Why AI becomes governable only when a decision binds its cost, its accountability, and its value.
Enterprise AI in mid-2026 is bought by the cent per token: rates per million tokens, GPU hours, a frontier model weighed against an open one on a unit-cost grid. This governs one thing well, infrastructure consumption, and nothing beyond it, least of all the decision. What follows does not price AI. It describes how an organization turns a technical event into something it can at last observe, impute and govern. The domain of validity is deliberately narrow: it holds where the organization has owners, a hierarchy and accountability, and not for conversation, ideation, or fully automatic systems where no human decision is attested. The restriction does not weaken the thesis; it solidifies it.
A cost is governable when an actor holds the institutional lever to observe it, price it, and impose the behavior attached to it. Of these three verbs the token serves one: it observes consumption with real precision, prices no value, and constrains nothing correlated with value. To observe without pricing or constraining is not to govern; it is to bill. FinOps governs the infrastructure layer perfectly and stops there. The enterprise can govern further only because it controls the point of access to the models, and that control is the enabling condition of governance, never its sufficient one. Holding the door is necessary. It is never enough.
Coase explains why the firm internalizes an activity; Williamson explains how it governs what it has internalized. The point of access is a governance structure in Williamson’s sense, the frontier where a firm can observe, price and constrain. But a structure governs nothing until it has an object to act on, something observable once, priceable once, imputable to someone. The door is necessary and empty; what fills it is an accounting unit. That unit is the decision, taken here not as a metaphysical primitive, a fragile ontological claim, but as a unit of commensuration: an accounting unit, because it renders comparable what was not.
Searle’s distinction between brute and institutional facts fixes the move: a model call is consumed, a validated decision is endorsed. We call binding (le nouage) the institutional operation that ties, inside one governable object, a resource consumption, an organizational accountability and an economic consequence. Before the binding these three live in separate registers with no common denominator: consumption in the technical logs, accountability in the org charts, value in the operating results. The binding is the mechanism; governance is its consequence. The decision is simply the first object in enterprise AI where the binding appears clearly, on one condition that is part of the thesis, that a human endorses it. From the binding follow five properties of an accounting unit, as consequences and not axioms: identifiable, attributable, closable, auditable, additive. Each eliminates a rival unit, the prompt is not closable, the workflow not identifiable, the case file not attributable, the result not auditable at the right instant.
The institutional cost is not the value of a decision; it is what the organization commits to produce it and to stand by it. It aggregates two realized costs, compute and human supervision graded by criticality, and one provisioned exposure: the risk charge the owner accepts in endorsing, treated as a provision for decision risk, estimated on entry, adjusted periodically, settled only when one finally knows what the decision actually produced. Knight grounds the closure, ex ante uncertainty settling only ex post; Hayek adds that the knowledge to evaluate it is dispersed and reveals itself only in use. One does not price a decision; one settles it. Additivity, taken literally, would be false: one does not add decisions, heterogeneous in risk and horizon, but their institutional cost, which the binding has reduced to a common metric. That is what turns traceability into an instrument of management.
Three operations are habitually collapsed into one and must be held orthogonal: route to the right model, qualify what counts as a decision, allocate accountability. Layer 1, technical optimization, is a solved commodity, as FrugalGPT and RouteLLM show. Layer 2, institutional qualification, is the heart of the apparatus, more than Layer 3: a binary question, does this event acquire an institutional status, on which everything downstream depends, since without qualification there is no inscription, hence no imputation, no attestation, no accountability. Layer 3 allocates accountability, and the guardrail is strict, the ledger records, it does not govern: register, then measure, then management decision, then governance. The instrument is a ledger of decisions, not an ERP, and the word is chosen: an account that aggregates, reconciles and budgets institutional cost, not a journal that merely stores a history.
This shifts usage optimization from ex ante filtering, which prices a supposed intention, to ex post commitment, which settles an assumed one. Throttling the model against the presumed laziness of a prompt judges the user before the act, manufactures resentment and flight, and revives the very shadow AI it claims to reduce. The ledger counts after the act, with no opinion on the user. Human attestation is not a satisfaction rating granted to the AI but a choice of endorsement: endorse, and you trigger the pricing of institutional cost; rework, and the correction time counts as iterative supervision; leave the output neither endorsed nor reworked, and it settles as technical waste imputed to infrastructure. One cannot use a deliverable and then declare it had no value. The thesis produces testable predictions, which distinguishes it from a posture: at identical compute, different institutional qualification yields different institutional cost (P1); at constant qualification, institutional cost moves independently of inference cost (P2); technical optimization shows diminishing returns while qualification is absent (P3).
A solid theory names its own failures. The ledger governs the cost and the accountability of the decision, not its quality: pricing responsibility does not guarantee discernment, which still depends on workflows, incentives and rights. It fails if qualification is blurry and lets everything in or out, if no decision has a named owner, if the granularity is wrong, or if its implementation cost exceeds the cost it lets one govern. The field excludes conversation, ideation, fully automatic systems, and organizations without a clear owner, where the unit loses its attributability. And the whole apparatus holds on one condition: the enterprise controls the point of access, the one place where the door exists by default. Remove the door and there is neither routing, nor accounting, nor attestation, which is precisely the situation at the scale of the ecosystem, where no actor holds this position by default and governability must be created rather than presumed. That is the object of the next article.
Enterprise AI becomes governable neither when one understands it, nor when one counts it by the token. It becomes governable when a call ceases to be a consumption and becomes a commitment for which someone answers.
Full argument, with all references (Coase, Williamson, Searle, Knight, Hayek, Simon, Espeland and Stevens, Akerlof) and the complete apparatus, in the PDF below (8 pages).
Doctrinal notes and explorations on AI in regulated systems. Once or twice a month. One-click unsubscribe.