Share with your CFO
ManageEngine CEO Rajesh Ganesan is making a direct case that enterprise AI budgets are breaking not because models are expensive but because companies are routing every workload to the most powerful, costliest option available. The supporting numbers are stark: agentic workflows, which chain multiple model calls to complete a task, burn five to thirty times more tokens than a single query, and EY estimates one customer-service interaction that cost four cents in 2023 now runs $1.20. An analysis of 2.4 billion enterprise API calls found tiered model routing cut per-token costs from $18.40 to $2.31, an eight-fold gap.
What this means for your business
Budget blowouts at Uber and Microsoft are the clearest signal that AI cost overruns aren’t a governance failure at careless companies; they’re the structural result of applying frontier-model pricing to commodity tasks. The CFO most exposed here is the one who approved a flat AI budget in 2024 and is now watching agentic adoption accelerate inside engineering or customer service without a corresponding mechanism to control token consumption. Goldman Sachs projects total token volume will multiply roughly 24 times by 2030, so the exposure compounds every quarter you don’t have a routing discipline in place.
The cost structure of agentic AI has a specific shape worth understanding. Token price per unit is falling fast, roughly 98% since early 2024 by Ramp’s data, but volume is rising faster, driven by every agent that calls a model multiple times to reason through a task. The companies that avoid the blowout aren’t negotiating better rates; they’re separating workloads by capability requirement and routing summarization and classification tasks, which represent the bulk of enterprise volume, to cheaper models while reserving frontier capability for genuinely hard problems. That’s a routing and governance architecture decision, not a procurement one, and it has to be built before the volume arrives, not after the invoice does.
The FinOps function, which spent the last decade rightsizing cloud infrastructure costs, is now being handed AI spend with no established playbook: the share of FinOps practitioners responsible for AI budgets jumped from 31% to 98% between 2025 and 2026. If that function is sitting inside your organization without tooling for model-level cost attribution, the current AI budget is essentially a black box. The falsification condition for Ganesan’s thesis is straightforward: if model prices fall fast enough over the next 18 months, undisciplined routing becomes cheap enough to tolerate. But nobody defending a budget renewal this year can plan around that hope.
Concept deep-dive: Model routing
Model routing is the practice of directing each AI task to the cheapest model capable of completing it adequately, rather than sending everything to the most powerful option. Think of it like freight shipping: you don’t use overnight air for every package. Most enterprise AI tasks involve classification or summarization, work a smaller, cheaper model handles well. Routing reserves frontier models for genuinely complex reasoning. The business case is an eight-fold cost reduction on identical output volume, which makes it the single highest-leverage cost control available to any organization running agents at scale.
Based on reporting from As Agentic AI costs break enterprise budgets, ManageEngine says the fix is spending smarter, originally published 2026-07-09 19:24:00.
