ai cornerstone-consultants working on tablet

An AI token is a small unit of text processed by a foundation model, roughly 750 words for every 1,000 tokens. Vendors typically bill in increments of a million tokens, with separate rates for input (the data and context a model reads) and output (the content it generates). Output almost always costs more, because generating new text takes more compute than simply reading it.

This distinction matters more than it sounds like it should. In the early days of generative AI, a prompt and a reply only cost a few hundred tokens or fractions of a cent. Agentic AI behaves more like a running meter. An agent takes a high-level goal, retrieves external data (input tokens), reasons through the problem and generates a plan through a process known as inference, which results in billable output tokens, hits a dead end, loops back, and tries again. A single token is cheap. The compounding, recursive nature of this loop is not, and it can process millions of tokens in minutes without anyone watching it happen.

That's the mechanic behind the AI token economy, and it’s what we at CGI treat as a core discipline of agentic AI cost management. In this article, we examine why AI is no longer just an IT cost. It requires an operating-model change that touches architecture, finance, procurement, governance and culture. 

Here's what the AI token economy means for enterprise leaders in finance, procurement and IT.

Who's actually driving consumption?

AI usage is no longer concentrated in IT or engineering. Every function can now initiate complex AI workflows, often with limited visibility into what's happening behind the scenes, and as adoption spreads across finance, HR, operations and customer service, this creates what we describe as agentic sprawl.

In practice, that sprawl tends to take three forms:

  1. Shadow agents built and run outside any central inventory. 
  2. Duplicate agents solving the same problem in different departments. 
  3. Unbounded agent loops that retry, replan, or use tools more often than a task actually requires.

Some companies have actively encouraged this behavior. Meta employees reportedly built an internal leaderboard, nicknamed “Claudeonomics,” to track who consumed the most tokens, while Amazon openly pushed staff to maximize usage. In both cases, consumption became a badge of honor before it became a line item anyone could see.

At the same time, agentic AI spending is becoming decentralized in a way it never was under the old licensing model. An employee's cost to the business used to be predictable, capped by a fixed software seat regardless of how heavily they used it. In the token economy, the cost of a single workflow depends on its complexity, the model selected and how many reasoning steps it takes to finish. This means that technology spending is no longer shaped solely by centralized procurement, but by thousands of operational choices made across the business every day.

That's not a reason to slow adoption. It's the reason the three functions below need to be looked at now, not once the bill arrives.

AI cornerstone, patterns

Finance: Approving a number nobody can pin down

For a CFO, AI token consumption behaves less like a fixed IT overhead and more like a volatile supply-chain cost. The immediate problem is business case approval: How does a finance committee sign off on a project when the underlying cost estimate is, by nature, a range rather than a number?

The math makes this harder over time. Gartner expects the cost of running a frontier-scale model to fall by nearly 90% between 2025 and 2030, yet still forecasts higher enterprise AI spending because agentic workloads are growing faster than unit prices are falling. Goldman Sachs similarly projects a 24-fold increase in global token consumption by 2030 as agent adoption spreads. Based on Jevons Paradox, falling prices and rising bills are not a contradiction.

The practical answer is that the definition of a “safe” business case has to change. A single static ROI calculation doesn't hold up when usage is this variable. Finance teams need wider estimation ranges going in and continuous tracking of value against spend once a workflow is live, not just a pre-launch estimate. If a workflow costs $10,000 a day in tokens but saves $50,000 in operational overhead, that's a good trade, but only if finance can actually see both sides of it on a dashboard.

According to Fortune magazine, Uber is a clear example of what happens when that link is missing. The company reportedly burned through its entire 2026 budget for AI coding tools in four months, after encouraging adoption through internal usage leaderboards, and even after that spend, its COO noted that the connection between rising token costs and features actually shipped to customers “is not there yet.” If a company that far ahead on AI adoption can't draw that line yet, it's a fair bet most enterprises can't either.

AI cornerstone, patterns

Procurement and operations: Contracting for a moving target

Procurement teams are built around negotiating fixed-seat licenses or volume discounts. Token-based pricing breaks that model. COOs and CPOs are now negotiating hybrid, usage-based contracts where pricing itself can shift through cache-hit discounts, rate limits and repricing clauses that didn't exist in traditional software contracts.

This isn't a hypothetical scenario, either. Several leading technology companies have publicly discussed the challenges of managing rapidly growing AI consumption as enterprise AI adoption accelerates. This reflects a broader shift in how organizations are adapting procurement practices, pricing models and governance as usage-based AI becomes mainstream.

AI cornerstone, patterns

IT: Governing AI

CIOs sit between a business pushing for faster AI adoption and a CFO pushing for cost control. It's unrealistic to expect every developer, let alone every business user, to know which model is the cheapest adequate choice for a given task. Left to guess, people tend to default to the most capable and expensive model available.

That's why the more durable fix isn't training everyone. It’s architecting governance. Semantic routing gateways that sit between the user and the model intercept prompts, assess their complexity, and route simple tasks to lightweight models while reserving frontier models for genuinely complex reasoning. They’re becoming as central to AI governance as the models themselves.

The piece that doesn't show up on a dashboard: Culture

None of this works without one more thing, and it's the one enterprises tend to skip: knowledge doesn't travel on its own. If an operations team finds a prompt structure that cuts token consumption by 40%, or an engineer builds a more efficient routing path, that insight is only worth something if it reaches the rest of the business before the next budget cycle. Training built around a one-time workshop doesn't hold up when models and best practices shift monthly. Treating token efficiency as a shared, constantly updated practice, rather than a one-off certification, is what actually makes the estimate, manage and control framework durable over time.

We've run this experiment on ourselves

This isn't just advice we give clients. As enterprise AI adoption expanded across CGI, our own token consumption increased as advanced coding assistants and agentic capabilities became part of everyday work. Rather than slowing adoption, we applied the same FinOps-inspired approach we recommend to clients: 

  • Optimize access to AI capabilities
  • Introduce usage tiers
  • Coach high-consumption users
  • Monitor business value alongside token spend 

That experience is helping shape our approach to AI governance.

We work with executives across industries to establish the methodologies and guardrails needed to manage AI costs effectively.

Contact us to learn more

 

Sources: This article draws on the following report and articles.