Brian Wilson is CEO of Kion, whose FinOps+ platform unifies AI-powered governance and cost management for regulated enterprises and agencies
Powerful LLMs and AI agents are quickly driving AI usage costs through the roof. To combat it, organizations need to apply the same kind of discipline that FinOps brought to cloud infrastructure.
FinOps, which helped organizations get control of spiraling costs while maximizing the value of their cloud investments, now has a bigger job on its hands, courtesy of artificial intelligence.
Like in the early days of cloud, the emergence of large language models and agentic AI is radically changing the way enterprises operate, pushing organizations to find a way to understand and control their environment. Yet, AI is different, permeating organizations more pervasively than cloud did. Today, AI usage spans a wide variety of teams and departments throughout the enterprise, leading to extremely quick consumption of the new source of rising costs: tokens, and the infrastructure required to create, manage and govern them.
This sticker shock comes despite the fact that, on a per-usage basis, the operating costs of LLMs have fallen off a cliff in recent years. The Stanford HAI “AI Index 2025” found that the inference cost of a GPT-3.5-level capability per million tokens fell from $20 to 7 cents from 2022 to 2024. Yet enterprise spending on LLMs tripled in 2025, and “93 percent of respondents to a McKinsey survey reported exceeding their AI budgets.”
About 20% of organizations curtailed AI use because of escalating costs, particularly as AI projects scaled upward. The rest of the organizations are struggling to justify the spend with hard metrics. In my experience, the answer is not to abandon AI but to improve how organizations use and manage it. Organizations need a clear idea of the value being generated by AI, which starts with gaining clear visibility into AI spending at a granular level. That’s the foundation of a governance plan that enables organizations to justify their spending while deriving maximum value from AI investments.
How Frontier AI Costs Can Quickly Snowball
AI costs initially appeared as something of a surprise to many organizations. Token costs appear when an AI model requests and processes information. Input tokens, based on the instructions sent to the model, typically range from 5 or 10 cents or less to $5 per million tokens, or more for flagship models. Output tokens, which require more processing, run about three to five times more. It’s peanuts, at least on a small scale. But as AI models become more powerful, with agents initiating activities on their own and retrying queries without assistance, those costs quickly multiply.
I’m seeing organizations struggle more than ever to track and control token spend. Since AI is available to sales, marketing, finance, product, operations, engineering and executive teams, rather than just a relative few, AI costs are growing faster than teams can track.
If one thing is for certain, organizations need AI visibility to be more precise. Rather than weekly reports, organizations need real-time tracking and alerting to detect and identify anomalous AI activity before it gets out of hand.
How FinOps Can Help Control AI Spend
Like with cloud, FinOps can provide the discipline enterprises need to gain control over their AI spend. But implementing it does have challenges. Most significant is that, while nearly every team and individual has access to AI, most have no concept of how fast token costs can escalate or how AI agents can accelerate those charges.
I believe that for organizations to be successful in this endeavor, FinOps for AI must be an all-of-enterprise, culture-changing program, with visibility as the first step. Organizations need to know where and why AI is being used, as well as who is employing it. From there, an organization’s goal should be to aggregate AI use and spending into a singular view so teams can allocate spending at a granular level to answer the ROI questions executives are asking. This is a core objective for the FinOps Foundation as they delve deeper into Tokenomics and what it means for the modern enterprise.
This approach works best in environments where putting a ceiling on spending is critical, like AI sandboxes. An organization I spoke with recently had success with sandboxing an AI tool when they gave engineers a daily budget on how much they could spend. When they hit it, access for that day was shut down, allowing them to put a ceiling on how much each team was able to spend for each project.
Why It’s Not Just The Spend But Also The Value
It’s true that AI spend is only going to get harder to control, but I’ve seen great success within organizations that are proactive about educating their employees about AI use and how token costs work. This means educating your employees on what cost-efficient prompts look like, creating ways for employees to use simple language prompts without blowing the budget and making cost-efficient model choices. It also means implementing tracking tools that show your teams their daily, weekly and monthly token spend as they complete tasks, letting them see how their token use is trending toward budget.
When you can see and control token spend, you can create an environment where people can use AI tools without going too far.
Beyond controlling AI spend, the ultimate goal when employing AI is realizing value from its use. With FinOps policies in place, organizations can build the right guardrails, alerts, policies and awareness without slowing innovation. Token spend may be the new cloud bill shock, but with early visibility and governance, it does not have to become the next unmanaged enterprise cost problem.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?
