What Doesn't Compress

The middle is what automates

Automation never comes for the top of a job or the bottom. It comes for the middle, the mechanical, repeatable, teachable part that sits between deciding what to do and being accountable for whether it worked. Compensation, leverage, and career durability then relocate to the inputs the machine couldn’t reach.

This is not an AI claim. Run it against a case with a century of hindsight.

Before mechanized bookkeeping, an accountant spent most of a career on arithmetic. Ledgers were posted by hand, columns were footed by hand, trial balances were reconciled by hand. The person who could add a long column quickly and without error was scarce and paid for it. Then the adding machine arrived, then the spreadsheet, and the arithmetic went to roughly zero cost. The accountants did not disappear. The ones who owned the arithmetic and nothing else disappeared. The ones who moved up, to whether the numbers meant the business was healthy, whether the treatment was defensible, whether the client should be trusted, became more valuable. VisiCalc did not thin the ranks of people who understood a business. It thinned the ranks of people who could only foot a column.

The pattern is consistent across waves that have nothing to do with each other. The telephone operator’s switchboard work automated; the judgment about which calls mattered did not. The draftsman’s line work went to CAD; the architect’s decision about what to build did not. In each case the middle, the trained, mechanical execution that used to take years to master and command a wage, collapsed toward free. The scarce thing sitting on top of it got repriced upward.

Hold that shape in mind, because the current wave has the same shape. Drafting, summarizing, first-pass analysis, boilerplate code, the competent-but-generic memo: that is the middle of a knowledge job, and it is going to roughly zero cost. What sits above it is the question.

Abundance reprices its complement

When one input to a process goes to zero cost, the paired input that can’t be automated becomes the constraint, and the constraint captures the value.

This is the cleanest way to reason about any automation wave, and it is more reliable than guessing which tasks a model can do. Every task is a bundle of inputs. Producing a legal brief requires research, drafting, and the judgment to know which argument a specific judge will find persuasive. Producing a product spec requires synthesis, writing, and the taste to know which of 40 reasonable features is the one that matters. When the automatable inputs (research, drafting, synthesis, writing) fall to near-free, the whole bundle gets cheaper, but the value inside it does not vanish. It migrates to whichever input is still scarce. That input becomes the bottleneck, and bottlenecks price high.

Four inputs resist automation hard enough to absorb the repricing. Judgment, taste, context, trust.

Judgment is the call you make under uncertainty when the data underdetermines the answer. A model can enumerate the options and estimate the odds. It cannot own the decision, because owning a decision means being accountable when it goes wrong, and accountability is not a thing you can generate.

Taste is knowing which of many correct answers is the right one for this audience, this product, this moment. A model produces the average of everything it has seen. Taste is the deliberate departure from the average, the choice that a corpus, by construction, cannot recommend because the corpus is the average.

Context is the local, un-written knowledge that never made it into any training set: why the last redesign failed, which executive will veto this, what the customer actually meant in that call. The model has read everything published and nothing private, and most of what governs a real decision is private.

Trust is the accumulated record that makes other people act on your word. It is the least automatable input because it lives entirely in other minds. A model can draft the message. It cannot be the person the reader has decided to believe.

Why these four resist

The four are not a claim about human specialness. Nothing mystical is happening. The distinction is mechanical: some things can be learned from a corpus, and some things can only be earned through outcomes.

A model learns from a corpus, text that already exists, patterns already recorded. Anything that has been written down enough times becomes learnable, which is exactly why the middle of every knowledge job is the part that automates. The middle is the documented part. It is the procedure, the template, the worked example, the thing a senior person could hand to a junior with instructions.

Judgment, taste, context, and trust share one property: each is earned through accumulated exposure to consequences, not from reading about consequences. You develop judgment by making calls and living with the results. You develop taste by shipping things and watching them land or fail in front of real people. You accumulate context by being in the room for the decision that later turned out to matter. You build trust by being right, or by being wrong and owning it, over enough cycles that other people update toward you.

None of this is in the corpus, because the corpus is the output of the process, not the process itself. You can read every published post-mortem and still not have judgment, the way you can read every book on swimming and still drown. The gap between knowing the account of a consequence and having absorbed the consequence is the whole game. That gap is what doesn’t compress.

Packy McCormick framed the near end of this as “most human wins.” The mechanism underneath it is narrower and more useful: not the most human, but the input that had to be earned through outcomes rather than learned from a record.

What this predicts

A lens that doesn’t predict anything is a mood. This one makes calls.

Roles that are mostly middle thin out. The pure execution jobs, the ones where the value was the trained ability to produce competent, standard output on demand, compress hardest. Junior analyst work that is 80 percent research-and-format. Copywriting that is 80 percent producing serviceable prose to spec. First-line code that implements a well-specified ticket. Not because those people lack value, but because the specific input they were paid for is the one going to zero. The half of those roles that was judgment survives and gets more valuable. The half that was throughput evaporates.

Roles that are mostly the scarce four thicken. The senior operator whose job is deciding what to build and being accountable for it. The founder whose entire function is judgment and trust. The designer whose value was always taste and never pixel-pushing. The person other people call before they make a decision. These roles don’t just survive the wave; the wave makes them more valuable, because it strips away the mechanical work that used to dilute their time and leaves them doing only the part that couldn’t be automated.

The falsifiable version: over the next several years, compensation dispersion inside knowledge roles widens, not narrows. The gap between the top and median performer in the same title grows, because automation collapses the floor of competent-execution value that used to compress the range. If instead the wave flattens pay, if median and top converge as the tools equalize everyone, the lens is wrong. I don’t expect that. Watch senior-versus-junior pay ratios in software, design, and analysis. If they widen, the lens holds.

The individual move follows directly: get out of the middle. Move toward the inputs that are earned, not learned. This is why the memos circling this argument all pointed the same way, that a resume of skills is a depreciating asset, that the product role is splitting into the half that specs and the half that decides, that the value of a document was never the drafting.

The claim, narrowed

Here is the strongest objection, and it is a real one. Every generation said exactly this about its own wave. The bank tellers were supposed to relocate to relationship judgment; the ATM came, and mostly the tellers just left. The frame that says “the scarce human input survives” has been used to reassure people right up until the scarce input turned out to be automatable too, one wave later. Why is this time different?

It isn’t, and that’s the point. This is not a claim about a permanent floor under any specific skill. Judgment, taste, context, and trust are where value sits during this transition, where it relocates as this wave commoditizes this middle. Nothing here promises those four stay scarce forever. A later wave may learn to earn what today can only be learned from a corpus, and if it does, the value relocates again to whatever is scarce then. The lens describes the direction of the flow, not a final resting place. That’s the honest version of the claim, and it’s still the useful one, because you don’t get to skip the transition you’re standing in by pointing out that another one comes after it.

So the question is never whether the scarce input is permanently safe. It’s which input is scarce right now, in your job, this cycle, and whether the part of your work you’re proudest of is the part that got written down, or the part you had to earn.