The Thing AI Cannot Compress

01 — The Commoditisation Claim Is Too Coarse

The argument that AI commoditises expertise conflates three things that are not the same. The first is idea generation — producing an output that resembles the output a skilled person would produce. The second is discrimination — recognising which of many plausible outputs is actually correct for this situation, this client, this moment. The third is judgment — making a consequential decision under uncertainty with incomplete information and bearing accountability for the outcome. AI has replaced the first almost entirely. It is replacing the second at whatever rate model capability improves. It has not replaced the third.

What most people call "expertise" is a bundle of all three. When they observe that AI can now produce a first-rate memo, a credible analysis, a competent draft contract, they conclude that expertise has been commoditised. What has been commoditised is the generation step — the ability to produce a plausible output. The discrimination step — knowing whether the memo is actually right, whether the analysis captures what matters, whether the contract clause will hold under the governing law of the relevant jurisdiction in a dispute — is being commoditised more slowly, but it is being commoditised. And the judgment step — deciding what to do, and answering for it — has not been commoditised at all.

But "at all" requires a distinction that the prevailing discussion does not make. Judgment resists commoditisation for two reasons that should not be conflated. The first is epistemic: consequential decisions under genuine uncertainty draw on something models do not yet reliably have. The second is institutional: there exists no structure through which a model can bear accountability — no entity to sue, no licence to revoke, no career to end. The first moat erodes at the pace of capability. The second is a fact of law, and facts of law can be changed by statute. A professional whose value rests on the judgment layer should know which of the two moats they are actually standing behind, because they have very different durabilities.

The confusion matters because it leads to the wrong prediction about who is threatened. The routine executor — the person whose job was primarily to produce legible outputs — is genuinely exposed. But the person whose value was primarily in discrimination and judgment is less exposed than the prevailing narrative suggests. And there is a third category, less discussed, whose value is more defensible than either: the person whose value is accumulated context that cannot be extracted from the relationship in which it was built.

02 — What Context Density Is

Context density is not expertise in the abstract. It is expertise anchored to a specific system — a specific organisation, a specific market, a specific set of relationships — at a specific moment in that system's history. It is the difference between knowing how MiCA works on paper and knowing how a specific competent authority is reading it this quarter, given the political context it operates in and the enforcement action it is trying not to repeat.

It has three layers. The institutional layer: how the organisation actually moves — not its stated decision-making process, but the real one, the informal one, the one that depends on who is in the room and what they owe each other. The relational layer: who actually decides — not the person with the title, but the person whose discomfort prevents decisions, whose enthusiasm accelerates them, whose silence on a proposal means it is dead. The temporal layer: what the system is anxious about right now — the live constraint, the recent embarrassment, the unresolved tension that makes certain conversations impossible and certain proposals compelling at this particular moment.

Context density is perishable in a way that abstract expertise is not. A lawyer who knows contract law knows contract law indefinitely. A lawyer who knows how this client's general counsel approaches risk, what their board is anxious about after last quarter, and which counterparties they will not negotiate with regardless of commercial logic — that knowledge expires when the general counsel changes, when the board composition shifts, when the counterparty relationship is resolved. It cannot be stored. It can only be continuously renewed through engagement.

The difference between knowing how MiCA works and knowing how a specific regulator is reading it this quarter is not a matter of detail. It is a different kind of knowledge entirely — one that lives inside a relationship, not inside a document.

03 — Access, Not Capacity

AI has breadth without anchorage — but the constraint deserves precision, because it is usually stated wrongly. The constraint is access, not capacity. A model with persistent memory, fed a firm's transcripts, correspondence and meeting records, accumulates organisation-specific context perfectly well; anyone who has worked alongside one for months knows this. What it cannot reach is the layer nobody records. The institutional layer of context density is unwritten not because recording is technically hard but because writing it down would itself be a political act. Nobody minutes the real reason a proposal died. Nobody documents the chief executive's private anxiety. Nobody commits to text that the general counsel's silence means no. People navigate these things. The navigation accumulates in the person who does it repeatedly. It does not accumulate anywhere else.

This makes the argument contingent rather than categorical, and the contingency should be stated plainly. Firms record more every year. The unwritten layer shrinks at the margin, and the portion of context density that a machine can hold grows with it. What remains defensible is not everything undocumented — it is the subset that is undocumentable, the knowledge whose recording would change the politics it describes. That subset is smaller than most advisers would like to believe. It is also where most of the value sits.

The distinction between simulation and accumulation is the key. AI can simulate context density — it can produce outputs that look like the outputs of a person with deep system knowledge. A well-prompted model can write a memo that sounds as though it understands the organisation's internal politics. But simulation is not accumulation. The simulation is produced fresh each time, from general patterns. The accumulation is built over time, from specific engagements, and it is the history of those engagements — not the outputs they produced — that constitutes the value.

A client can extract the memo. They cannot extract the judgment about when not to send it, the read on how the counterparty is likely to respond given what happened eighteen months ago, the sense of which argument will land and which will create a problem that takes six months to resolve. Those are not outputs. They are the accumulated residue of sustained engagement — and they live only in the person who has been engaged.

04 — Why It Compounds — and Why It Decays

Legible expertise compounds too, but it leaks. Each refinement of craft eventually becomes describable; everything describable eventually becomes teachable, reproducible, automatable. The compounding accrues first to the profession, then to the market, then to the model. A lawyer who produces excellent contract drafts this year does not own next year's advantage — the patterns that made the drafts excellent are already in the training data.

Context density compounds differently because each engagement adds a layer that accrues to one person. The second year of advising a client is more valuable than the first — not because the adviser has become more expert in the abstract, but because they have accumulated a year's worth of institutional, relational, and temporal knowledge that the client cannot get from anyone else. The third year compounds on the second. The advantage is not visible in any single output. It is visible in the aggregate — in the quality of judgment that can only be exercised by someone who has been in the room repeatedly, across multiple decision cycles, through at least one crisis.

This appears to contradict the perishability claim made earlier. It does not — and the resolution is where the defensibility actually comes from. Context density compounds only under continuous engagement; the stock begins decaying the moment engagement lapses. The general counsel changes, the board rotates, the anxiety that organised last year's decisions resolves itself, and the accumulated layer thins. Compounding requires that renewal outpace decay. That condition is precisely what makes the asset defensible: a competitor cannot buy the stock, cannot acquire it by study, and cannot pause its own accumulation without losing ground. It is also what makes the asset worthless on a curriculum vitae. It exists only inside the engagement and dies outside it.

The invisibility of the advantage is not a weakness. It is the source of the defensibility. A client who can describe precisely why an adviser is valuable has implicitly described the output in terms that could be reproduced. A client who cannot fully articulate why they value an adviser — who knows only that the judgment is reliable, that problems seem smaller after the conversation, that the read on a situation consistently proves accurate — is describing something that cannot be specified in a procurement document and therefore cannot be replaced by whatever wins the procurement.

Tacit value is defensible precisely because it is tacit. The moment it becomes fully legible it becomes reproducible — and the moment it becomes reproducible it becomes a commodity.

05 — The Market Mispricing

The market currently prices expertise primarily through legible proxies: credentials, institutional affiliation, track record, output quality on observable transactions. These proxies worked reasonably well in a world where the legible components of expertise were the scarce ones — where producing a high-quality output required sustained human effort and the quality of the output was therefore a reasonable signal of the underlying capability.

That world is ending. When high-quality output can be produced by anyone with a good prompt and twenty minutes, the output itself is no longer a signal of the underlying capability. The track record of outputs becomes a record of what was legible, not what was valuable. The credential certifies that someone can produce the kind of work AI now produces for free.

The mispricing persists because buyers are in a transition period. They know the proxies are degrading but they do not yet have better ones. They continue to use credentials and output quality as selection criteria because those criteria are legible and comparable — they can be written into a request for proposal, evaluated by a committee, defended in a procurement process. The shape of the resulting failure is familiar to anyone who has sat on either side of a beauty parade: the committee scores writing samples and credentials, selects the firm with the cleanest deck, and discovers a year later that the deck was the product — and the product is now free.

The adverse selection dynamic in this transition is not obvious. Buyers who continue to select on legible proxies will systematically preference advisers who have optimised for legible output — which, in an AI-augmented world, means advisers who are good at using AI tools rather than advisers who have accumulated irreplaceable context. The best advisers by the new criteria will be underselected during the transition period precisely because their advantage is not legible to buyers who are still using the old criteria.

The illegibility defence has a dark twin, and intellectual honesty requires naming it. The same opacity that protects genuine context density also protects its counterfeits: capture, sunk-cost attachment, relationships sustained by comfort rather than value. That inarticulate client a moment ago — unable to say quite why the adviser is worth keeping — may be describing irreplaceable judgment, or may be describing an incumbency rent, and from the inside the two can feel identical. The distinguishing test exists, but it is not a credential. Context density is opaque in mechanism and testable in output: it generates specific, falsifiable reads — how the regulator will respond, where the counterparty will move, which board member will quietly kill the proposal — that are subsequently confirmed or not. Incumbency rent generates reassurance. An adviser whose predictions can be logged and checked against events is selling the former. An adviser whose value survives only as a feeling is, at best, untested.

Who gets exposed first is the legible executor — the professional whose primary value proposition was the production of high-quality outputs at speed. Not because they are unskilled, but because the specific skill they have is now available at near-zero marginal cost from a different source. The transition period is kind to them in one respect: buyers are still paying for output quality. It is unkind to them in another: the window during which that pricing persists is closing.

— Implications

For individuals, the implication is about where to invest attention. In an AI-augmented career, the return on investing in legible capability — learning to produce better outputs faster — is declining. The return on investing in context density is increasing — but the asset cannot be targeted directly, because by its nature it cannot be fully introspected even by its holder. What can be targeted are the conditions under which it accumulates: sustained engagement with one system rather than shallow engagement with many, presence across repeated decision cycles, availability through at least one crisis. The asset accrues as a byproduct of those conditions or not at all. Technical competence is now the floor, not the ceiling — and the ceiling is set by something that cannot be acquired by study or practice in the abstract.

For firms, the implication is about what to protect and what to let AI absorb. The processes and outputs that can be described precisely enough to be automated should be automated — the cost of not automating them is competitive disadvantage. What should be protected is the client relationship depth, the institutional knowledge, the judgment capacity of the people who have been engaged with specific systems long enough to have accumulated genuine context density. Those are not replaceable by AI and are not replicable by competitors who have not made the same investments in sustained engagement.

For clients, the implication is about how to evaluate advisers when generic execution is free. The question is no longer whether an adviser can produce good work — that bar is now low enough that almost any competent professional clears it with AI assistance. The question is whether an adviser has accumulated the kind of system-specific knowledge that makes their judgment on this particular situation, at this particular moment, worth more than a well-prompted model. Credentials and output samples cannot answer it. The test from the previous section can: ask for the adviser's specific reads in advance, log them, and check them against events. Context density predicts. Its counterfeits reassure.

Generic execution is now free. Anchored judgment is not — and the interval before the market learns to tell the two apart is where the next decade's professional fortunes will be made and lost.