Businesses ditch token-heavy AI as cost pressures reshape enterprise use

OpenAI charges by usage and by model, and Microsoft Azure’s OpenAI pricing follows the same pattern, turning every extra token into a measurable cost for enterprises. That billing structure is pushing companies to trim prompts, narrow use cases, and choose cheaper models when a task can be completed without a larger system.

The shift follows an early rush in which many workplaces opened generative AI broadly, testing long prompts, large chatbots, and broad deployments to see what would stick. Workers have used AI to save time and develop ideas, which helped fuel adoption. Gallup found AI use at work continued to rise in 2025. The result is that many firms now face the bill after the experimentation wave, with executives asking whether the outputs justify the spend.

Stanford Digital Economy Lab’s Enterprise AI Playbook, published in April 2026 and based on 51 successful deployments, reflects a market no longer focused only on model capability. Menlo Ventures’ 2025 State of Generative AI in the Enterprise, published on December 9, 2025, captures a similar pivot as buyers shift from broad enthusiasm to implementation lessons. The Federal Reserve’s Measuring AI Uptake in the Workplace, published on February 5, 2024, tracks AI use where the work actually happens, not just in labs.

The cost question has become central for procurement and finance teams. Pricing models matter more than headline cost, and AI tokens are driving new spend dynamics. A model that looks cheap at the prompt level can become expensive when employees use it heavily across routine work.

The practical response is a more disciplined enterprise style of AI use. Companies are favoring smaller, specialized models for targeted tasks because they can be faster, cheaper, and easier to manage. They are also limiting AI to higher-value workflows such as document summaries, draft emails, internal knowledge search, and repetitive administrative tasks, instead of allowing open-ended use that drives up cloud bills.