Enterprise teams cannot measure ROI on AI agent spend because token costs are not attributed to business outcomes

Enterprise AI agent spend is now a trillions-of-rows-per-month data problem with no shared measurement standard. Uber burned its entire 2026 AI budget by April (HN #47976415, 347 points). A TechCrunch investigation (June 5, 2026) documented CTOs paralyzed: one engineer spent $40k in a month and the CTO could not tell if he should stop him or replicate him across the team. 96% of enterprises report AI costs exceeding estimates; 40% of agentic projects fail due to hidden costs. The specific unsolved problem: token costs in agentic pipelines are not attributed to business task outcomes (PR merged, ticket closed, customer replied). Current tools (Datadog, Braintrust, AgentOps, Pay-i) give per-token observability traces but cannot answer the question 'what did this $40k buy us in delivered business outcomes?' No product maps agent token spend to task completion rates, compares cost-per-outcome across model configurations, and surfaces which agent prompts are burning budget without shipping work. AgentBudget (105 stars) enforces circuit breakers but does not do per-outcome attribution. Claude-code-router does cost routing but has no business-outcome tracking. The Linux Foundation Tokenomics Foundation was announced but its first deliverable is months out.