Loading…
One denominator. Five numerators. A single discount rate. A clear distinction between total ROC (is the program working?) and marginal ROC (where does the next dollar go?).
Discounted economic value created by AI compute over the discounted cost of committed compute. Same discount rate for both sides.
Value from AI systems already deployed. Counts contribution from incremental revenue, direct cost savings, retention improvements, and risk reduction. Subtracts vendor, implementation, and support costs.
Risk-adjusted future value from new AI products, agents, workflows, and model improvements. ROC forces you to subtract a counterfactual baseline so you only credit what wouldn't have happened anyway, and uses a ramped NPV so future-launch bets aren't credited as if they shipped today.
Saved hours are not automatically economic value. The capture rate estimates how much saved time becomes productive output. Most organizations land between 25% and 60% — the rest is reabsorbed by meetings, slack, or just earlier-end days.
Value from doing the same or better work with less compute. ROC prefers cost per useful output over cost per token, because tokens aren't the business outcome. Useful outputs are.
Value of having enough flexible capacity to serve upside demand without stranding cost on the downside. ROC produces a scenario-weighted estimate of missed value at current capacity, minus the cost of holding excess.
Total ROC tells you whether your AI program is working as a whole. Marginal ROC tells you where the next dollar of compute should go. The two often disagree, and the difference is where most allocation upside lives.