Agentic AI ROI: Track Cost Per Accepted Outcome, Not Tokens
A five-step framework for measuring agentic AI ROI through cost per accepted outcome instead of raw token or infrastructure spend.
OpenAI's guidance for managing AI investment in the agentic era lays out five moves: improve spend visibility, evaluate efficiency by outcome ROI, govern advanced workflows before scaling, fund compounding workflows, and match capacity to proven demand. The hard part is the denominator — saying an agent spent $800 tells you nothing, while a cost-per-accepted-outcome figure that includes review and rework actually supports a decision.
The proposed framework is a one-page outcome ledger: first define what counts as an accepted outcome (a draft a human must rebuild doesn't qualify), then roll model, platform, human review, rework, and allocated build costs into a single total to compute cost per accepted outcome. That number must be compared against a real manual baseline using the same unit and quality gate, since a lower cost-per-attempt can mask lower completion rates, and a faster median can hide an unacceptable tail.
Before buying capacity, teams should run a three-scenario sensitivity analysis — downside, expected, upside — because a decision that only holds up in the best case isn't ready for an annual commitment. Governance also needs to be treated as a costed requirement: a control without a named owner and expiry date is not a real control. For engineers, the takeaway is a shift from infrastructure consumption to workflow outcomes — a cheap model doesn't guarantee a cheap workflow, and an expensive model can still be economical if it reliably eliminates a costly, high-value bottleneck.