Microsoft, Google, and OpenAI have recently revised their AI contracts, moving from a per-seat subscription model to one that charges based on how much AI is used. Previously, companies paid a fixed fee per user, regardless of how much they used the AI tools. Now, the cost is tied to usage, which changes the risk of high-volume consumption. On September 25, Microsoft announced updates to its Copilot pricing model. The new model has two parts: "Everyday AI," a monthly license per user that covers basic tasks like summarizing documents, and "Advanced AI," which includes features like Cowork, Code, and Autopilot. This advanced part is paid for using Copilot Credits, which are based on how much AI is used, while the per-seat license is still required for access. OpenAI also updated its Enterprise contracts in mid-September to use token-based billing. Services like Chat, Work, and Codex are now billed based on the number of tokens processed, in addition to the fixed cost per seat. Even if a seat is not used, the company is still charged for it. Anthropic uses a similar system, charging $20 per seat per month plus the cost of API usage. Google has taken a different approach with its Gemini Enterprise Standard or Plus subscriptions, which no longer include Gemini Code Assist. Current users can keep the tool until the end of their contract, but they will need to find an alternative after that. The cost of high-end AI models has decreased. Anthropic's Opus 5.5 is now priced at $4 per million input tokens and $20 per million output tokens, a 20% drop from Opus 5. This is the first time the Opus series has seen a price reduction. The shift in pricing models changes who bears the risk of high usage. With the per-seat model, the provider estimated average usage and absorbed the costs of heavy users. Now, with usage-based pricing, the customer takes on the risk. However, the usage growing this year is not limited to simple tasks like summarizing documents but involves delegating complex tasks to AI agents over extended periods. Microsoft's post suggests that metering routine tasks could be a "bad deal" for companies, as they might limit access to AI to control costs. This applies to half of the features Microsoft has metered. For a Chief Information Officer (CIO), the key new feature is the ability to set spending policies. At Microsoft, the usage-based part is disabled for enterprise users until an administrator creates a spending policy. Before that, no charges are made. Budgets can be set per tenant, per group, or for individuals, with alerts and the ability to link costs to an Azure subscription. OpenAI offers similar features with a monthly workspace budget and usage limits per user and group. This setting is like an administrative parameter that sets a budget line. The concern is that an administrator might activate a policy to unlock a pilot, set a limit arbitrarily, and the finance department might later discover unexpected expenses. Until a business manager signs off on a budget for a process, the spending policy remains inactive. Once usage is enabled, stopping an AI agent is like shutting down a service. Contracts also include other hidden details: at Google, a simple online renewal can remove a tool that developers use daily, and license renewals are rarely reviewed. At OpenAI, a spending commitment can cover both ChatGPT Enterprise and the API, but under-consumption is still billed according to the contract. However, the ChatGPT dashboard does not show API usage, so tracking consumption falls to the customer, often with the help of spreadsheets. Microsoft compares its model to a rechargeable hybrid, with the license acting as a battery and credits as a fuel tank. While the analogy is clear, it doesn't specify who decides when the "engine" starts. For twenty years, CIOs have bought software based on the number of users. Now, AI is charged based on the work it performs, similar to how cloud services are billed. Companies that spent years learning to manage cloud costs now have a limited time to adjust. They no longer have years to prepare; they have a renewal date.