Gartner projects $207 billion in enterprise agentic AI spend in 2026. VentureBeat's investigation of corporate deployments found that most organisations cannot demonstrate concrete returns against that spend — a pattern being described as "tokenmaxxing," where token consumption grows rapidly without linking to measurable business outcomes. Uber deployed Claude Code company-wide in December 2025, exhausted its full 2026 AI coding budget by April, and subsequently imposed $1,500 monthly per-employee usage caps. The clearest ROI data point in the analysis comes from legal tech company Everlaw: a $27,000 token expenditure cut estimated engineering project time from 90–100 months to 19 months. Root causes for the broader measurement failure include premium models running at maximum reasoning effort by default, and employees reverting to familiar workflows regardless of cost settings. The prescribed remedy — LLM gateways with intelligent routing, task-appropriate model selection, and token costs treated as planned infrastructure rather than expensed software — requires engineering investment to implement. Organisations that have deployed agents without that measurement layer are running an expensive experiment with no feedback signal.