文 | 影子备忘录
这两年了解观察过不少企业，无论是上市多年的大厂，还是才创立几年的小厂，从几百人的中型公司到只有十几个人的初创团队。都有一个越来越强烈的感受：都在聊AI叙事，都在拥抱AI。
帽子是个很妙的东西。戴上它，你立刻有了某种身份感。会议PPT里不出现“大模型”“智能体”“AI原生”，似乎就矮了一截；老板在年会上不喊两句“全面拥抱AI”，好像就交不了差。
这顶至少看起来便宜，它比起重构一条产线、改造一套流程，买一个AI账号、签一份“智能体”合同，在财务报表上确实体面了不少。
但帽子终究是帽子。它戴在头上，不在身体里。风一吹会掉，人多的时候要扶，回到家第一件事就是摘下来。
问题在于，很多企业把戴上帽子当成了长出的新器官。
帽子是怎么被戴上的
先看一组不太体面的数字。
德勤中国2026年10月发布的白皮书调研显示，92%的企业已经开展了AI应用试点，但只有8%的企业认为AI投入实现了财务和战略收益，真正认为完成了组织级转型的，只有4%。
麦肯锡的口径稍温和一些：全球88%的企业在用AI，吃到真金白银的只有39%。BCG给了一个更锋利的分母：实现规模化AI价值的企业，占5%。
换句话说，帽子几乎人手一顶，但真正靠它遮风挡雨的，是个位数。
那92%为什么还要戴？答案不在技术里，在空气里。FOMO与错失恐惧，是这轮AI叙事最有效的燃料。
一家做建材生意的老板，花80万上了一套“AI智能营销系统”，供应商承诺获客效率提升10倍。一年下来，提升了0.8倍，亏了60万。
另一家餐饮连锁企业上AI客服，理由只有一句：“竞品都在用。”结果答非所问，客户投诉翻了三倍，最后灰溜溜切回人工。
还有东莞一位老板，听完一场AI峰会热血上头，花8000块买了大模型企业版账号，在销售群、客服群、研发群全面推广，却在一个月后让行政把账号删了。
最极致的版本来自一家五线城市的软件公司：老板花约70万采购所谓“FDE+本地化部署”的AI系统，最终拿到的是标准化软件加高价Token和代运营服务，账号被封、工资数据泄露、员工离职，供应商失联。
这些故事有一个共同的起点：不是因为疼了才去找药，是因为看见别人在吃药才去的。
更隐蔽的一层发生在组织内部。当“用AI”变成一项自上而下的任务，帽子就从装饰变成了考核。
国内某头部大厂强制推行自研AI编程助手，规定程序员“AI代码生成占比必须达到30%以上，否则扣减季度绩效”。
为了达标，工程师开始故意把10行能写完的逻辑扩写成上百行冗余代码，附带毫无用处的注释，AI采用率瞬间飙到45%。
Meta内部一度上线“Token消耗排行榜”，消耗越多排名越高，垫底面临裁员风险，第一名一个月烧掉近50万美元的Token，随后榜单被悄悄下架。
Uber给5000名工程师全量开放顶级AI编程工具，四个月烧光全年AI预算，被迫紧急出台配额制。
老板想用AI制止摸鱼，员工却用AI学会了更高明的摸鱼。帽子戴上了，但帽檐底下的人，和帽子之间没有任何真实的连接。
戴上之后呢
我无意把AI说得一无是处。恰恰相反，那些真正跑出来的案例，数据漂亮得让人没法忽视。
有一家钢铁企业是一个让我重新理解“传统企业”这个词的样本。2023年10月，这家民营钢企拿出10亿元启动全流程智能工厂，2024和2025两年又追加超过7亿元。
一期智能应用上线后，每年降本近2亿元，全员劳动生产率提升30%，产品不良率降低13.89%。
单是轧钢成品检测环节的智能化升级，一年节约成本就超过800万元。真正有意思的不是数字，是数字背后发生的事情：炼钢这行有句老话，“火候全凭眼力”，炉口火焰什么颜色对应钢水多少度，这种本事只有干了几十年的老师傅才摸得透。
日照钢铁的做法不是让AI去“替代”老师傅，而是用视觉大模型把火焰喷溅、溢渣这些信号识别出来，准确率稳定在99%以上，把老师傅的手感“翻译”成算法可以执行的动作。
设备运维也一样，本地化部署大模型之后，维修方案编制时间从3小时压到0.5小时，备件错配率从30%降到5%，维护知识复用率从40%提到85%。
友达光电苏州工厂的数据同样扎实：342个AI项目覆盖研发、生产、质检、运维，关键设备数控化率和联网率100%，生产效率提升29%，产品不良率降低6%，生产成本降低15%。
昆山一家做纺纱智能化的企业，把断纱检测准确率做到99.99%，吨纱电耗下降5%，方案推广到全球50余家纺织企业。
这些企业的共同点，不是“用了AI”，而是AI在他们身体里找到了一个具体的、疼了很久的位置。那家钢铁企业不是为了“拥抱AI”才上项目，是因为高炉配料和火焰判断这件事，靠老师傅的经验传承已经快撑不住了。
友达不是为了在PPT里写“智能体协同”，是因为342个场景一个一个抠出来，每个场景背后都有良率、能耗、停机时间这些硬指标在推着走。
反过来看那些“戴不住”的案例，问题也从来不在AI本身。
某银行为了拥抱AI，激进地采用纯Agent方案处理所有标准化客服查询。结果运营成本飙升3.2倍，多个Agent协同反而让响应速度变慢，大模型幻觉导致错误率增加30%，客户投诉不断。
红旗连锁2025年高调宣布机器人无人售货系统研发，公告次日股价涨停，一年多之后黯然终止，公告里写得很坦诚：技术适配和运维管理经验不足，算法稳定性、数据安全存在研究风险，机械臂抓取失误率高，视觉识别系统易出现偏差。
更早的教训来自大洋彼岸：IBM Watson Health投入超过40亿美元，最终以不到10亿美元贱卖；Zillow Offers因AI决策失误损失5亿美元、裁员2000人；Klarna用AI替换700名客服，15个月后重新招人。
把这些案例和日照钢铁放在一起看，分野不在于技术先进程度，甚至不在于投入金额。
IBM的算法不差，Watson在隔离环境中的诊断准确率相当高。问题在于，IBM把Watson当“工具”嫁接在既有的医疗流程上，医生的工作方式没有变，只是多了一个“建议系统”，而大多数医生选择忽略它。
RAND研究员的那句结论值得每一个准备“戴帽子”的老板抄在办公桌上：大多数AI失败不是因为算法不好，而是因为企业没有准备好让AI成功的组织基础。
企业花了90%的精力在模型上，但模型只占成功的10%。
帽子底下空着的那块地方
我越来越觉得，“工具思维”是这轮AI热潮里最温柔也最致命的陷阱。它让人产生一种错觉：AI是一个可以“加上去”的东西，像给旧房子装一台新空调，墙不用动，线不用改，遥控器一按，凉风就来。
但AI不是空调，它更像是一次装修，你得先决定哪些墙要拆，水电往哪走，人住在里面怎么活动。
如果你只是把一台最先进的空调搬进一间没有布线的毛坯房，它连电都插不上。
MIT NANDA的报告把这件事说得更直白：60%的企业评估了企业级AI工具，只有20%进入试点，最终只有5%到达生产环境。
大多数失败源于工作流程僵化、缺乏情境学习能力、以及与日常运营需求不匹配。
换句话说，不是AI没能力，是AI被放到了一个没有能力让它发挥作用的结构里。
数据是另一块空着的地方。有数据显示，85%的AI项目失败根源在于数据问题。
一家金融公司重金打造“业务分析师Agent”，结果内部各个数据库是典型的数据孤岛，指标定义混乱，Agent分析出来的东西结构都是错的，结论根本没法用。
你让一个再聪明的脑子去读一堆互相打架的账本，它输出的只能是更精致的混乱。
组织和人，是第三块空着的地方。长江商学院2026年对2016家规模以上工业企业的调研有一个细节很说明问题：实体企业的AI渗透率约为10%，但大企业采用率15.5%，中型企业9.1%，小企业只有5.4%。
中小企业不是不想用，是试错资源、数据基础、人才储备都不在一个量级上。而在已经用了AI的企业里，50%报告AI对岗位是“补充/增强”，31.4%感受到“替代”压力，真正发生的是任务重组，不是简单的裁人或增效。
这意味着，AI要落地，人的工作内容必须跟着变。但大多数企业的做法是：买个工具，发个通知，让员工“自己琢磨”。
一个习惯了Excel的老会计，你把他扔进一个AI系统里，他不会觉得被赋能，只会觉得被惩罚。
德勤的调研里还有一个数字我一直忘不掉：92%试点，8%看到财务回报，4%完成组织转型。4%和92%之间的那道沟，不是技术沟，是流程沟、数据沟、组织沟。
帽子可以一天戴上，沟却要一锹一锹地填。
不是“用不用”，是“从哪里长出来”
其实我不反AI，我只是反对把AI当成帽子来戴的那种做法。
友达的342个AI项目，每一个背后都有一个可以指认的产线痛点。它们的AI不是“加上去”的，是“换出来”的，用算法换掉了一部分靠经验硬扛的环节，用数据换掉了一部分靠人盯人的流程。
帽子戴上去风吹会掉，长出来的东西不会。
而那些踩坑的企业，往往在买工具之前没有回答三个问题：我到底要解决哪一个具体的疼？这个疼AI真的比现在的方法更划算吗？为了让AI起作用，我的流程、数据、人，需要先动哪一块？
答不上来，8000块的账号和80万的系统，结局是一样的。
AI当然值得用，但“值得用”和“适合你现在用”，是两个完全不同的问题。
一家十来个人的初创公司，最划算的AI可能就是一个写得好的提示词模板，而不是一套需要专人运维的智能体中台。
一家流程还没跑顺的中型制造企业，最该做的可能先把数据底子理干净，而不是急着上“AI原生”。
一家老板连自家客服最常被投诉什么都不知道的公司，买再贵的客服Agent，也只是把答非所问自动化了一遍。
风停的时候，帽子会掉，站在原地的还是原来那个人。
而真正跑出来的企业，风来的时候帽子只是顺手一戴，风走了，AI已经长在了产线里、流程里、人的日常判断里，摘不掉了。
这才是“拥抱AI”和“被AI拥抱”之间的区别。
本文由 AI宝库自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
The Shadow Memorandum
Over the past two years, a number of enterprises, whether large or small, have been on the market for many years, ranging from medium-sized companies of hundreds of people to start-up teams of just a dozen. There's a growing feeling: they're talking about AI, they're hugging AI.
The hat is a wonderful thing. Put it on, you immediately have a sense of identity. The absence of “LLM” “AI agent” “Ai native” in the PPT for the meeting seems to be short; the owner does not shout “A full hug” at the annual meeting, as if it would not be bad.
This, at least, would appear to be cheap, and it would indeed have been quite decent in the financial statements, compared to the re-engineering of a production line, the re-engineering of a process, the purchase of an AI account number and the signing of a “AI agent” contract.
But the hat is the hat. It's on the head, not inside. When the wind blows, when there are more people to help, the first thing to come home is to take it off.
The problem is that many enterprises have seen their hats as new organs.
How did you get your hat on?
Let's start with a set of unorthodox figures.
The White Paper study publish of October 2026 in Deloitte and China showed that 92 per cent of enterprises had implemented AI application pilots, but only 8 per cent believed that AI inputs had achieved financial and strategic benefits, and only 4 per cent believed that the organizational transformation had been completed.
McKenzie's caliber is slightly mild: 88 per cent of businesses worldwide use AI and only 39 per cent eat real money and silver. The BCG gave a more sharp denominator: 5 per cent of the enterprises that achieve a scale-up of AI values.
In other words, the hat is almost one-man, but it really covers the rain.
Why do you wear 92%? The answer is not in technology, in air. FOMO and lost fear are the most effective fuel for this round of AI narratives.
An owner of the building materials business, who spent 800,000 on an “AI smart marketing system”, promised a tenfold increase in efficiency. Over the course of the year, the increase was 0.8 times, resulting in a loss of 600,000.
The other restaurant chain was in the AAI service for the sole reason: “The competition is in use.” As a result, the client ' s complaint tripled and was finally turned back to the manual.
An owner of Dong-chun, who had heard the blood of an AI summit, bought the LLM Enterprise account at $8,000, was fully promoted in the sales, passenger and R & D groups, but asked the administration to delete the account after a month.
The most extreme version came from a software company in a five-line city: the owner spent about 700,000 dollars on the purchase of an AI system known as “FDE+ Localized Deployment”, which eventually received standardized software with higher prices for Token and alternative operating services, with accounts sealed, payroll data leaked, employees separated and vendors lost.
There is a common starting point for these stories: it is not because of pain that they seek medicine, but because they see people taking them.
A more hidden layer occurs within the organization. When "using AI" becomes a top-down task, the hat changes from decoration to testing.
The introduction of self-study AI programming assistants in a major plant in the country requires programmers “to produce more than 30 per cent of the AI code, or to reduce their quarterly performance”.
In order to meet the target, the engineers began to deliberately expand the logic of the 10 lines into hundreds of lines of redundant codes, accompanied by useless notes, and the AI adoption rate jumped to 45 per cent in an instant.
At one point, Meta went on the “Token Consumption List”, and the more it consumes, the higher the ranking, the lower it runs the risk of being laid off, and the first one burned down nearly $500,000 token in the month, and the list was quietly dropped.
Uber opened up the top-level AI programming tool to 5,000 engineers and burned the budget for the entire year for four months, forcing the introduction of an emergency quota system.
The boss tried to stop the fish with AI, but the employees learned to fish with AI. The hat was on, but there was no real connection between the person under the hood and the hat.
What happens when you put it on?
I didn't mean to make AI useless. On the contrary, the data are so beautiful that they can't be ignored.
There's a steel business that's a sample that's got me rephraseing the word “traditional business”. In October 2023, the private steel company set aside 1 billion yuan to start full-process smart factories, with an additional 700 million yuan in 2024 and 2025.
After the first period of smart application went online, it dropped by nearly $200 million per year, increasing labour productivity by 30 per cent for the entire workforce and reducing product failure by 13.89 per cent.
An intelligent upgrading of the steel-finding chain alone will save more than $8 million a year. What's really interesting is not the numbers, but what's happening behind the numbers: the old saying in the steel industry, "Flammation is eye-driven" and how much color of the fire on the mouth corresponds to the steel water, something that only teachers who have worked for decades can find out.
Rather than "replace" teachers, the solar-skinned steel approach is to identify the signals of flames and slags with visual LLM, with an accuracy of more than 99 per cent, and to translate the teacher's sense of hand "translation" into actions that can be performed by algorithms.
The same applies to equipment transport, where after the local deployment of LLM, maintenance programming time was reduced from 3 hours to 0.5 hours, the replacement rate for spare parts was reduced from 30 per cent to 5 per cent, and the re-use rate for maintenance was increased from 40 per cent to 85 per cent.
The data for the Sundarsu plant are equally solid: 342 AI projects cover R & D, production, quality control, transport, 100 per cent control and networking of critical equipment, 29 per cent improvement in production efficiency, 6 per cent reduction in product failure and 15 per cent reduction in production costs.
In Kunshan, a swagger-in business, with 99.99 per cent of the swab detection accuracy and 5 per cent of the ton of gas consumption, the programme was extended to over 50 textile enterprises worldwide.
The common denominator of these businesses is not “use AI”, but that AI found a specific, long-suffering position in their bodies. The steel company was not on the project for “a hug” but for the fact that the furnaces and the flames judged it, the memory of the teacher was barely enough.
Toshida did not write “AI agent Synergy” in PPT, because 342 scenes came out one by one, each with hard indicators of probabilities, energy consumption, and time off.
In turn, in those cases of “failure”, the problem was never in AI itself.
In an effort to embrace AI, a bank has been radical in using a pure Agent programme to handle all standardized client queries. As a result, operating costs have skyrocketed 3.2 times, multiple Agent synergies have slowed the response, and LLM hallucinations have led to a 30 per cent increase in error rates and constant client complaints.
The red flag chain announced in 2025 that no-man-sale robots had been developed, announced that the next day ' s stock price would stop, and that after more than a year, it would end in an open letter: insufficient technical and operational management experience, algorithm stability, research risk for data security, a high rate of mechanical arm grabs and error in visual recognition systems.
The earlier lessons came from across the ocean: IBM Watson Health invested more than $4 billion and eventually sold less than $1 billion cheaply; Zillow Offers lost $500 million and retrenched 2,000 as a result of AI's decision-making mistakes; and Klarna replaced 700 customers with AI and re-engaged 15 months later.
To look at these cases together with sun-ray steel, the division is not based on technological sophistication or even on the amount of inputs.
IBM algorithms are not bad, and Watson has a fairly high diagnostic accuracy rate in an isolated environment. The problem is that IBM has attached Watson as a “tool” to the established medical process, and the doctor's working methods have not changed, except for an “advice system”, which most doctors have chosen to ignore.
The RAND researcher's conclusion deserves to be copied to the desk by every owner prepared to wear a “hat”: most AI failed not because of poor algorithms, but because the business was not prepared to make AI a successful organizational base.
Enterprises spend 90 per cent of their energy on models, but only 10 per cent of their success.
The place under the hat was empty.
I'm getting the feeling that “tool thinking” is the softest and most lethal trap in this round of AI. It creates the illusion that AI is a "plus" thing, like a new air conditioner for the old house, the wall, the line, the remote, the cooler.
But AI is not an air conditioner, it's more like a retrofit. You have to decide which walls you want to tear down, where you want the power to go, how people live inside.
If you just move a state-of-the-art air conditioner into a furry room without a wire, it won't even plug in.
The MIT NANDA report makes this matter even more straightforward: 60 per cent of the enterprises evaluated the enterprise-level AI tool, only 20 per cent were piloted and only 5 per cent eventually reached the production environment.
Most of these failures stem from the rigidity of the workflow process, the lack of situational learning capacity and the mismatch with day-to-day operational needs.
In other words, it's not AI that is incapable, it's AI that is placed in a structure that does not have the capacity to make it work.
Data is another empty place. Data show that 85 per cent of AI project failures are rooted in data problems.
As a result, the internal databases are typical island-specific data, the definition of indicators is confusing, and the structure of what Agent analyses is wrong and the conclusions are not valid.
You let a smart brain read a bunch of books that fight each other, and it's just a more sophisticated mess.
Organization and people, the third empty place. The study of 2016 industrial enterprises above the size of the Yangtze Institute of Business in 2026 provided a detailed account of the problem: the rate of AI penetration in real enterprises was about 10 per cent, but the rate of adoption by large enterprises was 15.5 per cent, medium enterprises 9.1 per cent and small enterprises only 5.4 per cent.
Small and medium-sized enterprises (SMEs) do not want to use them, but rather try to mistook resources, the data base, and the pool of talent at all levels. Of the enterprises that have already used AI, 50 per cent reported that their jobs were “supplemented/enhanced” and 31.4 per cent felt the pressure of “replacement” and that what really happened was a reorganization of the mission, not a simple downsizing or efficiency gain.
This means that if AI is to land, the content of people ' s work must change. However, the practice in most enterprises is to buy a tool, send a notice and let employees “think for themselves”.
An old accountant who is used to Excel, and you throw him into an AI system, he doesn't feel empowered, he just feels punished.
There is another figure in Deloitte's research that I can't forget: 92 per cent of the pilots, 8 per cent of the financial returns and 4 per cent of the organizational transformation. The ditch between 4 and 92 per cent is not a technology ditch, but a process ditch, a data ditch, and an tissue ditch.
The hat can be put on one day, but the ditch will be filled with a piece.
It's not "use or no." It's "where to grow."
I'm not against AI, I'm against the way I wear it as a hat.
There are 342 AI projects in Tomda, each with a identifiable line pain point behind it. Their AI is not “plus”; it is “replaced”, with algorithms that replace part of the experience-based and part of the process of people-oriented.
Put your hat on and blow it off, and the things that grow don't.
And those who stepped on the pit often failed to answer three questions before buying tools: Which specific pain should I solve? Is this Ai really better than the way it is? Which one do I need to move first to make AI work, my processes, my data, my people?
I can't answer that. The $8,000 account and the 800,000 system end up the same.
Of course, AI is worth it, but the “useful” and “fit for you now” issues are completely different.
A dozen individual start-ups, the most cost-effective AI, may be a well-written prompt template, rather than a set of AI agent mid-stages requiring specialized human resources.
In a medium-sized manufacturing enterprise where the process is not running smoothly, the best thing to do may be to clean up the data, rather than rush to “Ai native”.
A company whose owner did not even know what was the most frequently complained about was his own customer's dress, and which bought the most expensive customer's service, Agent, had only automated the answer.
When the wind stops, the hat will fall off and the same person will be standing there.
And the real runaway business, when the wind comes, the hat comes, the wind goes, and AI grows in the line of production, in the process, in people's daily judgment, and can't take it off.
That's the difference between “a hug” and “a hug”.
Auto-fetched and translated by AI Gem Box. All rights belong to the original author.