Most AI business cases rest on a number the person building them does not control and has little reason to trust. The token price, and the number of tokens needed to produce one useful outcome. Together they decide whether the whole thing pays for itself. Both are set outside your organisation, both are easy to underestimate, and both are hidden by pricing models designed for adoption more than durability.
The maths works today. The question almost nobody asks is whether it still works at the price you will be paying once the cheap phase ends.
You can see how far below cost it sits when the flat rate breaks. When GitHub moved Copilot toward usage-based billing, developers who had paid a flat thirty or forty dollars a month started posting projected bills of seven hundred, eight hundred, even a few thousand. The work had not changed. Only who pays for it. A quick chat question and a multi-hour autonomous coding session had been treated as the same subscription. The flat fee was never the real price. It was the company quietly paying the difference.
The cheap price is the strategy
This is not a glitch in someone's pricing model. It is the model. You price below cost, you win the workflows, you let them harden into something nobody wants to rip out, and you move the price once leaving is expensive. Cloud, streaming, ride hailing, food delivery, every one of them was cheap until you had nowhere else to go. AI is running the same play.
The same pattern showed up at Anthropic. Claude's Max plan gave heavy users far more room than ordinary subscriptions, but Claude Code quickly exposed the economics. Anthropic later introduced weekly rate limits after saying some users ran it continuously in the background, and that a few outliers were very costly to support. The exact outlier is not the business case. The signal is that flat pricing was hiding real inference cost.
The economics underneath are starker than those earlier industries ever were. Per-token prices have fallen sharply since 2022, yet enterprise AI bills have climbed, because agentic tools turn one request into many. A simple prompt becomes planning, context loading, tool calls, retries, validation, and output generation.
The unit got cheaper. The outcome got more expensive.
The meter is arriving
Flat subscriptions hid something useful. They capped your exposure. Whatever you did inside the plan, the bill at month end was the same. The shift now underway removes that cap. GitHub's new model keeps the headline fee but turns it into a finite allowance of credits priced from model usage. Spend the allowance and the meter runs against your card.
Read what that does to your incentives. Under a meter, the more your solution is used, the more it costs you. Adoption was the goal. Now adoption is the line that grows. If every successful workflow consumes more context, more retries, and more agent steps, success does not just increase usage. It changes the cost per outcome.
There is a smaller change inside the GitHub move worth noticing. The old plans had a silent fallback. When capacity was tight, Copilot could quietly serve a cheaper model so the request still completed. That fallback is being removed. The cushion you were leaning on without knowing it was a cushion disappears the day the economics change.
A business case is a bet on a price you do not set
When you signed off the solution, you modelled a return. That return assumed a cost, and that cost assumed both today's token price and today's token volume per task. So the business case is, underneath, a bet that a price you do not control stays where it is, and that the workflow will not become more expensive as it becomes more capable.
The counterparty owns the variable. It also has the incentive to move it, and once your processes are built around the model, the leverage to make it stick. A large enough buyer can negotiate a private rate. Most cannot, and even a negotiated rate is still a number the counterparty sets.
The obvious escape is to switch providers, and it is weaker than it looks. The serious alternatives are in the same below-cost phase, funded by the same investors, heading for the same place. You are not choosing between a cheap provider and an expensive one. You are choosing between two temporary prices and budgeting as though one of them is permanent. Different logo, same exposure.
Build for the price you will pay
The fix is not to avoid AI. It is to stop pricing the case on a number designed to move.
Model cost per outcome, not cost per token, and then stress it. Re-run the business case at three times today's cost per useful outcome, then five. Include the model price, but also the amount of model work each outcome consumes when the solution becomes agentic, popular, and load-bearing.
Route the work. Use cheap models for the routine majority, and frontier models only where they change the result. Set budgets, caps, and observability before adoption becomes success and success becomes an uncontrolled invoice.
The aim is a solution whose value still clears the cost when the economics normalise. If the outcome is worth far more than the real cost of the model work, you have something durable. If it only clears the bar at today's price and today's pilot usage, you have something on loan.
The test
Take the business case for your AI solution. Find the cost per useful outcome it quietly assumes. Then stress both parts of it, the price of the model and the amount of model work each outcome consumes. Multiply the cost by five and run the case again.
If it still pays for itself, you have a business case.
If it only works at today's price and today's usage, you do not have a business case. You have a free trial, paid for by someone who is not you, and they can end it the moment they decide they have won.