The biggest misconception among corporate leaders about AI: trying it out rather than using it to disrupt.

"We've also adopted AI, developed intelligent customer service and knowledge bases. Why doesn't AI seem as miraculous as it's advertised?"

"We established an AI team and developed an industry large - model, but we still haven't achieved the so - called cost reduction and efficiency improvement. The leadership has reduced AI investment this year, and the technology department will also lay off employees."

This is the current situation of enterprise AI application that Lao Yang recently saw in the group. He had an in - depth discussion on this issue with the CIO of a well - known enterprise, let's call him Mr. C for now. Mr. C finally said, "Actually, the biggest misunderstanding in current enterprise AI application is that leaders are just having a taste rather than making management subversion!"

Yes, this statement accurately hits the most hidden and common "false proposition" in current enterprise AI application - the "try - it - out" shallow - level application. And this "try - it - out" mentality is the root cause of AI projects being "much ado about nothing" and having many unfinished projects.

Why is current enterprise AI application just "trying it out"

In Lao Yang's view, "trying it out" is not something new. It has appeared in every stage of enterprise digital transformation (from cloud computing to big data and then to the middle - platform). Its core characteristics can be summarized into three "taken - for - granted" assumptions:

First, taking it for granted that achieving "intelligence" in an enterprise is just "tool replacement"

This phenomenon is very common. For example, when a leader sees others using AI to write copy, they ask the information department, "Let's also adopt an AI writing tool"; when they see other enterprises using AI customer service, they immediately launch an intelligent customer service system. Let's give a real - life example. A leader of an enterprise saw a self - media video saying that "AI customer service can replace human labor", so they immediately launched an AI customer service project, disbanded the human customer service team, and only retained two employees. As a result, after the system went live, due to inaccurate intention recognition and non - updated knowledge bases, the customer complaint rate increased by 50%, leading to a large number of old customers leaving.

This is a typical case of regarding AI as a new "plug - and - play" tool, aiming to be "quick" and "convenient" to replace a link in the existing process, rather than re - examining whether the entire business process is reasonable and needs to be reconstructed. Let's think back. Haven't we often encountered such problems in the digital age?

Second, taking it for granted that using "AI" will immediately lead to cost reduction and efficiency improvement

When an enterprise introduces an AI project, the questions leaders always ask are: "How many people can be laid off? How much money can be saved?" In the eyes of leaders, AI is like a money - printing machine, and they expect to see profits immediately after investment. This is a typical phenomenon of pursuing "short - term ROI" rather than "long - term capabilities". This phenomenon is common in most traditional enterprises. It should be noted that the core value of an AI project is not realized through "saving human resources" once, but through building a system capability of 'data + algorithm + human - machine collaboration' and continuously optimizing in long - term operations to achieve cost reduction and efficiency improvement. This is a capacity - building process, not a short - term investment.

Third, all construction aims to satisfy the leader's "taken - for - granted" assumptions

For example, when an enterprise leader sees others developing an industry AI large - model at an industry summit, they come back and ask the information department, "We must also develop one." Then the information department spends half a year or more, investing a large amount of manpower and material resources to complete the so - called "large - model". Then the leader has something to show off in the industry, but in fact, the business department doesn't use this model at all, or there are a lot of problems when using it.

So it's not hard to see that AI projects have become "image projects" to satisfy the leader's "trend - chasing" psychology, rather than "substantive projects" to solve real business pain points. The success criterion of a project is "the leader thinks it's okay", rather than "the business department thinks it's useful and can improve efficiency". Don't we also find this phenomenon familiar? For example, when visiting an enterprise for exchange and learning, the leader's presentation materials are very impressive, but when communicating privately with the business department, they always shake their heads and sigh, saying that the system has various problems and deficiencies.

What is subversion?

The word subversion is not unfamiliar to everyone in the AI era. Just like the most common saying before: The essence of digital transformation lies in transformation, not digitalization itself. That is to say, if an enterprise wants to truly enjoy the dividends of the AI era, it's not as simple as buying a few tools and launching a few projects. It's a systematic change. Only a thorough change can achieve subversion in the true sense.

So the question is, in which aspects should an enterprise use AI for subversive change? Lao Yang gives the following examples:

The most common area where enterprises apply AI currently is process automation. For example, using OCR to recognize invoices and then automatically submit them for approval. The "try - it - out" approach is to automate the manual review step to replace the manual review process. It seems to improve efficiency, but the truly subversive change is that an enterprise doesn't use AI to "optimize" the old approval process, but directly questions the necessity of this process. For example, why is approval needed? Is it because of credit concerns? Then can we use AI to build a "credit scoring system" based on real - time data and historical behavior analysis, so that 95% of low - risk orders can be automatically approved without approval, and only the 5% high - risk orders trigger manual intervention? This is process re - engineering, and the efficiency improvement is on a different level.

Another example is at the data application management level. The most common "try - it - out" approach for most enterprises is to export data from existing Excel, ERP, and OA systems and feed it to the AI model for training. The subversive change measure is that when designing the business system, we must consider 'whether AI can understand and use it'. Data is no longer the result of business, but becomes the "fuel" to drive business decisions. This is to fundamentally change the way data is produced and its flow, so that AI can not only analyze data but also "independently discover" new business opportunities in the data, re - define the business model.

So it's not hard to see that the real value of AI is not about improving efficiency or laying off employees, but about reconstructing the entire business model of an enterprise. Reconstruction means subversion.

How difficult is subversion?

This is a very real problem that deeply troubles enterprise managers. I think CIOs who have experienced digital transformation construction should have the same feeling. In the AI era, not to mention the redistribution of power, interests, and risks, just the iteration and upgrade of the historical technology platform are difficult for most traditional enterprise leaders to bear.

For most traditional enterprises, after years of traditional digital construction, there are numerous systems and platforms criss - crossing. It's very difficult to implement AI on the basis of traditional digitalization. Some people say it's simple, just connect the interfaces to get the data. But the problem is that some software companies are reluctant to provide data interfaces because they are also promoting their own AI products, which creates technical barriers. Even if the software companies provide the interfaces, the enterprise will have to pay a high access cost. After accessing, they will find that the data quality cannot meet the requirements of AI. What should they do? At this time, they have to do the data cleaning and governance again, which they were reluctant to invest in before. If they don't want to use historical data, they can reconstruct a new system that meets the requirements of AI. But the problem is that if they don't build a platform base, there will be various isolated islands later. And the cost of reconstructing an AI platform base is also a significant investment, not to mention the planning and design risks of this platform. Hundreds of thousands of dollars can be easily spent, which is difficult for most current traditional enterprises to bear. So it's not hard to see that even technical problems are difficult to solve, let alone the deep - rooted management problems in enterprises.

Advice for enterprises

Lao Yang has also mentioned the correct way for enterprises to apply AI in previous articles. Here is a simple summary:

• Put aside the "try - it - out" mentality and establish "long - termism".
• Start with "data governance" instead of "model selection".
• Position AI as an "enabler" rather than a "replacement".
• Tolerate trial - and - error and make small and rapid progress.
• Cultivate "translator - type" talents.

Most current enterprises are still at the stage of "watching the show" or "trying it out" when it comes to AI. They are being carried forward in the media storm, lacking independent and correct judgment. If they want AI to truly play its value, enterprises need the courage to "subvert", put aside their obsession with "coolness", and tackle the tough issues of data governance, process re - engineering, and organizational change. AI will naturally subvert those managers who don't understand or know how to use it.

This article is from the WeChat official account "Xiangjiang Digital Review" (ID: benpaoshuzi). The author is Lao Yang, and it is published by 36Kr with authorization.