The boss makes a wish to AI, while the employees tell AI to get lost | Shenke
By | Zhao Rui
Edited by | Zhang Yuxin
The Bosses' "Wishing Well"
"Do you know how Elon Musk manages his employees?" At this question, Hu Xinlan sensed something wrong immediately.
The boss sitting opposite her came from a heavy industrial enterprise with more than 10,000 employees. On social media, he had read the news that Musk used Agents to replace human employees for work, and the idea of doing the same thing rose in his mind. He said frankly to Hu Xinlan: "Can you help me lay off the entire information technology department? I will only communicate with Agents every day."
Among the large number of bosses Hu Xinlan has contacted, the wish of "laying off employees" is widespread. They come from different industries such as finance, manufacturing, and the Internet, with company sizes ranging from a few hundred to tens of thousands of people. Some tentatively brought up the topic: "Don't you think there is some redundancy in our company's staff?" Others threw out the ultimate question: "What's the maximum number of employees we can lay off?" The boss who followed Musk's example was the most radical one she had ever met.
There are also some bosses who did not explicitly mention layoffs, but hoped that AI could monitor every move of their employees. "Can you install an Agent for me to record what each employee does every day, what output they produce, and what stage the project has progressed to? It would be best if I can monitor and call these Agents at any time, and then calculate the input-output ratio of each person based on the complete logs."
Hu Xinlan knew clearly in her heart that the bosses who put forward such demands just haven't made the layoff plan public for the time being. They hoped to collect enough data on human efficiency first, and then "lay off employees as accurately as possible" to avoid "cutting off the main artery of the company by mistake".
Hu Xinlan, who deals with "layoff" demands every day, does not work in human resources. Her job is one of the most popular positions in the AI circle right now — FDE, which stands for Frontline (Frontier) Deployment Engineer.
At present, AI is developing at an amazing speed: models are iterating extremely fast, applications are emerging in endlessly, and financing for start-ups has not slowed down. However, this speed and popularity have not been successfully transmitted to those enterprises that hope to use AI to improve efficiency and innovate production. Accenture once surveyed 163 Chinese enterprises in seven industries, and the results showed that only 9% of the enterprises have achieved improvements in production efficiency, revenue and profit from AI applications.
The profession of FDE emerged exactly from this "gap between technological development and technology implementation". It was first proposed by the Silicon Valley company Palantir, then carried forward by OpenAI and Anthropic, and was brought to China from Silicon Valley in 2026. This job is like a bridge between AI and enterprises. FDEs are stationed at the client's company, deeply understand the client's business needs, and then implement AI capabilities such as large models and Agents for them in a targeted manner, so as to narrow the gap between AI and real work scenarios.
But in China, the work of most FDEs is awkward — many demands from enterprises are as unrealistic as "making a wish".
Once, Hu Xinlan was introducing her company's AI products to clients from different industries at a meeting, when a boss suddenly took over the conversation and described the great significance of AI to human beings with deep emotion: "It's not just work. When can our physical and life health data be handed over to AI? Let silicon-based creatures replicate and benefit us carbon-based creatures. You should figure out a way to make this happen."
After the words, no one in the room spoke. After Hu Xinlan tried her best to manage her expression, she took control of the situation: "This requires careful planning for the long term." Then she ended the meeting in a hurry, preventing the demand from crossing over from work efficiency improvement to transforming human life.
Julie, an FDE entrepreneur, has similar troubles. Not long ago, she received a promotional video of a product called "AI CMO" from an e-commerce client: as long as you enter the company's website, the product can realize the whole-process marketing work from copywriting, operation to delivery, and the process is extremely smooth. The soundtrack of the promotional video is exciting, and the boss's attitude is even more enthusiastic than the video: "If this can be realized, I am willing to buy it at a high price!"
Julie's heart tightened, "I never thought that the pie drawn by AI companies would now have to be realized by me." She knew clearly that if AI was only shown the information on a company's website, it would never be possible to achieve the effect in the promotional video, "It can automatically post advertisements, but it cannot guarantee that the information sent to users is valuable."
This is exactly the embarrassment brought by another kind of mismatch: enterprise bosses do not know what they really need, nor what AI can achieve. "For customers, we are just a wishing well," Julie said.
In fact, behind these "wishes" lies the deep anxiety of a large number of enterprise bosses: there are more or less inefficiency problems in many production links of the company. More importantly, they are extremely afraid of being left behind by the times for missing the AI train.
Not only enterprise bosses, but on the other side of AI, those AI companies with technology and products also have their own anxieties.
For example, Tencent launched WorkBuddy in March this year, and ByteDance and Alibaba quickly followed up later. ByteDance launched Doubao Work, and Alibaba launched Qianwen Office, focusing on the enterprise-level market. The three giants have all directed their competition to the AI office entrance, but in fact, the functions of office Agents are not very different, and there is no technical barrier. How to make customers choose themselves has become one of the most headache problems for the big factories at present.
As a result, the big factories began to set up FDE teams intensively, hoping to use this service model to bring their AI into enterprises. FDE has also become a new card for big factories to ease their commercialization anxiety.
A sales manager from Alibaba Cloud told 36Kr that in the face of WorkBuddy's first-mover advantage in the market, his way is to give enterprises free usage quotas first, and then send more people to do FDE work. "Now it's all about who is more willing to bet. Many employees in the company, including technical staff and pre-sales staff, have been transferred to FDE positions. Products with strong strategic positioning such as Bailian (Alibaba Cloud's MaaS platform) also have dedicated FDE teams."
WorkBuddy has also recently established its own FDE team. A pre-sales manager at Tencent was initially unwilling to let his team do FDE work, but because he was worried that customers would be poached by competitors, he also chose this way to "hold on" to customers.
Big factories want to use FDE to compete for the market, and small and medium-sized enterprises want FDE to turn AI into real efficiency. The anxious two sides are betting on FDE at the same time.
The problems faced by FDE are not unique to the AI era. They are just a repeat of the SaaS era 20 years ago. The tricky part is that as technology can take over more work, the gap between technology and implementation will also become larger.
It seems that the problems of the past are reappearing in a more complex way.
The Tug-of-War Under the Shadow of Layoffs
Facing the wishes of customers, Hu Xinlan never refuses, but uses "magic to defeat magic".
For the customer who wanted to follow Musk's example, Hu Xinlan first nodded sincerely, and then deduced along the other party's logic: "I understand, but in this case, you need to read tens of thousands of AI execution logs every day. Do you think you can spare the energy for that?" The boss hesitated for a moment, and Hu Xinlan kept adding weight: "Moreover, once employees use other tools, you will not get the context anymore."
If the reasoning is not enough to calm the other party down, Hu Xinlan will use her killer trick: "If implemented in this way, the daily Token cost may reach 10,000 US dollars." After this sentence, the customer calmed down instantly. Hu Xinlan knew it was time to talk about realistic solutions.
The company Hu Xinlan works for is small in size. While working as an FDE, she also undertakes sales work. In order to win orders, she does not have much stance to argue, but needs to use a way acceptable to the customer to make the other party understand that "it's not that it can't be done, it's just that it's not suitable for you right now."
Lin Ling is an FDE at a leading big factory. In the past six months, she has found that customers' requirements for the construction period of AI projects are gradually shortening. The project delivery period that used to be calculated in years no longer exists. The four-month delivery period is often pressed down to three months by customers. And as the project deadline approaches, customers will have more and more new ideas they want to realize. In response to these, she will not satisfy all of them, but turn to the original emails and contracts to check one by one what was agreed at the beginning, "just like checking chat records when couples quarrel".
Compared with small start-ups, big factories do have more ways to handle their clients: customers are usually already bound to their cloud services, and the migration cost is relatively high, so they will not easily cancel the cooperation. When she is so angry by the customer that she is about to lose her temper, Lin Ling always reminds herself: We are already an old married couple, how could we get divorced?
After aligning the expectations with the boss, FDEs encounter new difficulties: the boss's eagerness for AI cannot be empathized by the employees.
Pan Kelin is the second-generation owner of an auto parts factory. Troubled by the inefficiency of the factory, he tried to use AI to help transform it. At first, every time he developed an AI tool, he would send the plan to the WeChat group first, then call more than a dozen department heads to a meeting to solicit opinions, but the meeting usually went to two extremes, "either there is no problem at all, or there are nothing but problems".
Sometimes, employees all nodded and said, "No problem, we will all follow the leader's instructions." But when it comes to actual implementation, problems pop up everywhere: some people are busy explaining how this will increase their workload and difficulty, some say there are not enough resources and inter-departmental cooperation is difficult, and even bring up difficulties at home, in short, "it is really impossible to push forward".
Before becoming the second-generation factory owner, Pan Kelin never worked in any company, so he "did not understand the workplace". After several meetings, he gradually realized the perspective difference between the boss and the employees: the boss cares about the efficiency and revenue of the company, while the employees care whether the work will be increased, whether they will take more responsibilities, and more directly, whether they will be laid off because of AI.
Before becoming an FDE, Hu Xinlan never expected: "What is more difficult than making AI tools is the human game between bosses and employees."
When asking the client's technical staff for cooperation, Hu Xinlan often runs into obstacles. The first time she went to the R&D department, she asked everyone what needs they had for AI to improve work efficiency. The R&D staff all stared at their computer screens, and only the crackling sound of keyboards responded to her. Two minutes later, the employee at the far end replied: We have a lot of work recently, no time to migrate the tools.
Later, at the demand alignment meeting, the R&D staff asked sharply: "Is your product as easy to use as Claude Code? After switching to it, how can you guarantee the work progress?" Hu Xinlan soon realized that "being busy" was more like an excuse, the main reason might be the R&D staff's "tool obsession", they already have their own favorite tools in mind.
There are also people who simply give up cooperation. Once, Hu Xinlan proposed to provide a free Agent seat for the IT employee sitting next to her. The employee spoke calmly and replied: No need, Doubao is good enough for me.
If the AI efficiency improvement measures touch the jobs of employees, it will even lead to a dead end.
Julie once served a small bank. The boss had a clear demand: he hoped AI could optimize the loan approval process.
To achieve this, Julie first needed to understand the business process from front-line sales staff, make an operation pilot, and then get the budget from the business manager with the results. In the loan approval process, different banks have different risk preferences and calculation standards. Julie realized that this is more critical than the general process, and it is highly dependent on human experience.
But the salesperson Julie asked for experience seemed unwilling to cooperate: the more eagerly and specifically Julie asked, the more hesitant and vague the salesperson's response was. Until a rhetorical question: "What on earth are you doing this for?" Julie then realized that what she wanted to understand was the most important knowledge barrier of the salesperson in their work. When trying to let AI optimize this work, the salesperson immediately had the fear of being laid off.
Later, the meetings between Julie and the employees of this company began to be postponed for various reasons. The first version of the Agent ran into an error, and Julie sent it to a salesperson asking for correction, but never received a reply.
The boss wants to embrace AI but does not understand the business; front-line employees understand the business but do not want to embrace AI; most of the middle managers show support, but the premise is that the effect is obvious and they do not have to take any risks.
Being torn between different roles, Julie found that a boss's one sentence of "supporting AI" is far from enough. The biggest contradiction here is: the boss who wants to use AI to replace people exactly needs those employees who can be replaced to teach AI "how to work" first.
Paying Off Old "Debts"
On the first day after FDE entrepreneur Edison connected the Agent for the client, he received nearly 300 error messages.
His client is a garment manufacturing company. In the past, the boss needed to spend at least two hours every day checking production information. All the production information of the factory was counted in a huge Excel sheet. All orders of hundreds of garments were arranged in rows, and the production status of different links such as fabrics, zippers, and hangtags were distributed in columns, which were updated and maintained by dozens of production line leaders together.
Although the factory has an ERP system, it only records the most basic static data: product type, customer, amount, and delivery time. The data involved in actual production, such as the production information mentioned above, are not in this system. These data are often fragmented, changeable, and scattered in local Excel sheets, chat records and human brains of different departments.
Edison originally planned to connect an Agent to this Excel sheet to help the boss identify production risks. As a result, because the leaders always failed to maintain the sheet in time, the Agent kept raising alarms. When employees are busy and do not update it immediately, AI will judge that the order will be delayed; if there is one less space in the text, AI will judge that there is a production anomaly... The flexibility that human beings can understand is all regarded as a problem by AI.
To let AI enter the core production/work process of an enterprise, one of the keys is to have data that can be called at any time. But for many enterprises, this is an old debt that has been buried for a long time.
When Pan Kelin first used AI to transform the factory, he also found that there were data loopholes everywhere in the factory. For example, ten workshops had ten sets of daily report forms, and everyone wrote them differently. If the recording methods are not unified, the information aggregation across workshops will be inefficient, and a series of subsequent links such as production scheduling will also slow down. When AI intervenes, let alone improving efficiency, it is even difficult for AI to accurately read this information.
As a result, the AI efficiency improvement originally envisioned by the enterprise is stuck, and it has become "paying off debts" — going back to the "data cleaning and governance" link.
For Pan Kelin, this means re-sorting out the standards of daily report records from the source. He spread out the ten versions and compared them item by item, asked the workshop leaders while discussing with AI, and spent two weeks re-organizing the recording specifications for the workshop daily reports. And this work is only the beginning, "It is not known how long it will take to sort out all the scattered data in the past."
What surprises many FDEs is that before considering accessing AI, a large number of enterprises never thought that "non-standard data" was a problem at all.
An FDE