What the industrialisation of knowledge does not transfer: a minimal axiom, five layers of residue, and the embedded deployment of AI in regulated systems as the domain of demonstration
Five notes published over recent months on the embedded deployment of artificial intelligence in the enterprise were written separately, on separate ground, none of them claiming to explain the others. They converge nonetheless. This note advances the hypothesis that the convergence is not an authorial effect but a structural one: the five observations are not five results but five manifestations of one, produced by a single cause. That cause is neither AI, nor the cloud, nor the strategy of any one vendor. It is a property of every industrialisation of knowledge. The reasoning does not depend in any way on the actual performance of the models; it bears on the economic and organisational logic of the deployment apparatus, not on what the models can do.
To industrialise knowledge is to produce a representation stable enough to be replicated independently of its context of origin; every representation of that kind discards context; the discarded context does not vanish and must be carried by an actor who is necessarily local, and who thereby inherits control rights, recurring costs and an institutional exposure no one has contracted for. That remainder has a name: contextual residue.
The temporal anomaly is an asymmetry of windows: the vendor recognises its return near go-live, while the client can establish its own only after the full cost of permanence has been observed. The institutional anomaly is a defect of subscription: an agent may decide within a process with a perfectly identifiable operational engagement while no authority has subscribed ex ante to the consequences. The patrimonial anomaly is a displacement of the asset: what is actually deployed is not the model but the operating capability constituted at the client. The metrological anomaly is a confusion of units: adoption at go-live is not deployability, and billed consumption is not demand. The strategic anomaly is a choice of response: when the logic of platform return on investment seizes up, an infrastructure vendor answers with a deployment apparatus rather than with a feature.
Five domains, five anomalies, one author. The temptation would be to write that the corpus proves them. It proves nothing: it observes. A theory does not ground the observations that precede it; it accounts for their co-occurrence. The question is therefore not to demonstrate each anomaly again, but to explain why they appear together.
To industrialise knowledge means here to produce a representation stable enough to be replicated at scale independently of its context of origin. The definition separates four operations that discourse conflates: to standardise, to deploy, to replicate, to scale. Only the first is at issue here, the other three being its operational consequences. What matters is the claimed independence from the context of origin. Without it, there is no industrialisation, only displaced craft.
Contextual residue designates the set of determinants of a decision that cannot be captured, transferred or represented by the abstraction used to industrialise a body of knowledge. The critical distinction plays out here, and it is counter-intuitive: residue is not missing information that could be collected later. It is what the representation, by virtue of the very thing that makes it replicable, cannot carry. An additional collection budget does not reduce it; it moves the boundary of the abstraction without abolishing its outside.
A single statement, as primitive as possible. Every abstraction loses the context it does not represent. The statement is information-theoretic before it is organisational: to abstract is to retain a representation and discard the rest, and what is discarded is not destroyed, it ceases to be carried. It speaks neither of AI, nor of embedded deployment, nor of liability. It is credible to anyone who has never heard those terms, which is precisely what one expects of an axiom.
The intellectual debt is explicit. Herbert Simon showed that organisations do not handle complexity by representing it in full, which their bounded rationality forbids, but by constructing hierarchical and deliberately incomplete representations. Near-decomposability is not a property of the world; it is what the organisation imposes on the world in order to process it. The axiom states the price of that imposition: what does not enter the carving is not absent from the real, it is absent from the representation. Simon does not come along to prove the axiom after the fact; he explains why an organisation that industrialises has no option but to produce residue.
An objection must be raised before a reader formulates it, because it is the only one that could kill the text: if the axiom were merely an abstract restatement of the anomalies it claims to generate, the theory would be circular. The safeguard is twofold. The axiom must be statable without any of its consequences, which has just been done. And every link in the derivation must import a premise external to the axiom. Were a corollary to follow by mere renaming, the axiom would fall.
To industrialise is to abstract at scale. By the axiom, that abstraction loses context. The lost context is the residue, and it is local by nature since it is attached to the place and the moment that produced it.
From this follows a theorem, to be distinguished from the axiom on pain of confusing the cause with its manifestation: for an activity whose value is contextual, the residue cannot be reduced to zero without cancelling the abstraction, hence without cancelling the industrialisation itself. This property, contextual irreducibility, separates decomposing without loss from decomposing with remainder, and everything that follows plays out in the second term.
Since the residue can be neither standardised nor eliminated, some actor must carry it, and that actor is necessarily local. First external premise: the theory of incomplete contracts, from Grossman and Hart to Hart and Moore, establishes that when future states cannot all be specified, residual control rights determine who decides in unforeseen situations. The actor who carries the residue holds those rights de facto, being alone in a position to arbitrate what the representation did not anticipate.
One must stop here, because the next step is where a demonstration substitutes words for proof. That the actor is local does not say that the actor is the client. A systems integrator, a managed operator, an industrial partner, an operating subsidiary or a regulator may occupy that position. Identifying the local actor with the client is an additional hypothesis, true within the domain studied, where the vendor takes neither operating responsibility nor residual control rights. Elsewhere it will have to be shown rather than assumed. The restriction does not weaken the theory; it delimits its field, which is not the same thing.
A concept that explains everything explains nothing. Residue, as just posited, would account for return on investment, governance, liability, contracts, value capture and deployability. That fecundity is suspect: it signals an omnibus concept, and an omnibus concept is hard to refute, therefore uninformative. The remedy is not to reduce the ambition but to stratify. The stratification is not a typology appended after the fact to save the theory; it is the chain of derivation, read across.
Informational residue is what does not pass into the representation. It is the only layer that follows directly from the axiom, and it would be refuted by a representation of contextual knowledge that discards nothing, which would amount to abolishing abstraction.
Organisational residue is what remains to be coordinated locally. It arises from co-evolution: an architecture of knowledge evolves continuously with processes, data, teams, regulation and objectives, and each local evolution regenerates a remainder. The problem is not knowing the business, which a vendor’s experts are often perfectly able to do, but evolving with the business, which does not replicate across thousands of simultaneous engagements. Domain depth does not run up against competence, it runs up against the scale of co-evolution.
Decisional residue is what remains to be arbitrated. Two movements are conflated by dominant discourse. The vendor externalises an operational risk by placing its engineers in contact with the real. The client internalises a decisional risk, that of unspecified situations. It is not the vendor abandoning a risk; it is the client inheriting another.
Legal residue is what remains to be assumed institutionally. To sign is to subscribe ex ante to consequences; yet the consequences that matter are those the representation did not specify. A signature bearing on the residue does not standardise, since the residue does not specify. It is therefore omitted. One term must be fixed, as two neighbouring notions are commonly confused: an organisation may be liable without having signed, and may sign without being wholly liable. The term retained is the institutional signature, the act by which a designated authority converts a technical delegation into a commitment of the institution. It is the signature that is missing, not the liability, which always ends up being attributed, generally by a court and generally late.
Economic residue is what remains to be captured as value. What is industrialised is measured, and what is measured becomes the proxy on which return is computed: consumption, billing, metered usage. These proxies are standardised, hence captured early; net value resides in the residue, hence reveals itself late. It is the same cut that separates the cost of the token from the price of the decision. Second external premise: Teece showed that creating value and appropriating it belong to two distinct regimes, governed by appropriability and complementary assets. The vendor does not hold the residue; it holds the infrastructure without which the residue produces nothing. A reservation is in order: this does not establish that the vendor captures value at the client’s expense. Both may gain. What the structure renders indeterminate is the distribution of the value created, not its existence.
Each layer comes with its falsification condition. An apparatus supplying, across thousands of simultaneous engagements, a depth equivalent to that of the client’s internal experts would refute organisational residue. A contract designating ex ante the institutional signatory for the behaviour of the agents would refute legal residue. A demonstration that billed consumption tracks symmetrically the net value appropriated by the client would refute economic residue. No layer is obtained by renaming the axiom, subject to the reservation that its external premises are themselves debatable, as every premise is.
A theory of residue must be able to show it somewhere. Proponents of embedded deployment speak abundantly of governance: trust, tooling, responsible AI, agent management. They never name, ex ante, the institutional signatory for the behaviour of the agent deployed in a regulated process. Microsoft Frontier Company, announced on 2 July 2026 with two and a half billion dollars and six thousand people embedded at clients, insists on measurable outcome and on trust without attributing to anyone the legal responsibility for the decisions produced. This occurrence is an illustration, not a demonstration: the theory would hold without it, and that is precisely how an illustrative example is recognised.
What that silence does not establish must be named. It establishes no intent. Three competing explanations account for it just as well: commercial simplification, since legal matter weighs down a pitch; contractual prudence, since one avoids committing in writing to a liability; immaturity of the framework, since no one yet knows, in law, who signs for an agent. None has been eliminated. The silence is therefore an empirical signature compatible with the theory, an indication, never a proof. That modest status is what makes it useful, because it comes with its condition of refutation: let a single proponent of embedded deployment publish a communication designating ex ante the institutional signatory for the agents it deploys, and the indication collapses.
What this note demonstrates bears on one domain: the embedded deployment of AI in regulated systems, where the five anomalies were observed, where the chain holds, where the local actor is the client. What the axiom licenses one to conjecture goes further. Every industrialisation of a cognitive service abstracts contextual knowledge, therefore produces residue, therefore leaves something to a local actor. Consulting, software publishing, medicine, law and education ought to exhibit the same stratification. This is not a result but a research programme, testable and not yet tested. A general theory proved on a single case would be an essay; stated general, proved on a single case, the remainder posited as conjecture, it is a programme.
The limits must be set without evasion. The axiom does not forbid embedded deployment; it forbids embedded deployment without residue, which is not the same thing. An apparatus with strong capability transfer brings the client’s break-even nearer and reduces organisational residue without ever cancelling it: residue is minimised, it is not suppressed. A managed operator remunerated on persistent performance does not refute the theory, it displaces the local actor. Finally, the axiom rests on an information-theoretic intuition this note has not formalised; it treats it as a point of departure, not as a demonstrated theorem, and this is the most attackable point of the whole.
From this follows one prescription, architectural and not moral. Legal residue is not filled at the level of governance, where the tools operate, but at the level of signature, where no one operates. That is therefore where to act. A contract of agentic governance must designate the institutional signatory ex ante, as a precondition of deployment and not as documentation produced after the fact. Designating the signatory does not eliminate the residue. It names who carries it, which is the only operation available. That is the difference between a risk assumed and a risk inherited.
The gesture is the one running through this corpus: name the distinction that separates two regimes of operation, one viable under a regulated frame, the other not. Here the distinction is ontological before it is organisational, since it separates what a representation carries from what it discards. It explains why AI governance is a problem of architecture rather than of declarative policy, why residual economic risk stays on the client’s balance sheet even when the system is delivered conforming, why benchmark performance is not deployability, why persistence is not a property of revenue but of the capability that generates it, and why the responsibility frontier moves without ever disappearing. You can industrialise a representation. You cannot industrialise what made it right. That remainder has a name, a bearer and a price: the residue, the local actor, and the commitment no one has signed.
Full text freely available in the PDF below (7 pages).
Doctrinal notes and explorations on AI in regulated systems. Once or twice a month. One-click unsubscribe.