Marty Kausas, CEO of customer-support software company Pylon, recently highlighted an AI budgeting challenge that many enterprises are likely to face. Pylon's annual Anthropic bill, he noted, was projected to surge from roughly $400,000 to $1.4 million once the company surpassed 150 seats. The increase was not driven by a sudden spike in usage. Instead, it stemmed from a structural pricing change: above 150 seats, Pylon would transition from a plan that bundled usage to an enterprise plan where tokens were billed separately at standard API rates.

Kausas was blunt in his assessment: after years of urging employees to embrace AI tools, Pylon had started imposing spending caps and requiring approval for additional consumption.

Pylon's situation illustrates a wider transformation underway. Many companies initially adopted AI under highly favorable terms — bundled usage, introductory pricing, generous enterprise discounts, and relatively limited deployment. As those arrangements lapse and pilot projects evolve into production systems, the true cost of enterprise AI is coming into focus.

Simultaneously, the underlying technology itself is becoming less expensive. Inference costs have dropped significantly, competition among model providers remains fierce, and organizations now have more options to route tasks to smaller, more efficient models. Major providers including OpenAI, Anthropic, and Google have repeatedly reduced API prices for their flagship models over the past two years, while open-source alternatives from Meta and others have added further downward pressure.

The paradoxical result is that AI can become cheaper to use in isolation while growing considerably more expensive to operate across an organization. As models grow more capable, companies assign them more work, extend them to more employees, and embed them into more products. The resulting increases in volume and complexity can easily outpace any savings from lower per-unit prices.

This phenomenon is not unique to AI. Computing grew cheaper, and organizations consumed vastly more of it. Storage prices declined, and companies retained far more data. Bandwidth costs fell, and video came to dominate internet traffic. Efficiency consistently expands markets — a pattern economists have observed since the nineteenth century, when more efficient coal-burning engines paradoxically increased overall coal consumption rather than reducing it. AI is poised to follow the same trajectory — but with a critical distinction: much of its consumption will be difficult for executives to observe.

Traditional enterprise software is typically priced around a visible unit: a seat, a transaction, or a customer account. Agentic AI systems, by contrast, can generate costs autonomously and unevenly. Two employees holding the same license may consume radically different amounts of compute. A system may grow more capable while simultaneously requiring longer contexts, additional reasoning steps, and more calls to external tools. Unlike SaaS subscriptions, where cost scales linearly with headcount, token-based pricing introduces variability that most enterprise budgeting processes were not designed to handle.

This creates an entirely new management challenge. Enterprises can no longer equate purchasing access to AI with controlling its cost.

During the first phase of enterprise AI adoption, this issue was easy to overlook. Most organizations were running experiments and distributing a limited number of licenses. These initiatives revealed whether employees would use AI tools, but they did not address the question that now matters most: whether that usage generates sufficient economic value.

As AI budgets transition from experimental to material, enterprises must replace adoption metrics with unit economics. For a customer-support system, the relevant measures might include cost per successfully resolved case, resolution time, and escalation rate. For Smartling, an AI-powered translation company, meaningful metrics could include the cost of producing content at an agreed quality level, the time needed to enter a new market, or the volume of content a company can economically make available in each language.

These metrics are inherently more difficult to calculate than simple utilization figures. They demand that companies define the desired outcome before deploying a technology, establish a baseline, and account for both the cost of the technology and the human labor that remains around it.

Over the past several years, the process of building Smartling around AI has revealed that the central challenge was never introducing AI into the organization or persuading people to use it. The real challenge was identifying where AI materially altered the economics of the work.

A cheaper translation, for instance, offers limited value if the human correction it requires eliminates the savings. The relevant metric is not the number of words that pass through an AI system, but the cost and business value of producing usable multilingual content.

The same principle extends across the enterprise. Every AI deployment should begin with a clear economic hypothesis: which cost will decline, which constraint will be lifted, or which revenue stream will grow? Each deployment should then be evaluated against that hypothesis — not against the raw volume of AI consumed.

The next phase of enterprise AI, therefore, will not be defined by which companies achieve the highest adoption rates. It will be defined by which companies genuinely understand the economics of what they have adopted.

Organizations that can connect AI spending to revenue, margin, capacity, or strategic advantage will continue to invest — even as their absolute expenditure rises.

AI does not need to become inexpensive to justify its place in the enterprise. It needs to become economically legible. The companies that achieve this will not necessarily spend the least on AI, but they will know precisely what each additional dollar is buying.

The opinions expressed in Fortune.com commentary pieces are solely the views of their authors and do not necessarily reflect the opinions and beliefs of Fortune.

This story was originally featured on Fortune.com.