Imagine checking out of what you thought was an all-inclusive resort, only to find a bill for every activity, snack, smoothie, towel, pillow case, shampoo sample or beach umbrella. No rate card was posted. No advance notice was given. And nobody can tell you what anything cost until after you already used it.
That is roughly where enterprise AI teams find themselves today.
This month, GitHub Copilot, OpenAI, and others completed their transition from flat-rate subscriptions to usage-based token billing. The all-inclusive era of AI is over. What replaced it is not a straightforward menu with visible prices. It is a mystery menu — everything has a cost, but nothing has a price tag. You cannot optimize against a price you cannot see. And “be careful” is not a strategy when no one can tell you what careful costs.
If this sounds disorienting, it is. But here is the harder truth: we were not on solid ground before the bill arrived.
The Foundation Was Already Shaky
Nine in ten employees are already using AI tools at work. Yet fewer than 40% of the organizations they work for have a formal, comprehensive AI policy — and one in four has no policy at all.
I work with large enterprises navigating this gap every day, and what I observe is more specific than the statistics suggest. The organizations that do have policies have typically invested heavily in restriction. Employees receive long, emphatic lists of what they cannot use — which models are prohibited, which data cannot be shared, which tools violate privacy requirements. What they do not receive is equal clarity about what they can use, how to use it well, or what good looks like. The message is loud on the guardrails and nearly silent on the road.
The predictable result has a name: shadow AI — employees quietly using unapproved tools to get their work done. It sounds like a covert operation. It isn’t. It is what happens when people who want to do their jobs well are given a prohibition list instead of a playbook. The fault lies with the guidance gap, not the employee.
That was already the environment. Then the pricing model changed too — adding financial mystery to an operational mystery that was never resolved.
When Incentives Flip, Everything Follows
To understand the vendor side of this shift, it helps to know what flat-rate pricing actually was. For several years, AI companies absorbed compute costs under unlimited subscriptions to drive adoption, build market share, and establish stickiness. Flat-rate pricing was not a philosophical choice. It was a growth strategy — an adoption accelerant designed to remove the friction of financial calculation from every interaction.
It worked. Usage exploded. And critically, it removed the conditions under which organizations would have been forced to ask hard questions about cost, efficiency, and utilization governance. When everything is included, nobody builds discipline around using it carefully.
That era is over. Vendors have passed the compute risk back to the enterprise — and in doing so, flipped their own incentives entirely. Like fee-for-service medicine, where physicians earn more for every test and procedure ordered, AI vendors now have a direct financial interest in boosting your consumption. More prompts, more agent calls, more tokens: all equal more revenue. The all-inclusive model required vendors to manage their own exposure. The metered model does not. Your usage is now their profit.
Which means the enterprise is now the only party in this relationship with an incentive to be efficient. And most enterprises are not yet equipped to act on (or advise their workers about) that incentive.
The New Competency Divide
What makes this moment particularly disruptive is that the pricing shift does not arrive in a vacuum. It arrives inside organizations already navigating confusion about policy, tools, skills, and what AI is even supposed to accomplish at the team level.
The employees who will navigate this well are not the ones who used AI the most during the free era. They are the ones who used it most intentionally. People who approached AI with a clear objective, a structured thought process, and an understanding of how to guide a conversation toward a useful outcome. Those people were already getting better results. Now they will also be getting better value. In a metered environment, the difference between a well-constructed prompt and a vague one is not just output quality. It is cost.
This is the new competency divide. It will not show up on a job description for some time. But it is already separating teams that work with AI purposefully from those that are guessing, and in a token economy, guessing is expensive.
Closing that gap requires deliberate investment in training — not generic AI awareness, but specific instruction on model selection, prompt engineering as a cost discipline, and architectural choices that contain compute costs before they compound. It also requires people who can translate technical signals into business decisions and business priorities into technical constraints. In a metered environment, translation is not a soft skill. It is a financial control.
The Governance Problem That Just Got Harder
Privacy risk. Model risk. Data exposure. Regulatory uncertainty. These governance concerns existed before the pricing shift and will exist after it. What is genuinely new is unanticipated financial exposure at scale. Uber’s AI spending revelation was a warning that most boards did not internalize: compute costs can compound faster than any traditional budget cycle is designed to catch.
I have served on three boards — two of them publicly traded — and I say what follows with genuine respect for the dedication most members bring to the role. I have seen that dedication firsthand. I have also seen the insularity that forms naturally when people who have known and trusted each other for years are asked to honestly evaluate one another’s continued relevance. Board membership is well-compensated, socially cohesive, and professionally desirable. These are not conditions that encourage rigorous self-assessment.
Now, consider this. Fewer than 10% of board members have strong AI expertise. Plus, fewer than one-third of AI practitioners (people responsible for escalating problems) believe their boards understand the issues their company faces with AI. Adding an invisible, fast-moving cost risk to a governance body already struggling with technical literacy does not simplify that problem. It compounds it.
CEOs should be challenging their Board Chairs directly: do we have the expertise in this room to govern AI’s financial exposure honestly, not just its ethical and privacy dimensions? Where restructuring is not possible, the answer is: advisors with hands-on AI experience brought in as genuine participants in governance, not window dressing.
The goal is not a board that can code or calculate token costs. It is a board that can ask the right questions, recognize an insufficient answer, and trust the people in place to tell them the difference.
Right now, most organizations — from the front line to the boardroom — are navigating AI in the dark. The pricing shift did not create that darkness.
It just made it expensive.
