“A single employee generated a $38,000 AI bill in three hours”
As AI spreads across the enterprise, organizations are scrambling to control token costs without slowing adoption.
Until recently, organizations could calculate their monthly labor costs with remarkable accuracy. Headcount, salaries, benefits and employment expenses formed the foundation of budgets and workforce planning.
Artificial intelligence is changing that equation.
As AI tools spread throughout the workplace, companies are facing a new and rapidly growing cost center that barely existed a few years ago: tokens.
Tokens are the units AI models use to measure consumption. Every prompt submitted to a chatbot, every document analyzed and every response generated is converted into tokens, which determine how much an organization pays. The more advanced the model, or the more complex the task, the higher the token consumption and the larger the bill.
1 View gallery
From left: Dor Atias, Inbal Namir, Oded Tahori.
(Photos: Sponsor, Almog Sogbaker, Nikki Truk)
Once a technical term understood mainly by software engineers, tokens are becoming a concept that every executive will need to understand.
An increasing number of AI services are priced on a pay-as-you-go basis, making token consumption a significant operating expense. Organizations must now encourage employees to use AI while ensuring they do so efficiently. In effect, managers are no longer responsible only for human employees, they are also managing a growing digital workforce.
"Tokens are the smallest unit of consumption AI systems use," says Inbal Namir, Managing Director at Deloitte and leader of the firm's Future of Work practice. "Whether it's a question you ask or an answer you receive, everything is measured in tokens."
According to Namir, the shift is far more significant than simply adopting a new technology.
"Until now, organizations have managed their economics around labor hours and salaries. They know how many employees they have, what they cost and what their payroll will look like each month.
"Now a second production engine has entered the organization, digital intelligence. Unlike payroll, its cost isn't entirely predictable. Companies can estimate it, but they cannot yet forecast it with the same level of confidence."
The result, she says, is that organizations now operate with two productive resources: human labor and AI.
"One has predictable costs. The other is still evolving, and organizations are racing to build the tools needed to forecast and manage it."
The shift is also changing how companies think about efficiency.
Traditionally, organizations measured productivity against employee time. Now they are beginning to measure output relative to AI resources consumed.
"Employees now control this new productive resource," Namir says. "Some know how to use AI efficiently and achieve excellent results with relatively few tokens. Others consume far more tokens than necessary because they lack experience or don't realize those tokens represent real money."
As a result, companies are beginning to evaluate employees not only by what they produce, but also by the efficiency with which they use AI.
"Organizations initially rewarded adoption, they simply wanted employees to use AI tools," she says. "Now they're realizing that usage alone isn't enough. They also need to measure efficiency and business value."
Some AI providers have also shifted from fixed-price subscriptions to usage-based pricing, making token efficiency increasingly important.
The changes extend well beyond employees.
"We see the biggest impact on middle managers and team leaders," Namir says. "They now need entirely new management capabilities."
Managers must decide which AI models employees should use for different tasks, monitor token consumption and investigate unusually high costs.
"If an employee consistently chooses the most expensive model, managers need to understand why," she says.
Organizations are responding by investing heavily in AI training, establishing networks of internal AI experts and creating governance frameworks to guide employees.
"Continuous learning has become critical because AI changes constantly," Namir says. "Many organizations are creating networks of AI champions who stay at the forefront of new developments and share that knowledge across the company."
Another emerging challenge is allocating AI resources fairly.
"It's not one size fits all," Namir says.
Teams building complex software may require AI budgets worth thousands of dollars each month, while others need far less.
"Organizations have to define token allocation policies and communicate them clearly. Otherwise it can create tensions between teams."
For Oded Tahori, CEO of Jeen.ai, token management has already become a core enterprise challenge.
"We started building token management into our platform more than a year ago," he says. "If you want employees to have the freedom to innovate with AI, you also need security controls and financial controls."
Without those controls, costs can escalate rapidly.
"We've seen cases where a single employee unknowingly ran an AI process overnight that generated a $38,000 bill in just three hours," Tahori says.
In another case, he says, a cloud provider contacted one of its customers after detecting an unusually large spike in AI spending caused by an automated loop.
"Multiply that across thousands of employees, and costs can quickly reach millions of dollars."
As AI adoption accelerates, Tahori argues, organizations need formal policies governing AI spending, including approval workflows when employees exceed their allocated budgets.
"We allow managers to allocate AI budgets just like traditional departmental budgets," he says. "If one employee is away on vacation, that budget can even be reassigned to someone else."
Increasingly, companies are weighing whether the next dollar should fund another employee, or additional AI capacity.
Not everyone believes token consumption should become an employee performance metric.
Dor Atias, founder and Chief Product Officer at Cycode, argues that measuring people by how many tokens they consume misses the point.
"Measuring employees by token spending is complete nonsense," he says. "People should be measured the same way they always have been, by what they deliver, the quality of their work and the value they create."
AI, he says, has changed how work gets done, not how performance should be judged.
"The strongest engineers before AI are generally still the strongest today. They simply use AI to accomplish much more."
What has changed is hiring.
Companies increasingly prefer experienced engineers who can supervise AI-generated output, while junior developers face greater challenges.
"Today we're hiring more experienced and versatile people," Atias says. "Everyone needs to know how to work effectively with AI."
Cycode has also automated many internal processes using AI agents, including documentation.
"We're constantly looking at which repetitive tasks can be automated and how to reduce token consumption by improving prompts and workflows," he says.
