The Unit Economics of Intelligence

In a fifteen-day window in mid-2026, the major platforms moved agent workloads onto consumption meters — per-token, per-workflow, billed to the organization rather than the user [3]. The accounting consequence is structural: AI is now a variable cost coupled to usage and outcome, and a seat-based budget cannot govern a per-token bill. The question is no longer “how many licenses?” but “what does each unit of work cost, and what does it produce?”

For an agentic workload, the cost of one successful outcome is approximately:

Every term is an operating lever. Tokens and price fall with model choice and prompt discipline; steps fall with better decomposition; the retry term and the success-rate denominator are governed by reliability engineering. The denominator is the one most budgets ignore — and the most punishing.

Because a failed run still consumes tokens, cost per successful outcome scales as 1 / (success rate). A workload that completes 50% of the time costs roughly twice per delivered result; at 33%, three times — before counting the human rework a failure triggers. This couples directly to the compounding-error curve of the engineering paper [1]: low per-step reliability does not merely reduce output, it inflates the unit cost of every output that does land.

Not every step needs the frontier model. The largest, most defensible cost reduction comes from routing each step to the cheapest model that clears its reliability bar — reserving the expensive tier for the judgment that requires it. Vendors already do a version of this inside their own bills; the operator who does it deliberately, measured against per-step success, captures the margin instead of paying for it. Routing is a unit-economics decision before it is an engineering one.

Margin discipline is enforced at the workload, not the seat. Each recurring agent workload gets a budget line, a named owner, a target cost-per-successful-outcome SLO, and the business number it exists to move; a workload that cannot name its number does not get a meter [3]. This converts an opaque, growing platform line into a set of governed units, each answerable for its own economics — the precondition for AI that is profitable rather than merely impressive.

For the CFO, the discipline is concrete: carve the token line out of the platform budget and project it on its own; track cost per successful outcome, not gross spend; and hold a gross-margin target on AI-delivered work the way the business holds margin on anything else. Falling per-token prices are not the relief they appear — usage is projected to grow faster than prices fall [2], so the bill rises while the rate drops, and only unit discipline keeps margin intact. The firm that meters by outcome scales profitably; the firm that meters by hope finds out in April.

The unit-cost expression is a planning model, not a billing reconciliation; real bills include caching, tool costs, and vendor-side model substitution that the customer does not control [2]. Success rate must be measured against honest acceptance criteria, or the denominator flatters itself. And token accounting at production scale is a trillions-of-rows data problem in its own right — the measurement system is itself an investment.

Intelligence is now a metered input, and inputs are governed by unit economics. Model cost per outcome, route to the cheapest model that holds, govern by workload, and protect margin deliberately — or watch a falling price produce a rising, ungoverned bill. AI strategy ends where the bill begins; unit economics is how the bill is read.