Companies: Budgets and Safeguards for AI Use

After a phase of widespread encouragement, large companies are now imposing limits and budgets on AI tools. Spending is increasing, risks are being identified, and access is now adjusted to the needs of the business rather than being completely open.

Costs Estimated in the Hundreds of Billions for Limited Value

According to Menlo Ventures, investments in AI deployment within companies reached $177 billion in September, up from $33 billion in 2025. Meanwhile, a survey conducted by Gartner between May and July among approximately 2,000 technology leaders found that only 13% reported deriving significant value from AI. This trend is prompting organizations to prioritize optimizing return on investment rather than simply multiplying use cases.

Security: Unauthorized Uses and Potential Leaks Documented

Broad access to AI models can generate risks if the tool is misused. A KPMG survey conducted in 2025 among over 1,000 American workers reported that 44% used AI without authorization or inappropriately, and 46% had uploaded sensitive information or intellectual property to public platforms. In this context, companies are strengthening their safeguards while investing more in AI.

Caps and Budgets: From Open Access to Conditional Use

In September, JPMorgan implemented a monthly limit of $2,000 for certain engineers using Claude Code, linking spending to productivity, quality, risk reduction, and business impact. At Cisco, after a learning period during which development teams had unlimited access to tokens, managers of the 32,000 employees must now manage token budgets. This general movement is reflected in the establishment of budgets and restricted access for certain employees or tools.

Differentiated Access to Models: The Goldman Sachs Case

Goldman Sachs is rapidly expanding the range of AI tools available to its 45,000 employees, while considering it inappropriate for everyone to have access to the most advanced models on a daily basis. According to David Solomon, cutting-edge models are required for certain tasks, particularly in cybersecurity, while others do not necessitate them. This approach is part of a broader context where AI usage is increasing significantly and safeguards are deemed necessary, according to Lari Hämäläinen from McKinsey.

Managing Adoption: Dashboards and Incentives, Then Refocusing

To make usage visible and sometimes stimulate adoption, some companies have deployed dashboards. At Disney, the largest users were making hundreds of requests per day and consuming tens of millions of tokens. KPMG US indicated in May that it wanted to encourage more frequent and sophisticated usage among about 10,000 employees in its consulting business via a dashboard, before ceasing this monitoring when AI became integrated into daily operations. This management followed a period where some employers highlighted massive token consumption.

Token Spending: From Celebrating Volumes to Fine Control

Visa announced it had doubled its monthly token consumption from 1 trillion in February to nearly 2 trillion in March, accompanied by rewards for the highest-performing users. In March, Jensen Huang stated that Nvidia aimed for up to $2 billion in annual token spending for its engineers and expressed concern that an engineer costing $500,000 consumed less than $250,000 in tokens. These highlights of high volumes preceded a phase of increased control, motivated by consumption deemed significant by workers.

Measured Effects of Caps: The Rippling Example

Rippling implemented monthly limits on the use of its AI tools in August after anticipating that token-related costs could reach 40% of the personnel budget allocated to R&D. The company indicates that with the introduction of these caps and more effective management of model routing, this estimate has been revised down to between 10% and 15%, while maintaining a token usage volume close to the previous maximum. According to Matt MacInnis, the goal is not to reduce costs but to accept their growth when it is productive.

Why Tighten Regulation: Incentives and Usage Behaviors

According to Alex Sobol, AI bills can skyrocket because users have little incentive to limit their consumption, and the absence of limits encourages extensive use of the tools. Companies, which still wish to capitalize on AI, are therefore tightening access and formalizing rules, not to mechanically reduce spending but to make the most of it while incorporating security considerations.