceed without an answer.
They can already be asked to do this, and they do it well when asked. But somebody has to know to ask, and has to recognize a good answer from a fluent one. The knowledge does not disappear. It relocates to whoever holds the prompt, which is a smaller group than the one that currently writes the analysis.
Reason 04 weakens
if organizations begin granting genuine authority to automated systems, with budget, liability, and a party that answers for the outcome.
Rules already run unattended in production, and have for decades. What has not changed is that a named person owns them. That change would be legal and institutional rather than technical, and it would be visible years before it mattered.
Reason 01 weakens
if the supply of analysis does not in fact rise: if organizations turn out to consume roughly the same volume of studies as before and simply produce them faster.
This is the one I would watch. It is an empirical claim about organizational appetite, not a theoretical one, and the early evidence runs the other way.

What would not change the ranking, and is worth naming because it is the objection usually offered first: a model becoming substantially more capable at analysis and writing. That is the premise of this page rather than a challenge to it. Every reason above is stated in terms of what happens when the production of analysis becomes better and cheaper, so a better model strengthens the case rather than weakening it.

Where to start

The method is free, complete, and readable in an afternoon

Eighteen entries covering the whole sequence, from refusing the brief to installing a rule somebody owns, with worked examples from replenishment, expediting, and allocation rather than from generic case studies. Nothing on those pages is gated.

Read the method Take the 5-minute diagnostic