From Buzzwords to the Bottom Line
For years, artificial intelligence was the domain of innovation labs and pilot projects, funded by experimental budgets. Now, AI is rapidly moving into core business functions like customer service, marketing, and software development. As it scales, so does
the cost. A recent Deloitte survey found that 93% of organizations now use AI in their key operations. This transition from experiment to essential infrastructure means finance leaders are shifting their focus from potential to performance, and they're discovering that AI doesn't come with a simple price tag. Instead of predictable monthly subscriptions, they face a complex, consumption-based model that can be difficult to forecast and control.
Decoding the 'Token': AI's Currency
At the heart of this new cost structure is the 'token.' A token is the basic unit of data that a large language model (LLM) processes. Think of it as a piece of a word; for English text, one token is roughly three-quarters of a word or about four characters. An entire sentence is broken down into these tokens before an AI model can 'read' it, and the model generates its response one token at a time. This matters immensely because AI service providers like OpenAI, Google, and Anthropic price their services based on the number of tokens processed. Every query sent to the model (input tokens) and every word it generates back (output tokens) adds to a running tab.
The Meter Is Always Running
The cost of a single token is minuscule, often priced per million tokens (MTok). However, these costs accumulate with astonishing speed. A recent analysis shows that what looks like a cheap productivity tool in a pilot can balloon into a six-figure monthly bill when scaled across thousands of employees. Compounding the issue, input and output tokens are often priced differently, with output tokens typically costing three to five times more because generating text is more computationally intensive than reading it. This variable, metered model is a radical departure from the fixed-seat licenses of traditional Software-as-a-Service (SaaS) products, making AI spend highly unpredictable and a significant concern for CFOs.
The CFO's New Headache: Unpredictable Spending
The primary challenge for finance leaders is cost uncertainty. Forecasting AI spending is notoriously difficult because usage can fluctuate wildly. A poorly designed workflow, a verbose chatbot, or an automated agent stuck in a loop can burn through a budget at machine speed. Unlike traditional IT spending, AI costs are driven by engineering behaviours and architectural decisions, not just headcount. This has led to a phenomenon where the per-token price of AI is falling, but enterprise AI bills keep climbing due to exploding consumption. This forces CFOs to balance intense pressure to deploy AI quickly against the very real risk of spiraling costs, with 59% of finance leaders citing this as their top challenge.
Taming the Token Tide: Strategies for Control
In response, a new discipline known as 'AI tokenomics' or 'FinOps for AI' is emerging, focused on managing these variable costs. The goal isn't to reduce AI usage, but to connect spending to business value. Key strategies include implementing smarter model selection, where simple tasks are routed to cheaper, less powerful models, reserving the expensive, cutting-edge AI for complex problems that require deep reasoning. Companies are also developing better prompt engineering practices to be more concise and caching common queries to avoid paying for the same answer repeatedly. Other technical controls include setting hard spending caps per request or per user and using AI gateways that enforce rate limits based on token consumption, not just the number of requests.
