What now holds patrimonial value is not the software but the reproducible capability to operate it.
For thirty years the strategic asset was technology, then data. Autonomous systems force a third shift. What now carries patrimonial value is neither the model nor the license, but the reproducible institutional capability to operate an automated decision over time, at a controlled cost, and to answer for it. The model has become an input. The capability to run it under an auditable regime has become the asset.
The engineering claim that opened this cycle was narrow: a model that scores on a benchmark is not, for that reason, operable. It held on architectural grounds. What changed in 2026 is not the claim but its audience. A new reader now sits on the executive committee, and that reader does not read accuracy. The chief executive, the general counsel and the chief risk officer no longer ask whether the system works in a controlled setting. They ask whether the organization can sustainably bear responsibility for the decisions it automates, and at what recurring cost. Deployability, understood as the bare technical capacity to reach production, no longer names the problem. The lock has stopped being technological. It has become institutional.
The cycle rests on three notions that look like synonyms and are not. Their order carries the argument. The sustainable operating capability is an emergent property of the organization: its ability to absorb, steer and audit the risk of an automated system over time, at a cost it controls. The institutional regime is the set of rules, formal and informal in the sense of institutional economics, that make this capability possible: lines of accountability, arbitration procedures, evidentiary obligations. The auditability regime is the embodiment of those rules inside a given system, the concrete mechanisms that keep it controllable and accountable at every moment. A property, the rules that ground it, their implementation. The order is not decorative: it settles, at the end, which of the three deserves the name of asset.
The market long read the cost of artificial intelligence as a cost of acquisition: buy a license, call an API, and let the capability curve absorb the rest. Confusing the cost of the model with the cost of the regime around it was the founding error of the first deployments. The acquisition cost is the model’s, paid once. The operating cost is the auditability regime’s, paid every month: continuous monitoring, control-data engineering, human arbitration, drift management, legal liability. What has been commoditized is not delivery in general (integrating an SAP core, a core banking platform or a hospital information system remains a difficult art) but a precise segment: access to models, baseline integration, standard deployment. The delivery of the model is now a commodity. The delivery of the regime is not.
A counter-intuitive law follows, one that leadership is discovering this year: operating cost does not fall when the model improves. It often rises, because a model judged more capable is immediately pushed onto more critical processes, which widens the impact surface of any anomaly. Value has not changed source, it stays co-produced by model, data, processes and people. What moves is the bottleneck, and rent follows the bottleneck. Yesterday the bottleneck was raw model capability, and rent accrued to the frontier labs. Today generic capability is commoditizing, inference prices are collapsing, and the binding constraint becomes the organization’s inability to absorb, steer and audit operational risk. Capture migrates toward mastery of the institutional regime, and the field engineers deployed en masse by the Big Four are only the provisional embodiment of that transfer.
Sustainable operating capability is not a compliance layer added by external auditors. It lives in the technical architecture or it does not exist. To be institutionally operable, a highly automated system must answer to a strict grammar of auditability built on three invariants: a rigorous split of data flows, an explicit calibration of uncertainties, a dynamically controlled applicability domain. Validation then stops being a stamp obtained once at release. PREDICARE offers an instance of this, not a general proof: in connected health and triage, it treats the applicability domain as an active architectural object rather than a passive specification, re-measuring its own confidence boundaries in real time and suspending execution as soon as the context drifts outside its safety envelope. Auditability becomes a systemic property, not a frozen document.
Before a board, management no longer answers for having bought the most advanced model on the market. It answers for the regime under which it operates that model, and for its ability to suspend it without breaking business continuity. The benchmark is still purchased by IT, but value capture no longer happens at the model’s level. It happens at the level of the organization able to answer for it.
Name it precisely. It is not the auditability regime, which is only an implementation. It is not the institutional regime, which is only a set of rules. It is the sustainable operating capability itself, understood as the reproducible institutional capability to operate autonomous systems over time. The qualification is not rhetorical. A strategic asset shows five properties: it accumulates, it resists imitation, it transfers only by degrading, it generates future flows, and it is valued independently of its substrate. Operating capability meets all five. The model has lost them one by one, having become interchangeable, imitable, transferable, commoditized. The substrate is now an input. The capability to operate it is the asset.
The thesis is falsifiable. At comparable models, two organizations should show durable gaps in economic performance correlated with the quality of their auditability regime, not with the intrinsic quality of the models they run; and once models become interchangeable, valuation gaps should be explained more by institutional capability than by algorithmic assets. If durable gaps tracked the models rather than the regimes, the thesis would be false.
One question stays open. If architectural governance, regulatory compliance and financial risk management obey the same grammar of auditability, how will organizations restructure their departments to unify these competing frameworks? That is next week’s subject.
[Series: The Hidden Costs of Deploying AI / Sunday synthesis, W11/12]
Doctrinal notes and explorations on AI in regulated systems. Once or twice a month. One-click unsubscribe.