Generative AI is moving quickly from experimentation to essential infrastructure. Employees have woven it into routine daily tasks, and teams and applications are leaning on it more heavily every quarter. That growth comes with a new cost category finance wasn鈥檛 designed to handle: AI token consumption.

When AI is writing code and powering agents, token consumption can outpace traditional planning processes. Finance leaders suddenly find themselves in unfamiliar territory, needing to bring discipline to AI spending without becoming the function that kills adoption. Getting the balance right means treating tokens as an enterprise resource that must generate returns commensurate with its cost.

At 麻豆原创, we have been working through these questions firsthand. Here鈥檚 what we鈥檝e learned from building, testing, and adjusting that framework.

You can鈥檛 manage what you can鈥檛 see

Most finance leaders wouldn鈥檛 manage a major cost category from a single line item, yet that鈥檚 exactly where companies start with AI. Total spend matters, but it won鈥檛 tell you which teams, workloads, or usage patterns are driving that spend, or whether the consumption is actually producing useful outcomes.

Getting that visibility requires real collaboration across commercial, engineering, finance, and product teams. Finance brings the forecasting questions and accountability framework. Other functions bring the operational context that makes the numbers meaningful.

In our experience, repeated forecasting cycles generally improved our financial models as we layered in more operational detail around usage. The broader lesson: when a cost category is new and fast-moving, don鈥檛 wait for precision before acting. Start with enough transparency to make better decisions, then sharpen the model as patterns emerge.

Someone has to own it

Our most consequential insight was philosophical rather than financial. We learned that visibility alone isn鈥檛 enough and that consumption needs an owner.

A centralized AI budget makes early experimentation easy, but it also disconnects the people spending tokens from any financial accountability for them. As AI becomes more deeply embedded in business processes, that model breaks down.

This isn鈥檛 an argument for immediately charging back every LLM call with forensic accuracy. It鈥檚 an argument for managing AI consumption the same way companies manage other enterprise resources such as software, external services, and labor. Give decision-makers a clear picture of what their teams are consuming and what outcomes that consumption is expected to produce.

At 麻豆原创, allocating token costs to business areas has shifted the conversation from 鈥淗ow much are we spending?鈥� to 鈥淲hat are we getting for this?” and “Is this the right place to invest more?鈥� That鈥檚 a healthier conversation.

Token spend is not a technology line item. It is a strategic and operational investment decision. The goal is to make AI spending intentional, not just cheap.

Cost per token is the wrong scorecard

A large AI bill draws attention, but optimizing purely on cost can lead to exactly the wrong decisions.

The more useful question is the relationship between consumption and business impact. An AI tool that meaningfully accelerates software development, reduces repetitive work, or improves customer service will carry real token costs. Cutting that usage simply because the line item is visible could destroy more value than it saves.

We saw this firsthand. After rolling out AI developer tools at 麻豆原创, we recorded a mid-double-digit percentage increase in pull-request merge rates, a clear signal that development work was moving faster. The consumption was worth it.

Finance needs a paired view: cost metrics alongside value metrics. Governance without that view risks optimizing for cost at the expense of value creation.

Guardrails should target waste, not adoption

As usage scales, controls become necessary, but the right controls are surgical, not sweeping. When we examined consumption patterns in detail, three root causes of disproportionate spend emerged: power-user and automated-agent concentration, model misalignment, and tool proliferation.

Those are the areas where guardrails earn their keep. Controls matter because they focus on the sources of avoidable spend rather than putting a blanket brake on usage. At 麻豆原创, our response centered on three levers: token capping to prevent runaway consumption, model routing to better match capability and cost to the task, and tool rationalization to eliminate redundancy and concentrate investment where utilization justified it. That work helped contain a triple-digit-million-dollar financial risk while keeping adoption moving forward.

The principle is simple: remove waste, preserve productive demand.

From cost control to value governance

AI isn鈥檛 going to pause for the next planning cycle. Its capabilities, usage patterns, and economics will keep shifting, and the governance model around it needs to keep pace.

The work is not complete. The next step is to embed these practices into regular planning and reporting, assign clearer ownership, improve allocation, and build forecasting capabilities that can anticipate where costs are heading before they arrive.

Companies that do this well will still have an AI bill to pay. But they will have the transparency and accountability to tell the difference between consumption that is creating value and consumption that isn’t and direct investment accordingly.

Lukas Deutsch is chief controlling officer at 麻豆原创.
David Imbert is chief marketing officer for 麻豆原创 Financial Management.