Analyst coverage is most useful when it does more than validate a product announcement, and instead shows you how the market is changing.
That’s what stood out to me in IDC’s recent analysis of Flexera’s expanded AI Cost Management capabilities. IDC’s observation was bigger than Flexera: FinOps for AI has already outgrown token tracking.
That may sound obvious to anyone watching AI adoption accelerate, but it represents an important shift in how enterprises should think about AI economics.
AI doesn’t create one new cost. It creates a new cost stack.
For the first phase of generative AI adoption, much of the cost conversation centered on models and tokens. How many tokens are we using? What does each model cost? How do we monitor API consumption? Don’t get me wrong – those questions still matter. But there’s more we should be asking.
IDC points to GPUs, data platforms, APIs, networks and token-based charges as parts of the growing AI cost equation. Its recommendation is consequently broader: organizations need a platform approach that brings FinOps across the full technology stack rather than managing each layer separately.
An application consuming an AI model doesn’t exist in isolation. There may be cloud infrastructure underneath it, data moving through Snowflake or Databricks, model API charges sitting on top, agents performing tasks, and SaaS subscriptions supporting the teams deploying it.
If those costs live in different dashboards, budgets and ownership structures, nobody has a reliable view of what AI actually costs.
That’s why I think customers are beginning to expect AI cost management to deliver economic visibility, not simply token visibility.
The urgency is already showing up in the data
The numbers IDC cites are interesting:
- According to IDC research, 90.5% of enterprises were already over their 2026 token budgets by August
- Only 9% could proactively review AI costs weekly or more frequently
(Source: IDC’s September 2026 Intelligent Cloud and Edge Operations with FinOps Survey of 254 respondents)
AI adoption is moving at a speed that traditional cost-management processes were not designed to handle. A monthly report can tell you where money went. It cannot necessarily prevent an agent, model or workload from dramatically changing the economics of an application between reporting cycles.
IDC puts this particularly plainly: autonomous optimization, with the appropriate guardrails, is the right direction because agents capable of acting on recommendations can move considerably faster than static monthly reviews.
That shifts the question from “Can I see the cost?” to “Can I do something about it quickly enough?”
For enterprise technology leaders, that is a much higher bar.
Why agent sprawl is a concern
Perhaps the most interesting part of IDC’s analysis is its warning about agent sprawl.
IDC compares what’s happening with AI agents to previous eras of technology expansion, when virtual machines and unused cloud resources proliferated faster than organizations could govern them. IDC found that enterprises deployed 28.8 million agents in 2025, and those same enterprises expect to manage roughly 80 times that number by the end of 2026. When looking ahead, daily agent actions could approach 217 billion by 2029.
We saw it with SaaS, and we saw it with cloud. But AI could amplify the pattern because increasingly autonomous systems aren’t simply resources that sit idle. They can continually consume models, data and infrastructure as they perform work.
Cost management therefore cannot be something organizations bolt onto AI after deployment.
IDC recommends putting FinOps into every AI project before final project approval, forecasting costs before funding and requiring an AI total-cost-of-ownership model for pilots. It also recommends inventorying agents and establishing standards around technologies such as A2A and MCP.
The next phase of FinOps is about unit economics
There was another recommendation in IDC’s analysis that I think deserves more attention: show the CFO unit economics, such as cost per transaction.
Because ultimately, the goal isn’t simply to make AI cheaper: for instance, an AI application that costs $10 million and produces $50 million of measurable value may be an excellent investment. An AI pilot costing $100,000 that produces nothing useful is expensive.
Enterprises therefore need to move beyond questions such as “How much are we spending on AI?”
They need to ask:
- What does this AI service cost per transaction, customer interaction or business outcome?
- Which models deliver the right balance of cost and performance?
- Where are infrastructure and data costs accumulating?
- Which agents are consuming resources?
- Who owns those costs?
- And can teams identify and correct inefficient spending before the invoice arrives?
That is a fundamentally different discipline from monitoring a token bill.
What I think IDC gets right
The most significant takeaway from this report is that IDC frames the underlying problem as full-stack AI economics.
The firm recommends bringing FinOps into AI projects early, managing infrastructure alongside tokens, improving cost allocation, monitoring token costs frequently, experimenting with agent-assisted remediation and developing AI-specific total-cost-of-ownership models.
Flexera has been expanding its own approach to AICM accordingly, bringing together Token Cost Control and Cloud Cost Optimization with cost controls spanning AI cloud infrastructure and AI data-cloud platforms. IDC also notes Flexera’s strategic partnerships with AI gateways, observability, model routing and economic control systems, as well as the introduction of an AI Cost Management maturity assessment.
In my opinion, the best indication of maturity is whether you understand the economics well enough to turn all of the investment into business value and outcomes.
