Many companies thought AI would bring cost reductions, but instead, they first received a more expensive bill.

Over the past year, an increasing number of enterprises have implemented 'all-in AI.' However, by 2026, companies began to reflect on this approach.

Amazon took down its internal AI usage ranking, Microsoft started revoking internal Claude Code access for most employees, and Uber's CTO revealed that the company's AI programming tool budget for the entire year of 2026 had already been exhausted in the first four months.

Many companies initially wanted to measure whether employees were embracing AI, so they chose the simplest metric: token consumption. However, an absurd situation soon emerged: companies intended to reward productivity but ended up rewarding 'who could burn more tokens.'

The Absurd Token Ranking

Over the past year, many large companies have done the same thing: encouraging employees to use AI more. To make this measurable, some companies created rankings based on token consumption, AI invocation frequency, and activity levels for employees or teams.

The original intention of enterprises is easy to understand: management wanted to promote AI usage across the organization, but the results quickly went awry.

Internally at Meta, an employee-built AI token usage ranking called Claudeonomics emerged. This leaderboard tracked token usage among more than 85,000 employees, with a total consumption of approximately 60 trillion tokens in 30 days. The top individual consumed about 281 billion tokens alone.

The numbers are staggering, but it's hard to say how much effective output they represent.

Amazon's case is even more typical. Its developer platform, Kiro, once had an internal ranking called 'KiroRank,' which scored employees based on AI usage activity.

As a result, many employees performed meaningless tasks with AI to improve their rankings, directly driving up computational power consumption. Eventually, the ranking was taken down, and company management had to remind employees: 'Don't use AI just for the sake of using AI.'

An indicator originally meant to measure 'AI usage enthusiasm' ended up becoming a game of 'who could burn more tokens.'

It's like a company wanting to encourage employees to exercise, so they ranked everyone by daily step count. As a result, some people sat at their desks shaking their fitness bands, tied their phones to their dogs after work, or even bought step-counting shakers.

Tokens Reveal Organizational Inefficiency

However, the AI bill has a positive side. It exposes many previously invisible inefficiencies within enterprises and quantifies them with real money.

Previously, when an employee wrote low-quality requirements, held ineffective meetings, or repeatedly reworked tasks, the company only knew they received a salary.

Now, with AI involved, every inefficiency becomes token consumption.

Unclear requirements lead to agents repeatedly trial-and-erroring.

Dirty documentation causes RAG to perform extensive retrieval.

Restricted permissions force AI to guess in incomplete contexts, make mistakes, and then revise, burning more tokens.

So, the AI bill doesn't appear out of nowhere—it's a developer for enterprise inefficiency.

This is the financial dilemma of AI's 'copilot mode': after introducing AI, if enterprises don't optimize processes or restructure the organization but simply add a layer of AI on top of existing work methods, the result is: inefficiency remains, and the token bill adds another expense.

What Should Really Be Measured

In the future, enterprises should focus not on 'how much AI is used' but on three conversion rates:

1. Conversion rate from tokens to output. That is, how many tokens are required to complete each effective output.

For R&D teams, it's not about who consumes the most tokens but how much AI cost is incurred per merged PR, per fixed bug, or per completed code review.

For customer service teams, it's not about how many AI responses are sent but how many tokens are spent per resolved ticket.

For sales teams, it's not about how many emails AI generates but how much cost is incurred per effective lead.

2. Conversion rate from tokens to business results. This layer answers whether AI transforms output into results.

For AI-generated customer service responses, look at first-contact resolution rates, escalation rates, and customer satisfaction improvements.

For AI-assisted sales, examine changes in conversion rates and deal cycles.

3. Conversion rate from token costs to organizational cost reductions.

Has AI reduced manpower? Shortened cycles? Decreased rework? If the answers are all no, then AI is likely just a new layer of cost added to the existing organization.

Regarding token cost management, the FinOps Foundation, a non-profit industry organization under the Linux Foundation, has already begun discussing AI FinOps. It argues that AI cost management requires practices like model routing, feature-level budgeting, LLM invocation metadata, and cost-per-output.

AI is not free productivity, and cost reductions won't happen naturally just by integrating AI.

The most important thing is not 'making employees use AI more' but making enterprises increasingly aware of what they're getting in return for every token spent.

END

This article is an original work from 'Intelligent Evolution Theory.' Welcome to follow us.