Real LLM Usage Data: 7-Month Study Reveals Work Habits

A groundbreaking 7-month study has provided the first concrete data on how employees actually use Large Language Models (LLMs) in the workplace. Unlike analyst predictions or vendor presentations, this research is based on direct transaction counts from real user logs.

The study tracked 527 employees from June through December 2025, offering unprecedented insight into practical AI adoption. The findings challenge common assumptions about workplace AI usage and reveal significant cost implications for companies choosing between subscription and pay-per-token models.

The company decided to lead by example, providing all employees with a clear service to test various market models. The goal was to observe practical productivity gains rather than theoretical benefits. A critical decision was choosing pay-per-token pricing over fixed subscriptions, which proved crucial for accurate measurement.

Researchers tracked 122,346 queries over the seven-month period. The data revealed that 416 users out of 527 actively engaged with the tools at least once. This represents a 79% adoption rate among employees given access to the technology.

"If what, large models show users, but carefully hide the number of requests and traffic. Because it is extremely low there."

The study utilized an aggregator of neural networks that included image generation capabilities. This comprehensive approach allowed for tracking diverse use cases across different AI models, including Gemini 3 Pro Preview, the latest GPT models, and Anthropic's offerings.

The financial data reveals striking contrasts between pay-per-token and subscription models. Total expenses reached $6,851 (approximately 535,000 rubles) over seven months. This breaks down to 184 rubles per active user monthly—a figure that would be dramatically higher with subscriptions.

If the company had chosen standard $20 monthly subscriptions, costs would have been 8.5 times higher for the same usage level. This highlights the economic inefficiency of flat-rate pricing for variable AI usage patterns.

Key cost breakdown:

• 64% of budget allocated to image generation
• Text-only usage cost just 62 rubles per user monthly
• Annual cost per active user: $184
• 79% of employees actively used the tools

The data shows that image generation consumed the majority of resources, suggesting visual content creation is a primary workplace AI application. Text-based queries, while valuable, represented a smaller portion of overall usage.

A critical factor limiting AI adoption is the prompt formulation barrier. Research by Jakob Nielsen indicates that only 20% of the population can effectively structure prompts for optimal results. This skills gap explains why many employees try AI tools briefly and then abandon them.

The study observed that users who understood how to apply the models to their specific work needs became consistent, returning users. This pattern suggests that training and education are as important as tool access for successful AI integration.

"They try a couple of times and leave."

The 42-query average per user per month indicates moderate engagement rather than heavy reliance. This usage pattern supports the argument for flexible pricing models over fixed subscriptions, as most employees don't require daily AI assistance.

The study provides actionable insights for organizations considering AI implementation. First, actual usage differs significantly from vendor projections. Companies should base their budgeting on real transaction data rather than theoretical maximums.

Second, the image generation dominance (64% of budget) suggests that visual AI tools may be more valuable for workplace productivity than text-based models alone. Organizations should consider this when selecting AI platforms.

Third, the cost efficiency of pay-per-token models becomes apparent at scale. For 527 employees, the difference between subscription and usage-based pricing represents substantial savings without sacrificing access to state-of-the-art models.

Finally, the 79% active usage rate demonstrates that when employees are given clear, accessible AI services, adoption follows naturally. The key is providing the right tools with appropriate pricing models that match actual usage patterns.

This 7-month study provides the first concrete evidence of how LLMs are actually used in workplace settings. The data challenges several common assumptions about AI adoption and cost-effectiveness.

Organizations should consider pay-per-token models over subscriptions, as they align costs with actual usage. The 8.5x cost difference represents significant savings for companies with moderate AI requirements.

Training remains crucial for adoption. Only 20% of users can effectively prompt AI systems, suggesting that educational programs should accompany tool deployment.

Finally, the visual AI emphasis (64% of budget on images) indicates that workplace AI value extends beyond text processing. Companies should evaluate their specific needs before committing to any single AI solution.