Agentic AI Spend Needs an Outcome Ledger, Not a Bigger Token Budget

OpenAI's July 14 guidance for managing AI investments recommends five moves: improve visibility into usage and spend, evaluate efficiency by outcome ROI, govern advanced workflows before scaling, fund workflows that compound, and match capacity to proven demand. Primary source: OpenAI, “How to manage AI investments in the agentic era”. The hard part is the denominator. “This agent used $800” says little. “This workflow cost $14 per accepted reconciliation, including review and rework” can support a decision. Here is a one-page ledger I would require for an agent pilot. Define one accepted outcome Do not start with tokens, seats, or tasks launched. Define the business state that counts after review. workflow: vendor-invoice-reconciliation accepted_outcome: "invoice matched, exceptions reviewed, result posted" owner: finance-ops pilot_window_days: 21 minimum_sample: 100 invoices quality_gate: false_postings: 0 exception_recall: ">= 0.98" reviewer_minutes_p50: "<= 3" Enter fullscreen mode Exit fullscreen mode A generated draft is not an outcome if a person must rebuild it. An agent run is not successful if its result never enters the system of record. Capture the complete cost AI cost + orchestration and observability + human review + rework + incident handling + allocated implementation cost = total workflow cost Enter fullscreen mode Exit fullscreen mode Use a table with declared variables: Variable Meaning Example only C_model model and tool-call spend $600 C_platform workflow infrastructure $200 H_review reviewer hours 35 R_hour loaded reviewer rate $45 C_build pilot build cost allocated to window $2,000 N_accept accepted outcomes 850 total = C_model + C_platform + H_review * R_hour + C_build cost_per_accepted_outcome = total / N_accept Enter fullscreen mode Exit fullscreen mode With the illustrative numbers, total cost is $4,375, or about $5.15 per accepted outcome. These are not benchmark claims; replace every value with measured data. Compare against the real baseline The baseline must use the same unit and quality gate: Metric Manual baseline Agent pilot attempted invoices 1,000 1,000 accepted outcomes 920 850 false postings 0 0 total cost $6,000 $4,375 cost per accepted outcome $6.52 $5.15 median cycle time 18 min 7 min A lower cost per attempt can hide lower completion. A faster median can hide an unacceptable tail. Report attempted, accepted, escalated, rejected, and incorrectly completed counts. Run sensitivity before buying capacity The largest uncertainty is often human review, not token price. break_even_review_minutes = (baseline_cost_per_outcome - non_review_agent_cost_per_outcome) / reviewer_cost_per_minute Enter fullscreen mode Exit fullscreen mode Calculate three scenarios: Scenario Acceptance Review minutes Decision downside 70% 7 stop expected 85% 3 continue pilot upside 93% 1 prepare controlled scale If the decision only works under the upside case, it is not ready for an annual commitment. Add governance as a costed requirement Advanced workflows need named controls before scale: controls: approval_owner: finance-ops-lead permission_scope: draft-only audit_retention_days: 90 rollback: disable posting credential incident_owner: platform-oncall review_expiry: "2026-08-31" Enter fullscreen mode Exit fullscreen mode A control with no owner or expiry is a sentence, not a control. Draft-only access may reduce automation upside, but it also limits the pilot's blast radius. Price both review labor and risk reduction honestly. Hard stop conditions Stop or redesign when any condition is met: one unauthorized irreversible action; the accepted-outcome denominator cannot be reconstructed; more than 10% of runs lack traceable cost; reviewers routinely redo the work outside the captured workflow; the downside scenario exceeds the baseline after two iterations; capacity is being purchased to solve demand that has not been measured. Scale only after quality passes, unit economics survive the expected and downside cases, and operational ownership exists. The most useful part of agentic investment guidance is not permission to spend more on agents. It is the shift from infrastructure consumption to workflow outcomes. A model can become cheaper while a workflow remains expensive; a costly model can be economical if it reliably removes a constrained, high-value step. What denominator does your AI dashboard currently omit: accepted outcomes, reviewer time, rework, or failures?