Summary of the Key Points in Plain Language
This article directly addresses a common misconception in the AI community: everyone is focused on improving model accuracy, believing that once AI reaches 90% or 95% accuracy, it can replace human labor on a large scale. However, in reality, the implementation of AI in most industries has stalled halfway. The author, through in-depth discussions with professionals from various industries, realized that whether AI can be effectively integrated into business processes and become a real source of productivity does not depend solely on accuracy. Instead, three crucial, yet often overlooked issues are at the heart of the matter: whether there are supporting systems in place to ensure AI can complete tasks smoothly without failure; whether it is possible to quickly assess the correctness of AI’s work at a low cost without constant human supervision; and who will be responsible for any mistakes and how to mitigate them. The article provides a clear path for companies looking to implement AI: instead of waiting for a perfectly accurate model, it is better to start with small, well-defined tasks that are easy to check and can be corrected if necessary. This approach is more likely to lead to tangible results than attempting to create an “automated AI brain” from the start.
---
Detailed Explanation of Each Point
1. Why is AI becoming more intelligent, but implementation is slower? The lack is not the “engine” but the “transmission framework”
Many people think that buying a powerful AI model with large parameters means it can be put to work immediately. It’s like buying a 100-horsepower engine and expecting it to plow the field on its own—no matter how powerful the engine is, if it doesn’t have the necessary tools (plow, wheels, steering wheel), or if the field’s boundaries aren’t defined, it will get stuck. The “Harness” mentioned in the article is essentially a set of external rules and tools customized for your business, specifying what data AI can access, which tools it can use, and how to get feedback after completing tasks. With the same top-tier model, some AI agents can run tasks continuously without errors for five days, while others fail with just a few routine tasks. The difference lies not in the model’s capabilities but in the quality of the supporting framework. Moreover, more plugins do not necessarily mean a better framework; for longer and more complex tasks, it’s crucial to set clear boundaries for AI to stop working automatically when risks arise. Stability is always more important than complete functionality.
2. The hidden barrier to AI’s potential for profitability: the cost of verification
There is a critical industry boundary that is easily overlooked: whether it is cheap to determine whether AI’s work is correct. This directly affects whether an industry can be rapidly transformed by AI. For example, code-related AI was the first to be successfully implemented because its verification is almost cost-free—you can run the code and get immediate feedback on whether it compiles or tests correctly. In contrast, clothing AI faces significant challenges: a 99% accurate image may still be unusable if there are minor errors. If each image requires a senior designer to check for details, the time saved by AI is wasted on verification, making it more expensive than manual work.
3. The true value of industry experts is not their knowledge base, but their ability to articulate doubts
Previously, experts were asked to organize their knowledge into documents for the AI to learn from. However, this misses the core value of experts: the real challenge in many industries is not just finding information, but identifying logical flaws in AI outputs. Experienced experts can instantly spot issues—like changes in button materials or flaws in investment logic. These skills are not learned from textbooks but from years of experience. The new role of experts is to transform these vague doubts into clear, executable rules that the AI can follow. Only a few borderline cases require human intervention, turning industry expertise into usable productivity.
4. Can you use AI with 80% accuracy? The context matters more than the accuracy
The question of using 80% accurate AI is not about the number itself but about its role. For example, a 80% accurate medical AI might seem unreliable for making diagnoses, but it could be very useful for organizing medical records or identifying high-risk symptoms. Similarly, in investing, an 80% accurate AI can help find discrepancies in research reports. The key is to use AI for auxiliary tasks where it can provide valuable support without causing major mistakes.
5. Don’t rush to implement “grand AI solutions”; start with small, manageable tasks
Many companies aim for ambitious goals like creating an “enterprise brain” or a fully automated investment system, but these goals are poorly defined and often fail. A more reliable approach is to start with small, high-frequency tasks that are easy to verify and have minimal consequences if they go wrong. This allows AI to gradually improve its performance and become a valuable tool. By moving away from the idea of AI making final decisions, companies can leverage its potential more effectively.