Summary of the Core Content
This article highlights a common dilemma in corporate transformation during the AI era: in order to implement AI effectively, many companies link employees' usage of AI (measured by token consumption) with their performance and promotions. As a result, employees abuse AI to meet these metrics—doing mundane tasks like converting PDFs to PPTs—with the consequence of soaring AI-related expenses. Companies such as Accenture, Uber, Amazon, and Meta have all fallen into this trap. The underlying issue is the "Goodhart's Law"—a management principle that states when a metric becomes a goal itself, it loses its accuracy. Additionally, there’s a misalignment where technology (AI tools) outpaces the rules (evaluation methods). The article offers solutions by focusing on changes in evaluation systems, selecting specific use cases, and improving the behavior of managers themselves.
Detailed Analysis
1. From “AI for Everyone” to “Expensive Bills”: A Collective Embarrassment
Accenture spent $865 million on restructuring and linked AI usage to employee performance. To avoid falling behind, employees resorted to using AI for trivial tasks like converting PDFs to PPTs, leading to a dramatic increase in token consumption and unaffordable costs. This isn’t unique to Accenture: Uber burned through its annual AI budget in just four months, forcing it to set a monthly limit of $1,500 on tool usage; Amazon’s employees engaged in token-generating activities that were later promptly removed by the company; at Meta, someone consumed 28.1 billion tokens (worth millions of dollars) in 30 days. These companies initially encouraged more AI use, treating it as a sign of success, but only realized the cost spiral when control was lost—similar to giving employees cars and then expecting them to drive recklessly without considering the fuel costs.
2. Goodhart’s Law: When Metrics Become Goals, They Distort Reality
This is not a new problem in the AI era; it’s a well-known management flaw. When a metric is used as a goal, it no longer reflects true performance. For example:
- Programmers might write 50 lines of code instead of just 5 to meet usage targets;
- Employees may use AI for tasks like PDF conversion, which becomes an expensive waste of resources;
- In the Soviet Union, factories produced small nails to meet production quotas, resulting in useless products. In the AI context, token consumption serves the same purpose—being used merely to satisfy metrics rather than to solve real problems.
3. Technology Advances Faster Than Rules: The Old Ways Remain
The author cites an historical example from the 1910s when a U.S. island banned cars, and mailmen still used horses to deliver letters (the new technology was not allowed). Many companies today are using AI but still evaluating it based on outdated metrics. Employees end up using AI for unnecessary tasks, such as translating novels into English and generating videos just to meet usage targets, sometimes causing significant financial losses.
4. Three Steps to Overcome This Problem
The author suggests three practical solutions:
- Focus on Results, Not Tokens: Instead of counting tokens, focus on the actual outcomes—has productivity increased? Has speed improved? Can tasks that were previously impossible be completed now? Tokens are a means to an end; the real goal is the result.
- Reevaluate Evaluation Methods Before Implementing Tools: Many companies buy AI tools without changing how they assess employee performance, leading to excessive use. First, change the evaluation criteria to focus on outcomes, and then employees will use AI more effectively.
- Start with a Specific Use Case: A company-wide transformation can be daunting, so start with a clear use case (e.g., generating customer reports). Demonstrate success within three months, and others will follow suit.
5. The Role of Managers: Lead by Example
Managers must lead by example. If they don’t use AI themselves, employees won’t take it seriously. Only by experiencing its benefits (e.g., saving time with AI-generated reports) can managers identify which practices are effective and which are merely superficial. They should also question existing rules that hinder AI’s effectiveness.
In essence, this article emphasizes that AI transformation is not about a competition in usage but about creating value. Old rules must be replaced to allow AI to truly contribute to business success.