Why cost per execution
will replace thelicense per seat

The seat is a unit of measurement inherited from office software. We counted users because we only knew how to count that. With AI, this unit becomes absurd: it charges for presence, not for production. Cost per execution corrects this error. It measures what was actually done — and what it actually cost.

01 — The observationLicense per seat measures the wrong thing

A per-seat license is based on an assumption: the value is proportional to the number of users. It was defensible for a word processor. This is no longer true for an AI engine, whose value lies in the volume and quality of decisions produced, not in the number of active badges.

The result is known to all financial management. We pay for underused seats. We pay for capacity regardless of usage. And we discover, on the invoice, that the expense says nothing about what was produced. How much did Tuesday's refereeing score cost? Which ritual consumes the most? License per seat is unable to respond. It aggregates everything into an opaque package.

This opacity has a name in our models: the gap between paid capacity and executed capacity. It is immobilized capital that produces nothing. The bigger the bill gets, the more this gap becomes a budgetary blind spot.

What the seat makes impossible

Let us illustrate with simple reasoning. A management acquires two hundred seats from an AI engine. The package is comfortable, negotiated by volume. But in practice, only thirty people produce most of the decisions, and one of them, on a critical ritual, alone produces more than a hundred colleagues combined. What does the per seat bill say about this reality? Nothing. It distributes a uniform cost over two hundred badges, while the value – and consumption – are radically concentrated.

This inability to link cost to use has consequences beyond the budget. Impossible to compare two rituals on merit. Impossible to know if a ritual costs more than it brings in. Impossible to arbitrate finely, since everything is drowned in a package. The seat doesn't just measure the wrong thing: it prohibits the right questions.

02 — The mechanismCost per Run, or actual expenditure

Let’s pose the taxonomy. Ainferenceis the atomic unit: a submission to an engine. Arunis a coherent set of inferences — the complete execution of a ritual, attached to a single Run Receipt. THECost per Runis the exact cost of each of these executions, measurable in real time.

The consequence is immediate. Expenditure becomes a function of usage, not a fixed level. A little-used ritual costs little. A critical ritual, cast a thousand times, costs a thousand times its run — a cost known in advance, line by line. The IT department no longer manages a global subscription: it manages a cost structure that it can break down by ritual, by domain, by direction.

A taxonomy that makes cost understandable

To manage a cost, you must first know at what level you read it. Four rungs fit together. L'inferenceis the atomic unit — a demand on the motor. THErunbrings together the inferences of a complete ritual execution, attached to a single Run Receipt. THEprojectbrings together the runs of the same business objective over a period. THEjob, finally, designates all the projects in a governed domain.

This interlocking is not a coquetry of vocabulary. It allows you to read the cost at the scale where the decision is made. An engineer optimizes at the inference level; a ritual manager thinks at the run level; a business director who arbitrates at the project level; the executive committee pilots at job level. The license per seat only offers one level of reading – the global package – which does not correspond to any of these decisions. The Cost per Run, because it is rooted in inference, is recomposed at all the higher levels.

This move responds to a common objection: a variable cost would be less predictable than a flat rate. It's the opposite. A flat rate masks variability; he does not delete it. Cost per Run makes it visible and therefore governable. We can cap, arbitrate, optimize — precisely because we measure.

03 — The proofEach cost is attached to its run

Measuring a cost is not enough: you must be able to relate it to what caused it. This is where native proof makes the difference. The same Run Receipt that certifies the decision bears the cost. The cost is not an a posteriori estimate: it is native data from the run.

This cost traceability makes possible FinOps reasoning that per-seat licensing prohibits. We no longer ask “how much does AI cost?” » but “which ritual produces the most value per euro spent? ". The question finally becomes one of capital allocation, not bill reduction.

This shift has a discreet but profound organizational effect. As long as AI is a package, its cost is the responsibility of a single function — the one signing the contract. As soon as the cost is assigned to the run, it becomes the business of each management, which sees what its own rituals consume and produce. Financial responsibility is distributed where usage decisions are made. The cost is no longer incurred centrally: it is arbitrated as close as possible to the business that generates it.

04 — The profitA cost that becomes a lever

The move to Cost per Run is not just about savings. It is a change in the nature of the cost: from an incurred charge, it becomes a controlled lever. Each function of the executive committee finds its place there.

• For Finance.The AI ​​expense becomes predictable by the gesture and attributable to the cost center that generates it. The budget ceases to be a negotiated package to become a manageable structure, ritual by ritual.
• For the IT department.Cost per Run powers the Cockpit: costs, access and algorithms governed from a single interface. The fragmentation of subscriptions leaves room for a clear cost structure.
• For the Job.The cost of a ritual becomes an arbitration argument. We know what a certified decision costs — and we can decide whether it is worth the price, which the flat-rate seat prohibited.
• For Compliance.The assigned cost is also a trace. Knowing who executed what, at what cost, is part of the same evidentiary record as the decision itself.

05 — The objection“Isn’t a variable cost unmanageable? »

The financial management raises a legitimate objection: a package, at least, is known in advance. Doesn't a cost that follows usage risk slipping, surprising, escaping budgetary control? The objection deserves more than reassurance: it deserves a demonstration.

First, a package is only predictable in appearance. It sets the expense, not the value. We know what we are paying for; we don't know what we get. This “predictability” is only the organized ignorance of real variability, which nevertheless exists — it is simply invisible. Cost per Run does not create variability: it reveals it, and this is precisely what makes it governable.

Then, a cost measured per run is controlled with instruments that a package prohibits. We can set ceilings per ritual, trigger an alert beyond a threshold, compare the cost of a run to the value it produces. THEValidated Utility Rate— the share of runs approved without retakes — becomes a direct lever: a ritual whose runs are often retaken costs a lot of money for nothing, and we know it immediately. Variability ceases to be an experienced risk to become a regulated parameter.

Finally, the experience of the Virtuous Circle responds to the fear of getting carried away. The unit cost of a governed ritual does not increase with use: it decreases as the Vault becomes more dense. The natural trajectory of a well-managed Cost per Run is not drift — it’s decline.

06 — The trajectoryFrom seat to gesture

The software market has already experienced this shift. We left the perpetual license for the subscription, because the subscription suited the usage better. AI takes logic a step further: from headquarters to run, from presence to production. This is not a pricing fad. It is the alignment, finally, of price with value.

Each step of this story followed the same logic: bringing the invoiced unit closer to the value unit. The perpetual license charged one possession; subscription, access; the seat, a presence. None of these units measured production. Cost per Run takes this last step. It charges neither for possession, nor for access, nor for presence: it charges for the gesture that creates value. It is the culmination of a trajectory, not an isolated rupture.

The shift cannot be decreed: it is measured. The starting point is an existing ritual, from which we begin to measure the actual Cost per Run. The comparison with the fraction of license it consumes is, most often, enough to make the conversation obvious. The gesture is less expensive than the seat – and infinitely more controllable.

License per seat measured who could work. Cost per Run measures what was produced. Between the two, there is not a discount. There is a change in accounting times.