NewsMacroThe AI Token Economy: Why Companies Need AI FinOps Before the Bill Arrives

The AI Token Economy: Why Companies Need AI FinOps Before the Bill Arrives

Author: Bworldonline·

Key Takeaways

  • •Enterprise AI costs are shifting from fixed subscriptions to usage-based consumption measured in tokens, similar to how cloud computing billing evolved.
  • •The author predicts an 'AI FinOps' discipline will emerge to make AI spending visible, accountable and tied to business value, as cloud-era FinOps did.
  • •AI agents can turn a single employee request into many model transactions, making budgets based on user counts inadequate.
  • •The right metric for AI spending is cost per useful outcome, not cost per token, since the cheapest model is not always the best choice.
  • •Consumption controls such as caching, prompt shortening, selective document retrieval and AI gateway model routing are already supported by major AI and cloud providers.
The AI Token Economy: Why Companies Need AI FinOps Before the Bill Arrives

For the past two years, companies have been urging employees to adopt artificial intelligence. Soon, they will begin asking a very different question: who is paying for all these tokens?

The issue sounds trivial today because most employees encounter AI through a monthly subscription to ChatGPT, Claude or another tool. The cost appears predictable — buy licenses, assign them to staff and include the amount in the IT budget.

But that is not where enterprise AI is heading.

As companies embed AI into customer service, sales, marketing, software development, legal work, finance, analytics and internal processes, the technology becomes less like traditional software and more like a utility. It is consumed: the more you use, the more you pay. And the meter is increasingly measured in tokens.

A familiar story from the cloud era

The corporate world has been here before. More than a decade ago, companies rushed into cloud computing. Amazon Web Services, Microsoft Azure and Google Cloud offered exactly what corporate IT had long wanted: computing resources without buying servers or building data centers. Spin up a virtual machine, add storage, increase capacity, and pay only for what you use.

It sounded wonderfully efficient — until the bills arrived.

Developers created servers and forgot to shut them down. Test environments stayed alive for months. Storage accumulated. Departments purchased cloud services without central oversight. No one worried much when each decision involved only a few dollars, but thousands of such decisions combined eventually drew the CFO's attention.

That painful period gave rise to FinOps, a discipline built around making cloud spending visible, accountable and tied to business value — a practice now supported by an organized FinOps community and certification bodies. The author believes the same thing is about to happen with AI — call it AI FinOps.

From adoption metrics to consumption metrics

Today, companies are focused on AI adoption: How many employees use ChatGPT? How many AI projects have been launched? How many processes can be automated?

Those are useful questions, but another set is coming. How many tokens are we consuming? Which departments consume them? Which applications generate the most usage? Which models are being used? How much does an AI-assisted transaction cost? And, most importantly, what business result did the money deliver?

A token is simply a small unit of text processed by an AI model. Economically, however, companies should start treating tokens the way they learned to treat compute hours, storage and network traffic in the cloud: as units of consumption. Major model providers already publish usage-based pricing, typically quoted per million input and output tokens, which makes this kind of metered accounting technically possible today — the missing piece is usually visibility into who consumed what inside the organization.

Consider a company that processes 500,000 customer interactions a month using AI, with each interaction consuming an average of 4,000 input tokens and 800 output tokens. Management can estimate the monthly volume and multiply it by the applicable model prices. Now increase customer volume by 20%, add another AI application, give more employees access, and introduce an AI agent that performs several model calls to complete a single task. The budget can change quickly.

Why agents change the equation

The challenge becomes even more significant with AI agents. A person asks one question and receives one answer. An agent, by contrast, may receive one instruction, break it into several tasks, search documents, call different systems, check its work and invoke an AI model repeatedly. One employee request can quietly turn into many AI transactions.

That is why budgeting based simply on the number of AI users will eventually become inadequate. Companies need to measure tokens by application, department, workflow and business outcome. Finance needs forecasts. IT needs usage controls. Business units need accountability.

A marketing department running millions of personalized messages should know its AI cost. A contact center should know the AI cost per resolved case. Software teams should know the AI cost per development task. A bank using AI to review documents should know the cost per processed application.

Cost per useful outcome, not cost per token

AI FinOps should not become another corporate cost-cutting exercise, however. The cheapest model is not automatically the best model. If spending 10 pesos on AI saves 500 pesos of employee time, cutting the AI cost to five pesos may be meaningless. If a more expensive model produces fewer errors and resolves more customer problems, paying more may actually be the better decision.

The right metric is not cost per token. It is cost per useful outcome.

There are also obvious ways to control consumption. Companies do not need their most powerful model for every task — simple classification, extraction and summarization may work perfectly well with smaller models. Prompts can be shortened. Documents can be retrieved selectively instead of repeatedly sending hundreds of pages into a model. Long conversations can be summarized, repeated instructions and context can be cached, and applications can set token limits. AI gateways can route tasks to different models based on complexity and cost. Several of these levers — caching, prompt context reuse, model routing — are already supported by features and tools that major AI providers and cloud platforms have introduced for enterprise developers.

None of this sounds exciting compared with talk of autonomous agents or artificial general intelligence. But CFOs eventually make boring things important.

Cloud computing followed the same path: first excitement, then adoption, then uncontrolled spending, then governance. AI will probably follow it faster.

What comes next

It is easy to imagine companies soon deploying AI consumption dashboards showing token usage, cost per workflow, model utilization, budget variance and business value. Departments may receive AI budgets just as they receive travel, marketing or cloud budgets. And someone will eventually be held responsible for all of it.

That person may sit in finance, IT or an AI center of excellence. The title may differ, but the discipline will be the same: AI FinOps.

Companies should start building this capability now — before the AI bill grows large enough for the CFO to ask the question nobody wants to answer: where did all those tokens go?


The views expressed herein are his own and do not necessarily reflect the opinion of his office as well as FINEX.

Reynaldo C. Lugtu, Jr. is the founder and CEO of Hungry Workhorse, a digital, culture, and customer experience transformation consulting firm. He is a fellow at the US-based Institute for Digital Transformation. He is the Chair of the AI, Digital Transformation, and Governance program of the FINEX Academy. He teaches strategic management and digital transformation in the MBA Program of De La Salle University. He can be reached at rey.lugtu@hungryworkhorse.com.