For much of 2026, artificial intelligence inside large companies resembled an open-ended experiment. Employees were encouraged to use AI assistants as much as possible. They generated code, summarised meetings, analysed data and produced reports at unprecedented speed. In some organisations, using more AI became a badge of honour rather than simply a way to improve productivity.
That culture is now changing. New research from SemiAnalysis, based on conversations with more than 50 enterprise customers and discussions at the Databricks AI Summit, suggests companies are moving away from unrestricted AI usage and towards carefully managed budgets. Instead of rewarding employees for consuming more AI tokens, many businesses are now monitoring every prompt and placing limits on monthly spending.
The Rise of 'Tokenmaxxing'
Earlier this year, companies including Meta and Salesforce encouraged employees to make full use of AI tools to improve productivity. According to SemiAnalysis, Meta employees consumed more than 60 trillion AI tokens over a 30-day period. One employee reportedly created an internal dashboard known as 'Claudeconomics', ranking the company's 250 biggest AI users. Some employees competed for titles such as 'Token Legend' and 'Cache Wizard', while AI agents were reportedly left running for hours to consume additional tokens.
The internal leaderboard was removed two days after The Information reported on it. The incident became one of the most visible examples of what SemiAnalysis calls 'tokenmaxxing', where organisations encouraged extensive AI usage before introducing financial controls.
Budgets Replace Unlimited Access
Only months later, the focus has shifted. SemiAnalysis found that most organisations it spoke to have now introduced AI budgets, although the limits vary significantly.
Some companies cap employees at $250 or $500 per month, while technology firms such as Workday and Stripe allow budgets of around $2,000 per month. Other businesses use flexible limits based on seniority or the type of work employees perform.
Rather than allowing unrestricted access, companies are increasingly treating AI as another operational expense that must be managed alongside software subscriptions and travel costs.
The Biggest Headlines Do Not Tell the Whole Story
Reports about heavy AI spending at companies including Meta and Uber have attracted widespread attention. However, SemiAnalysis argues these examples do not represent the wider enterprise market.
The firm's analysis suggests the largest AI customers account for a disproportionate share of spending, while many businesses remain well below those levels. Citing data from Ramp Economics Lab, SemiAnalysis notes that median enterprise AI spending is far lower than that of the highest-spending technology companies. Its researchers also believe enterprise AI adoption still has considerable room to grow, particularly as more business functions begin using AI beyond software development.
Companies Are Finding Ways to Reduce Token Use
As spending comes under greater scrutiny, organisations are encouraging employees to use AI more efficiently. Some have switched default AI models from premium versions to lower-cost alternatives. Premium models remain available, but employees must actively choose them.
SemiAnalysis also found that some workers use Microsoft 365 Copilot for early drafting and brainstorming because that usage does not count against separate token budgets. They then move to paid services such as Claude or Codex only for more demanding tasks. The result is a growing focus on achieving the same outcome while consuming fewer AI tokens.
Businesses Still See Strong Returns
Although spending controls are becoming more common, companies continue to report significant productivity gains. An Amazon recruiter told SemiAnalysis that AI tools have reduced the recruitment process for principal engineers from six to nine months to roughly half that time by assisting with interview notes and reporting.
Meanwhile, an employee at a data and analytics provider serving 85 per cent of the Fortune 500 said work that previously took a week can now be completed in only a few hours. These examples help explain why businesses are introducing budgets rather than abandoning AI altogether.
AI Spending Continues to Grow
SemiAnalysis believes concerns about a sharp slowdown in enterprise AI spending are overstated. According to its Tokenomics Model, coding-related work continues to account for more than 70 per cent of annual recurring revenue across OpenAI and Anthropic. The firm also expects demand to expand into cybersecurity and broader knowledge work as AI tools become more deeply integrated into enterprise operations.
Its conversations with customers suggest most organisations are refining how AI is used rather than reducing investment.
A New Phase for Enterprise AI
The early months of 2026 were defined by rapid experimentation and unrestricted AI usage. Now, companies are entering a more disciplined phase.
Budgets, spending limits and model selection are becoming standard business practices. Organisations still expect employees to use AI to improve productivity, but they also expect them to justify the cost. For many businesses, the question is no longer how much AI employees can use. It is whether every AI prompt delivers enough value to justify its cost.
