The difference between chasing a trend and using it wisely

Image source: Internet (please contact us for removal if this infringes any rights). Partially AI-generated.

Over the past two years, I have observed and studied quite a few companies—from large corporations that have been listed for many years to small companies founded only a few years ago, from medium-sized firms with hundreds of employees to startup teams of just a dozen or so. I have developed an increasingly strong impression: everyone is talking the AI talk, and everyone is embracing AI.

A hat is a clever thing. Put it on, and you immediately acquire a certain sense of identity. If "large models," "agents," and "AI-native" do not appear in your meeting slides, you seem to be a notch lower; if the boss does not shout a couple of lines about "fully embracing AI" at the annual meeting, he seems unable to answer for it.

This hat at least looks cheap. Compared with restructuring a production line or overhauling an entire set of processes, buying an AI account and signing an "agent" contract does indeed look far more respectable on the financial statements.

But a hat is ultimately a hat. It sits on the head, not in the body. It blows off in the wind, must be held on when the crowd is thick, and the first thing you do when you get home is take it off.

The problem is that many companies treat putting on the hat as having grown a new organ.

How the Hat Gets Put On

First, consider a set of not-so-respectable numbers.

A white paper survey released by Deloitte China in October 2026 shows that 92% of enterprises have already carried out AI application pilots, but only 8% believe their AI investment has achieved financial and strategic returns, and only 4% truly believe they have completed organization-level transformation.

McKinsey's standard is slightly more moderate: 88% of enterprises globally are using AI, but only 39% are getting real money out of it. BCG offers an even sharper denominator: enterprises that achieve scaled AI value account for just 5%.

In other words, almost everyone has a hat, but those who truly rely on it for shelter are in the single digits.

So why do those 92% still wear it? The answer is not in technology; it is in the air. FOMO—the fear of missing out—is the most effective fuel of this round of AI narrative.

A boss in the building materials business spent 800,000 yuan on an "AI intelligent marketing system," and the supplier promised a tenfold increase in customer acquisition efficiency. After a year, it had increased 0.8-fold, and he had lost 600,000 yuan.

Another restaurant chain adopted AI customer service for only one reason: "Competitors are all using it." As a result, the answers did not match the questions, customer complaints tripled, and they sheepishly switched back to manual service.

There is also a boss in Dongguan who, after attending an AI summit, got fired up and spent 8,000 yuan on an enterprise-edition large-model account, promoting it across sales, customer service, and R&D groups—only to have administration delete the account a month later.

The most extreme version comes from a software company in a fifth-tier city: the boss spent about 700,000 yuan procuring a so-called "FDE + localized deployment" AI system, and ultimately received standardized software plus high-priced tokens and agency operation services. The account was banned, salary data leaked, employees resigned, and the supplier became unreachable.

These stories share a common starting point: they did not go looking for medicine because it hurt, but because they saw others taking medicine.

A more hidden layer occurs inside organizations. When "using AI" becomes a top-down task, the hat turns from decoration into assessment.

A leading domestic tech giant forcibly promoted its self-developed AI programming assistant, stipulating that programmers' "AI code generation ratio must reach at least 30%, otherwise quarterly performance will be deducted."

To meet the target, engineers began deliberately expanding logic that could be written in 10 lines into hundreds of lines of redundant code, with useless comments attached, and the AI adoption rate instantly soared to 45%.

Meta once launched an internal "token consumption leaderboard," where the more consumed, the higher the ranking, with those at the bottom facing layoff risk. The first place burned nearly 500,000 dollars' worth of tokens in a month, and the leaderboard was then quietly taken down.

Uber opened top-tier AI programming tools to all 5,000 engineers, burned through the entire year's AI budget in four months, and was forced to urgently introduce a quota system.

The boss wanted to use AI to stop slacking, but employees used AI to learn more sophisticated slacking. The hat is on, but under the brim, there is no real connection between the person and the hat.

What Happens After Putting It On

I have no intention of saying AI is worthless. On the contrary, the cases that truly broke through have data so good they cannot be ignored.

A steel enterprise is a sample that made me rethink the term "traditional enterprise." In October 2023, this private steel company invested 1 billion yuan to launch a full-process smart factory, and in 2024 and 2025 added more than 700 million yuan over the two years.

After the first phase of intelligent applications went live, it reduced costs by nearly 200 million yuan per year, increased overall labor productivity by 30%, and reduced the product defect rate by 13.89%.

The intelligent upgrade of the finished steel rolling inspection stage alone saves more than 8 million yuan in costs per year. What is truly interesting is not the numbers, but what happened behind them: there is an old saying in steelmaking, "the heat is all in the eye." What color the flame at the furnace mouth corresponds to what temperature the molten steel is at—only master workers who have done it for decades can grasp this skill.

Rizhao Steel's approach was not to let AI "replace" the master workers, but to use visual large models to identify signals such as flame splashing and slag overflow, with accuracy stable above 99%, "translating" the masters' feel into actions algorithms can execute.

The same goes for equipment operation and maintenance. After localized deployment of large models, the time to prepare maintenance plans was compressed from 3 hours to 0.5 hours, the spare parts mismatch rate dropped from 30% to 5%, and the maintenance knowledge reuse rate rose from 40% to 85%.

The data from AUO's Suzhou factory is equally solid: 342 AI projects cover R&D, production, quality inspection, and operations and maintenance; the numerical control rate and networking rate of key equipment are 100%; production efficiency increased by 29%, the product defect rate decreased by 6%, and production costs decreased by 15%.

A company in Kunshan doing intelligent spinning achieved a yarn break detection accuracy of 99.99%, reduced electricity consumption per ton of yarn by 5%, and promoted the solution to more than 50 textile enterprises worldwide.

The common point of these enterprises is not that they "used AI," but that AI found a specific position in their bodies that had been hurting for a long time. That steel enterprise did not launch the project to "embrace AI," but because in blast furnace batching and flame judgment, relying on the experience inheritance of master workers was almost no longer sustainable.

AUO did not do it to write "agent collaboration" in a PPT, but because 342 scenarios were carved out one by one, and behind each scenario were hard indicators such as yield, energy consumption, and downtime pushing it forward.

Conversely, in the cases that "could not keep the hat on," the problem was never AI itself.

A bank, in order to embrace AI, aggressively adopted a pure Agent solution to handle all standardized customer service inquiries. As a result, operating costs soared 3.2 times, collaboration among multiple Agents actually slowed response speed, large-model hallucinations increased the error rate by 30%, and customer complaints continued.

Hongqi Chain in 2025 high-profile announced R&D of a robot unmanned sales system, and its stock price hit the daily limit the next day. More than a year later, it was sadly terminated. The announcement was very candid: insufficient experience in technical adaptation and operations management, research risks in algorithm stability and data security, high mechanical arm grasping error rates, and visual recognition systems prone to deviation.

An earlier lesson came from across the ocean: IBM Watson Health invested more than 4 billion dollars and was ultimately sold cheaply for less than 1 billion dollars; Zillow Offers lost 500 million dollars due to AI decision-making errors and laid off 2,000 people; Klarna replaced 700 customer service staff with AI and rehired 15 months later.

Looking at these cases together with Rizhao Steel, the dividing line is not the level of technological advancement, nor even the amount invested.

IBM's algorithm is not bad; Watson's diagnostic accuracy in isolated environments is quite high. The problem is that IBM grafted Watson as a "tool" onto existing medical processes. Doctors' way of working did not change; there was just an additional "suggestion system," and most doctors chose to ignore it.

That conclusion from a RAND researcher is worth every boss preparing to "put on the hat" copying onto his desk: most AI failures are not because the algorithms are bad, but because the enterprise has not prepared the organizational foundation for AI to succeed.

Enterprises spend 90% of their energy on models, but models account for only 10% of success.

The Empty Place Under the Hat

I increasingly feel that "tool mindset" is the gentlest and also most fatal trap in this AI boom. It creates an illusion: AI is something that can be "added on," like installing a new air conditioner in an old house—the walls do not need to move, the wiring does not need to change, and with a press of the remote, cool air arrives.

But AI is not an air conditioner. It is more like a renovation: you first have to decide which walls to tear down, where the water and electricity should go, and how people will move around inside.

If you just move the most advanced air conditioner into an unfinished room with no wiring, it cannot even be plugged in.

The MIT NANDA report puts this more bluntly: 60% of enterprises evaluated enterprise-level AI tools, only 20% entered pilots, and ultimately only 5% reached the production environment.

Most failures stem from rigid workflows, lack of contextual learning capability, and mismatch with daily operational needs.

In other words, it is not that AI lacks capability, but that AI has been placed into a structure that lacks the capability to let it function.

Data is another empty place. Data shows that 85% of AI project failures are rooted in data problems.

A financial company spent heavily to build a "business analyst Agent," but the internal databases were typical data silos, metric definitions were chaotic, and what the Agent analyzed was structurally wrong, making the conclusions unusable.

If you ask even the smartest brain to read a bunch of ledgers that contradict each other, all it can output is more refined chaos.

Organization and people are the third empty place. A 2026 survey of 2,016 industrial enterprises above designated size by Cheung Kong Graduate School of Business has a detail that illustrates the problem: the AI penetration rate of real-economy enterprises is about 10%, but the adoption rate is 15.5% for large enterprises, 9.1% for medium-sized enterprises, and only 5.4% for small enterprises.

Small and medium-sized enterprises do not lack the desire to use it; their trial-and-error resources, data foundations, and talent reserves are simply not on the same order of magnitude. Among enterprises that have already used AI, 50% report that AI is a "supplement/enhancement" to jobs, while 31.4% feel "replacement" pressure. What is really happening is task restructuring, not simple layoffs or efficiency gains.

This means that for AI to land, people's work content must change accordingly. But most enterprises do this: buy a tool, issue a notice, and let employees "figure it out themselves."

An old accountant accustomed to Excel, if thrown into an AI system, will not feel empowered; he will only feel punished.

There is another number in Deloitte's survey that I can never forget: 92% pilots, 8% seeing financial returns, 4% completing organizational transformation. The gap between 4% and 92% is not a technology gap; it is a process gap, a data gap, and an organization gap.

A hat can be put on in a day, but the gap must be filled shovel by shovel.

It Is Not "Whether to Use It," but "Where It Grows From"

Actually, I am not anti-AI; I am only against the approach of treating AI as a hat to wear.

Behind each of AUO's 342 AI projects is an identifiable production line pain point. Their AI was not "added on," but "exchanged out": using algorithms to replace part of the links that relied on hard-won experience, and using data to replace part of the processes that relied on people watching people.

A hat put on will blow off in the wind; something that grows out will not.

Those companies that fell into pits often failed, before buying tools, to answer three questions: Which specific pain do I actually want to solve? Is AI really more cost-effective than current methods for this pain? To make AI work, which part of my processes, data, and people needs to move first?

If you cannot answer, an 8,000-yuan account and an 800,000-yuan system will end the same way.

AI is certainly worth using, but "worth using" and "suitable for you to use right now" are two entirely different questions.

For a startup of a dozen or so people, the most cost-effective AI may well be a well-written prompt template—not an agent middle platform that demands dedicated operations.

For a medium-sized manufacturing enterprise whose processes have yet to run smoothly, the priority should probably be to clean up its data foundation first, rather than rushing headlong into "AI-native."

For a company whose boss doesn't even know what its customer service gets complained about most often, buying a pricier customer service Agent will only automate answering beside the point.

When the wind stops, the hat will fall—and the person standing there is still the same person.

For enterprises that truly break through, when the wind comes, the hat is merely put on casually; when the wind leaves, AI has already grown into the production lines, the processes, and people's daily judgment—and cannot be taken off.

This is the difference between "embracing AI" and "being embraced by AI."