The first small-town business owners who paid big money for FDEs got badly burned

Wang Hang witnessed a farce in a fifth-tier small city in Shaanxi. He works at a software company with fewer than 100 employees. Since August this year, the boss has invested about 700,000 yuan in an AI system touted as capable of customer acquisition, customer service, recruiting, and decision support.

After deployment, accounts were repeatedly banned, payroll data leaked, and veteran employees left one after another. In the end, even the vendor that sold the system ran off. The boss was so angry that he was hospitalized at one point.

On the surface, this is a story of paying tuition for a lesson. It punctures one of the most common illusions today: after FDE and on-premises deployment became buzzwords, many small-town bosses treated them as miracle cures for cutting costs and boosting efficiency, jumping on the bandwagon without even grasping the basic concepts. In the end, the so-called deep business transformation was just installing a standardized software package.

The following is Wang Hang's account.

The FDE hired at great expense ran off

After the National Day holiday, the boss, who had disappeared for about ten days, finally showed his face. As soon as he arrived at the office, he held an all-hands meeting.

At the meeting, the boss announced two things: first, he had found the AI deployment company that had absconded and recovered part of the money; second, he was firing the company's vice president because the vice president lacked sufficient understanding of AI and had failed to identify risks in time during the company's AI transformation.

Everyone looked at one another, unsure how to react.

My name is Wang Hang. I am 37 years old. Four months ago, I left a big tech company, ended more than a decade as a Beijing drifter, and returned to a fifth-tier city in Shaanxi.

After returning home, I joined a fairly well-known local software company. The company mainly serves government and enterprise clients and performs well; reportedly, it has more than 20 stable clients each month and net profit of over 400,000 yuan. The boss, in his 60s, handles public-facing duties, while day-to-day operations are managed by the vice president.

The boss can barely operate a computer, but that does not stop him from being the most AI-anxious person in the company. He travels around the country to exhibitions and often browses industry peers' video accounts. Whenever he sees content about OPC, AI employees, Agent, AI marketing, and the like, he forwards it to the work group chat. At one point, he forwarded three or four times a day.

A colleague quipped: 'In his imagination, the company could keep only the boss. Customers would come to them on their own, AI would handle pre-sales and after-sales, automatically generate reports, and the boss would only need to keep a payment collection account.'

In this fifth-tier small city, there are plenty of bosses like him. Locally, there are many AI+OPC communities, with owners of supermarkets, storefronts, dental clinics, and businesses across all industries. Even industry association meetings cannot do without the topic of AI.

Once the idea of transformation took hold, all that was missing was someone to implement it. The boss also often forwarded articles about FDE and on-premises deployment to the group, but when it came to finding a service provider, he was a bit wary of big companies in Beijing, Shanghai, Guangzhou, and Shenzhen. He often said at meetings that those companies were indeed technically strong, but too far away to provide a sense of security.

In August, a company the boss considered 'reliable' appeared. At an exhibition, he saw a local company selling an AI system that claimed to have four major functions: AI customer service, AI advisor, AI customer acquisition, and AI interview.

Among them, AI customer acquisition appealed to him the most. According to the company, the system could manage more than 50 accounts at the same time, automatically generate videos, and automatically publish them.

Over the next two months, the boss frequently went to that AI deployment company to negotiate the partnership. I saw him go four or five times on my own.

During this period, many peer business owners came to our company to ask around. Most were cautious: they spent tens of yuan on a Doubao membership, hundreds of yuan on AI courses, and the bolder ones spent only 20,000 to 30,000 yuan to hire someone to train their employees.

Our boss firmly believed that 'the bold feast, the timid starve,' and spent 500,000 yuan to buy the full system.

Normally, he would hold up office supply reimbursements for weeks; his hotel standard for business trips was only 300 yuan. His entire prior investment in AI had been nothing more than forwarding a few short videos with titles containing FDE and OPC to the work group chat.

This time was different. That AI deployment company told him that if integration was slow, the company's AI transformation would be left far behind by peers. The boss, who firmly believed AI could change the world, panicked at once.

In September, he held an all-hands meeting and announced that the 500,000 yuan had been paid, and that from then on all efforts should be devoted to this system.

The audience was stunned. He had always been cautious in his work. When partnering with others, he required us to first understand the other party's background and hold several rounds of group meetings before considering signing a contract; any changes in the company would also be signaled in advance. Even when buying a printer, he wanted everyone to test it before agreeing to pay.

Two weeks after the meeting, the AI deployment company sent people over.

The boss was so angry that he was hospitalized.

After four brief training sessions, the legendary all-in-one system was distributed to everyone.

The first feature to go live was the AI customer acquisition tool the boss had been most eager to use. In plain terms, it let AI take over new media operations. Employees entered prompts, and AI generated images, edited videos, and automatically published them to accounts on various platforms.

Previously, the company accounts were managed exclusively by colleagues in the new media department. The content included screen-recorded operations and talking-head videos filmed for the boss and management. The results had always been good, bringing in dozens of qualified customers each month.

After the system went live, the boss required all platform accounts to be handed over to AI, and also opened more than 50 new accounts in bulk for simultaneous operation.

The company's editor spoke up, saying that using AI to mass-produce repetitive content could easily lead to account bans, and it would be better to pilot one new account first. However, the boss fired her on the spot, citing her as an obstacle to the AI transformation.

Just one week after the editor left, an incident occurred. When colleagues in the new media department logged into the accounts, they found that several accounts on major platforms had been banned. After contacting platform customer service, they learned that the content published daily on these accounts had been generated from the same set of prompts, resulting in extensive duplication and being judged as low-quality by the platforms.

When the matter reached the boss, he did not even collect his belongings and rushed to the AI deployment company to demand an explanation.

Faced with the questioning, the other side shifted the blame back: during training, they had stated that videos posted with AI should be differentiated and should not reuse the same set of prompts each time, otherwise they could easily be throttled or banned.

Seeing the boss's momentum weaken, the other side pressed their advantage and took the opportunity to recommend managed operations services.

Managed operations was not cheap: 2,000 yuan per account per month, excluding Token fees. With this system, Token also had to be purchased from the company at a price more than five times the market rate.

Normally, I would spend 10 yuan to buy a key from DeepSeek, enough for the whole company to use for a week. Based on that, the company's monthly cost would be less than 5,000 yuan. Buying from the AI deployment company, the same usage would cost more than 50,000 yuan per month.

The other side promised that managed operations would guarantee no account bans and also gave the boss membership that allowed him to attend their company's dinner event. The boss paid again. Including the cost of purchasing the system, the total spend was approximately 700,000 yuan.

After the payment was made, the boss returned and held an all-hands meeting, requiring everyone to make full use of the system. To avoid duplicate content, each person had to create videos with AI themselves. Since we were already unfamiliar with the tool, we could only repeatedly revise prompts and repeatedly generate content, which made Token costs even worse.

At the end of the month, the boss saw the bill, felt faint, and decided not to top up Token. But as soon as Token purchases stopped, the previously functioning AI customer service also stopped, and customer complaints followed one after another.

The boss went to confront them again, furious. This time, the other side gave the same explanation: computing capacity was insufficient, and once capacity increased, prices would naturally fall.

As complaints increased, the boss had no choice but to grit his teeth and continue topping up. To get his money's worth, he also created a digital avatar for himself and each member of management, hoping to use digital-avatar talking-head videos to replace the frequently problematic AI-generated videos.

The digital avatars were completed in just one afternoon. The boss posted a talking-head video generated from his own likeness, and two hours later received one comment: "Looks like facial paralysis." After that, he never mentioned digital avatars again.

The other managers were still using their respective digital avatars. They did save time, but no one cared about the results.

The problem with AI customer acquisition had not yet been resolved, and AI HR caused another incident. Just after September payroll was issued, the colleague next to me suddenly turned their computer toward me, and the screen clearly displayed: supervisor salary 35400.67 yuan.

He said that after checking his own salary, he casually asked how much the supervisor was paid, and unexpectedly the AI actually told him.

The news quickly spread throughout the company. Some colleagues followed up and found three unfamiliar names to whom the company paid salaries every month. When we repeatedly asked the AI, the note we received was: the boss's relatives.

It turned out that shortly after the system was connected, the boss had HR upload the payroll data so employees could check it themselves. But the data was not segregated, and the AI disclosed everything.

In the end, only the boss could clean up the mess. At the all-hands meeting, he said resentfully, "All salary information was fabricated by AI." In the audience, my colleagues and I exchanged smiles.

After this, the boss concluded that HR did not understand AI. He decided to hire a group of new employees who understand AI.

A new round of AI hiring began. To recruit AI-native talent, the boss followed the big-company playbook he had learned from short videos and made the AI interviewer the main assessment.

During interviews, the boss called candidates into his office, took out his phone, opened the AI interviewer, and had them answer questions to the phone.

That week, he interviewed more than 50 people, and almost everyone who walked out looked baffled.

After the interviews, he asked the AI to analyze each candidate. Because the questions only covered weaknesses, the AI's feedback consisted entirely of weaknesses. After looking them over, he liked none of them.

Only one video editor who understood AI passed, but she declined the offer before starting. I privately asked her why. She said that before signing the contract, she had asked the AI to review it, and the AI pointed out a pile of problems in the contract. In the screenshot she sent, the AI's summary was: 'This isn't a labor contract—it's an unequal treaty.'

Figure | AI's analysis of the contract

After all this trouble, the company's revenue and customer base had dropped sharply. At that point, the boss remembered the AI Advisor, a feature that automatically generates decision recommendations based on usage records and company data.

At the meeting, he demonstrated it on the spot, typing into the chat box: 'Predict the company's development over the next six months.' 30 seconds later, a line of text appeared on the screen—it predicted the company would go bankrupt within six months.

The next day, the boss's office was empty all day. The same happened the third day. Later, colleagues said the boss had been so upset by the AI Advisor that he was hospitalized.

700,000, just for a bang

The real drama was yet to come.

Before the National Day holiday, the AI deployment company suddenly became unreachable. The collaboration group chat was dissolved, and when a colleague reached the contact person, the other side said they had already left the company.

Whether out of embarrassment or realizing he had been scammed, the boss also disappeared for a while. When he reappeared, it was at that all-hands meeting.

I later thought about it for a long time: what did that 700,000 actually buy?

The boss had shared a post about FDE, but what he bought was completely different.

FDE, short for Forward Deployed Engineer (forward-deployed engineer)It refers to engineers stationed at the client site long-term, deeply involved in the client's business processes, customizing development as they understand requirements, and continuously iterating.

In plain terms, it means working alongside the enterprise to make its processes run end to end again. For a system to truly go live, the cycle often runs in years. It also requires a delivery team of at least seven or eight people: a project manager to coordinate, solution engineers to break down business needs and translate them for the technical team, data engineers to connect the data, backend and algorithm engineers to implement it, and someone stationed at the client site to provide training, ongoing support, and performance monitoring. On the client side, there also needs to be someone who truly understands the business to cooperate throughout.

This is completely different from selling a standard software package, conducting four training sessions, and then charging for Token and managed operations.

On-premises deployment, on the other hand, means installing the system and data on the enterprise's own servers so that data does not leave the company. It was originally an option for sensitive industries such as government and finance, at the cost of handling hardware, operations and maintenance, and upgrades entirely in-house. For a small company without even a dedicated IT staff, the day the system goes live is often the beginning of no maintenance.

What the boss bought with that 700,000 was software and anxiety. After four training sessions, the vendor's job was done; the remaining money was collected through Token and managed operations. The same deal was charged three times, and none of it was for helping the client make the process work end to end.

700,000 is a windfall for selling standard software, but not enough for real transformation. With a team of seven or eight people plus acceptance testing, the vendor would lose money. So that company never intended to do FDE from the start. Rather than saying it ran away, it was selling a different product all along.

A 2025 survey by the China Electronics Standardization Institute of 1,200 above-scale manufacturing enterprises showed that 62% of enterprises had already purchased or deployed AI systems, but only 38% were still operating routinely after three months of going live; among the reasons for discontinuation, nearly 30% cited 'no ongoing on-site support'.

Later, I saw many posts online from business owners who had successfully transformed with AI. Several conditions were indispensable: the boss must understand the business and be willing to dive into the details; the company's data must first be digitized—if the data is not included in a digitized system, rolling out AI office tools to all employees prematurely can cause major problems; the system must ultimately be handed over to the business units and used every day. If AI is treated only as a simple tool, it is hard to achieve real results, and sometimes it can even become a burden on employees.

Our boss met none of these conditions.

During the few days he was hospitalized, the company was unusually quiet, and my colleagues and I seriously discussed this system for the first time. As we talked, we found that we did not actually dislike AI.

When the AI customer service first arrived, we all thought it was pretty impressive. During the first test, we could not tell at all whether the person on the other end of the call was human. After the AI customer service was integrated, we also handed over our first round of phone breakthroughs to the AI.

It could also help us create some PPTs, retrieve company materials, and help new employees get familiar with the business. It even helped me make money.

When the system was first integrated, my boss asked me to use AI to create a query plugin for clients. Once data was entered, it could diagnose a company's strengths and weaknesses. After I completed it, my boss pitched it to more than ten clients. Two of them renewed their subscriptions because of it, and when he was pleased, he rewarded me 500 yuan.

In principle, for a small company with fewer than 100 people, deploying AI should have been easier than in a large company. But we simply could not push it through.

Because the AI needed large amounts of real business data for learning in the early stage, the AI deployment company told the boss, "The more data the AI receives, the more capable the AI becomes."

The boss firmly believed this. Soon, everyone in the company received a new task: teach the AI to learn their own jobs. Customer service uploaded Q&A materials, technical staff uploaded documents, sales uploaded case studies, and even work experience had to be organized and uploaded.

All of this had to be completed outside our regular work. During that period, after work every day we spent two hours training the AI, and on Sundays we still had to come to the office and continue working overtime.

We were all puzzled: we bought AI so that it could do work for people, but why did people stop their own work first to teach the AI how to work?

The boss's explanation was that the current AI was like an 8-year-old child. It did things imperfectly and needed to be raised; once it grew to 18, it would naturally be able to shoulder responsibility. At first, everyone just listened and let it go. Only when he said, "AI will eventually replace all of you," did we realize that the child we were training with overtime every day would, once grown, take our jobs as its first job.

Some veteran colleagues who could no longer stand it began to leave one after another. Every time someone left, the boss insisted that the person had been eliminated by the company, citing their obstruction of the AI transformation.

In the most extreme week, nearly one-third of the company left, and he was not worried at all. He believed that the experience of those people had already been "fully uploaded to the system." He also coined a trendy term for this: distillation.

Later, he hired two graduates with a monthly salary of 3000 and equipped each of them with three Agents, convinced that this was enough to replace veteran employees with a monthly salary of 8000. At first, it went fairly smoothly. When the new employees encountered something they did not understand, they could ask the AI for company materials from the past few years, and within a few days they could search through most of it.

Problems quickly appeared. Less than a month after the new employee started, the boss sent him to negotiate the final step with a client he had maintained for four months.

The opening went well, and the AI-generated report and PPTs satisfied the client. It was not until the other side began pressing for details that were not in the knowledge base that the new employee started to stumble and repeatedly looked down at his screen. The client then turned and gave the 10 万元 long-term contract to a competitor.

As usual, the boss blamed the people: they did not trust AI enough. He announced that the extent of AI usage would be included in performance evaluations, accounting for 60%.

When policies come from above, countermeasures emerge from below. To fill the performance gap, we began adding AI to our daily work reports. We renamed copywriting "AI-assisted content production," data organization "AI data cleansing," and proposal development "AI completed the client full lifecycle plan from 0 to 1."

In this way, AI became widespread in the company. The thing I used AI for most each week was having AI help me make up what new processes I had optimized with AI that week.

Later, I had a meal with the vice president who had been fired. Talking about this chain of events, he said that in the boss's eyes, whenever anyone said the system was hard to use, it was a personal rebellion against him.

"More than losses, what the boss feared most was being unable to control people," he said.

Source: 深潮 TechFlow