From cool features to real industrial applications, where is AI stuck?

Since the release of ChatGPT in November 2022, generative AI has developed at a high speed. The competition among large models has reached a white - hot stage, with performance indicators constantly being refreshed and multimodal capabilities continuously improving. AI agents can autonomously call tools to complete increasingly complex tasks. Many manufacturers of large AI models claim that the era of Artificial General Intelligence (AGI) is approaching.

In sharp contrast to the rapid progress of technology is the lag in commercial implementation. Data from the US Ramp AI Index shows that the proportion of US companies adopting paid AI products has recently stagnated and even declined.

A research report (The GenAI Divide: State of AI in Business 2025) by the Massachusetts Institute of Technology in July 2025 pointed out that 95% of generative AI application projects have poor results or are aborted midway. This report even triggered a shock in the US stock market.

When the bold claim that "all industries need to be redone with AI" encounters the reality of "high failure rate of AI projects", we have to ask: Where exactly is the bottleneck for AI to move from cool features to real - world industrial applications? And how can we break through the fog and achieve a real value closed - loop?

01

Business Process Reconfiguration and AI Path Planning

The performance indicators of AI models cannot be directly translated into commercial value. Currently, AI cannot provide end - to - end solutions in most cases. Therefore, the implementation of AI applications needs to prioritize the promotion of business processes where AI capabilities are relatively mature, enterprise data accumulation is relatively complete, and the value is most significant, based on the capabilities and limitations of AI, combined with the business scenarios, needs, and pain points of industries and enterprises. This requires finding the minimum viable flywheel of input - data - benefit at the intersection of technology and demand. While generating economic benefits, new data is generated to feed back into model optimization, forming a virtuous cycle of continuous iteration.

Therefore, at this stage, the implementation of AI applications requires a process of work - flow segmentation and business - process reconfiguration. The parts that AI is good at should be handed over to AI; the remaining parts, whether due to limitations in AI capabilities or insufficient data accumulation, still need to be completed by humans. Human work is to control AI, bridge process breakpoints, allocate tasks and resources, and evaluate and correct results.

We can compare the above - mentioned business - process reconfiguration process to path planning. For example, if you want to go from Caohejing High - tech Development Zone in Shanghai to Fudan University, the fastest route is not a straight line on the ground but to take the elevated road. AI is like the elevated road, which can greatly improve the speed of travel but cannot cover the entire journey, so ground roads are still needed to connect the two ends, and this is where human beings come in.

There are three similarities between the business - process reconfiguration required for AI implementation and path planning:

First, in path planning, you take the highway when it's available and the ground road when it's not. Sometimes, not only at the two ends of the journey but also in the middle, the highway may not be connected, and you need to take the ground road. Similarly, currently, AI can only handle some business processes. Enterprises need to first break down the existing workflow, hand over the parts that AI is good at to AI; the remaining parts, including the connection between different AI processes and the processes that require experience - based judgment and emotional interaction, still need to be handled by humans to ensure the completion of the entire task.

Second, path planning requires knowledge of the starting point, the destination, and the highway route map. Similarly, if an enterprise wants to optimize its business through AI, it needs to know its own needs (like the starting point and destination of a journey) and the current capabilities and limitations of AI (equivalent to the highway route map) to find value - creation points at their intersection.

Third, path planning needs to be adjusted dynamically. The progress of AI technology is like the continuous expansion of the highway: a section that is not covered today may be open to traffic tomorrow; a highway entrance that is on the east side today may have a closer one added on the north side tomorrow. Similarly, as AI capabilities improve, enterprises' process reconfiguration and the division of labor and cooperation between humans and AI also need to be continuously adjusted.

Based on my observation, most enterprises are still at the stage of directly applying AI tools. They have neither broken down the workflow nor evaluated the suitability of AI capabilities for business needs, and have failed to form the minimum viable flywheel of input - data - benefit, resulting in outcomes that fall short of expectations.

02

Who Should Lead the Implementation of AI

As mentioned above, the implementation of AI applications requires both knowledge of AI and industry insights. However, industries vary widely, and it is difficult to have both. So, either people who understand AI learn and transform industries, or people in the industry learn AI tools and bring the AI capabilities back to transform their own industries.

Path 1: Let people who understand AI "enter the industry" —

The rise of Forward Deployed Engineers (FDEs).

The "Forward Deployed Engineer (FDE)" model that has emerged in Silicon Valley in recent years is a representative of this path. This model was first explored by the data - analysis company Palantir: its core is to send engineers familiar with AI and data - analysis technologies to client enterprises, often for several months or even half a year. The task of these engineers is not to sell products but to go deep into the front - line of business, understand the information of enterprise production and operation, and finally find value - creation points that match the needs and pain points of the enterprise within the capabilities and limitations of AI.

Today, Palantir's FDE model has become the "AI implementation model" admired in Silicon Valley. These forward - deployed engineers, who have both AI technology and industry insights, have become the most favored group of entrepreneurs by investors.

Path 2: Let people who understand the industry "master AI" —

Difficulties and turning points.

Another path is for industry practitioners to learn and master AI tools and then bring AI capabilities back to their own businesses. The Massachusetts Institute of Technology report mentioned at the beginning of this article found that although only about 40% of companies are paying users of AI tools, more than 90% of companies' employees use AI tools at their own expense to improve work efficiency. The author calls this the "shadow AI economy".

The "shadow AI economy" is at the individual employee level for certain specific tasks, rather than a systematic application at the organizational level. It lacks coordination among employees and is not adapted to industries and enterprises. On the one hand, this shows that in the businesses of most companies, AI can indeed improve efficiency in many processes. It is imaginable that if these tools can be systematically adopted at the enterprise level, and their memory and context functions and adaptability to enterprise scenarios are enhanced, the effects can be further magnified. On the other hand, matching AI tools according to the needs of enterprise business processes requires the evaluation of business processes. This work can be completed in a distributed manner from the bottom - up, and a certain degree of adaptation and customization may be required during the process.

In the past, the technical threshold of AI was high, and the iteration speed was fast. It was very difficult for industry people to learn AI tools to empower and transform industries. However, in the past year, the explosion of AI programming has made this path possible.

03

AI Programming Activates Industry - Driven Self - Transformation

With the development of AI technology, AI programming tools have become more and more powerful, greatly reducing the threshold and cost of software development and making it "accessible to the public". In the past, development work that required professional programmers several months to complete can now be done by zero - based users who describe their needs in natural language and generate code through AI programming tools to develop a product prototype that can at least verify the concept and test user feedback.

Microsoft CEO (Chief Executive Officer) Nadella and Google CEO Pichai have both publicly stated that about 20% - 30% of the software code currently generated by their companies comes from AI. Amazon Web Services CEO Garman even said that 75% of AWS's code is now generated by AI. As AI technology advances, the proportion of AI programming will continue to increase. Industry leaders such as NVIDIA founder Huang Renxun and OpenAI CEO Altman have all predicted that in the future, programming will no longer require professional languages such as C++ and Python, and "natural language as code" will become the norm.

This change means that the core driving force for AI implementation is likely to shift from "pushed by technical experts" to "self - created by industry practitioners". Path 2 mentioned above becomes feasible. Industry people no longer need to wait for AI experts to "come and transform" but can actively learn, master, and use AI programming tools. According to the specific scenarios, needs, and pain points of the industry, they can find and build the minimum viable flywheel of AI applications in some business processes, solve specific problems, and create immediate and visible value.

In particular, AI programming is expected to enable small and medium - sized enterprises to become the main force in AI implementation. Compared with large enterprises, small and medium - sized enterprises do not need to coordinate multiple levels of departments to promote AI transformation. Often, one manager with two or three core backbones can determine the plan, and the decision - making and iteration speed are faster. Moreover, small and medium - sized enterprises have fewer business processes. Even if they need to make up for digital "lessons", they can directly build a digital system adapted to AI from scratch without transforming complex legacy systems, and the difficulty and risk are often smaller. Small and medium - sized enterprises may have been at a disadvantage in terms of talent in the past, but AI programming tools have greatly alleviated this problem.

04

Conclusion

The implementation of AI is not an overnight "disruption" but a process of gradual calibration and adaptation between AI technology and industrial needs. For enterprises, at this stage, there is no need to be obsessed with "full - process AI". They can choose to focus on scenarios with "small entry points, high adaptability, and high returns", find the minimum viable flywheel where AI and business match, and then use AI programming tools to test and refine functions, reduce implementation costs, and thus gain internal support.

AI programming tools are becoming more and more powerful. Even if they still have limitations and flaws today, they are improving at a fast pace, enabling more and more people to use programming to solve problems and create value. For individuals, in the AI era, the most important thing is no longer to master knowledge but to have vision and creativity. Vision means being able to see unmet needs, pain points, and opportunities in industries, work, and life, and creativity is to use new technologies to come up with better solutions to problems. One of the important paths for the implementation of AI applications is to encourage employees to learn AI programming tools to transform and improve their work, and then transform and improve their companies and industries.

When more and more industry practitioners can develop software in natural language and enterprises can quickly test and optimize AI solutions, AI can truly become a productive force driving the progress of all industries. However, even so, AI is still just a partner for co - evolution rather than a panacea.

(The author is the Managing Director of Liwa Information and the Chairman of the Stanford Growth and Innovation Circle)

This article is from the WeChat official account "Semi - cooked Finance" (ID: Banshu - Caijing). Author: Li Junjie, Editor: Mark. Republished by 36Kr with permission.