A fresh graduate achieved $15,000 in sales performance after only one and a half months on the job. No manager mentoring, no senior colleague teaching — all thanks to an AI system.

Veteran AI user Xu Zuwei told Paida that his company now directly bypasses the e-commerce supervisor role: "This kind of position is generally expensive, and whether it can solve my problems requires spending significant money and time to test." Therefore, after the AI wave hit, he tried using AI to build a workbench for capturing salespeople's daily chat records and conducting analysis and scoring based on scripts.

After obtaining these scoring results, HRBPs would manually organize them, conduct interviews with underperforming salespeople to determine the reasons, and then provide targeted training. Xu Zuwei said that under this logic, their talent screening and development efficiency is very high. Of the last batch of 7 fresh graduates recruited, only 2 were ultimately retained, and his decision on who stayed and who left largely referenced AI's suggestions.

Although this story sounds cruel — using AI to eliminate employees — Xu Zuwei is actually quite clear-headed: "People carry subjective biases, while AI is more冷静. As long as the standards, boundaries, and yardsticks are well-established, AI can completely replace humans in evaluation work."

Is this logic of using AI to evaluate and train newcomers actually reliable? How difficult is it to build this kind of workbench? Is the cost high? In a conversation with Xu Zuwei, Paida found some answers.

It's Not That They Don't Want to Use Newcomers — Training Costs Are Simply Too High

The most exhausting part of training newcomers is often not the process of answering questions, but finding where the problem lies.

Xu Zuwei's business is mainly printing and packaging, with both domestic and export operations. The export business is primarily conducted through Alibaba International Station, and the AI workbench mainly revolves around the export business.

Currently, Xu Zuwei's company receives approximately 100 customer leads per day. Due to the large volume and long transaction cycles, there is a need to train a large number of salespeople. But for some newcomers who haven't closed a deal for a long time, he sometimes feels powerless: Is it a problem with understanding the business? Or is it a problem with communication details?

Having interns self-review and self-evaluate often results in low communication efficiency. And if he were to teach them personally, he would first need to read through all 100-plus chat records — too time-consuming.

Xu Zuwei calculated for Paida: "One person's communication takes 16 minutes, so 7 people would take nearly two hours. I simply don't have that much time normally. The input-output ratio is too low. Hiring a supervisor in the market costs at least 12K starting, and expensive ones can reach 20K. Spending one to two hundred thousand a year — can this money match production value? It's actually very hard to quantify."

It was precisely based on this dilemma that after being exposed to AI, he made up his mind — to build a patrol workbench for evaluating work quality, tailored to his company's situation.

On the platform, he chose Coze, and the model used was Kimi3, currently the smartest domestically available. After the idea was formed, with essentially no programming experience, he began his journey of building the workbench by hand.

Initially, he exported chat records from over 10,000 real customers, had AI distill them, categorized and extracted keywords, set up 16 evaluation dimensions, and built the first version of the workbench.

The logic of this workbench is simple: The system automatically patrols daily to extract salespeople's chat records, submits the data to AI, and AI matches chat records against a keyword library based on the 16 preset evaluation dimensions, giving the salesperson a comprehensive score based on keyword matching.

Xu Zuwei explained that after local deployment, the script can conduct fixed daily patrols. After extracting salespeople's chat records, it matches and scores based on preset standards to identify where salespeople's problems lie.

"Shorter ones take just over 40 seconds, longer ones take a few minutes, and you don't need to watch it. Basically, it can be organized in about an hour or so."

But shortly after putting it into use, he discovered a problem: Keyword matching alone cannot reflect an employee's work quality. Some deals were closed without matching keywords; some with high keyword match rates still couldn't close deals.

After discovering the problem, Xu Zuwei changed his approach — he selected a batch of excellent examples as underlying data, had AI analyze these success cases to summarize common patterns, avoiding interference from messy data, so the output scores could be more accurate.

"This is a process of continuously finding certainty," Xu Zuwei said.

After this version was put into use, the results were noticeably better. Afterwards, Xu Zuwei updated several more versions based on problems that emerged in actual evaluations. As of now, it can basically be perfectly deployed locally and fits their business.

From initial experimentation to deployment to completed iteration, Xu Zuwei took less than a month, though the token consumption was indeed substantial: "It cost about 10,000-plus yuan in credits, but I think it's very worth it because it's closely integrated with our company's business. Previously, using a third-party CRM cost over a hundred thousand a year. Calculating it this way, it's actually not expensive."

When discussing the details of expenses, Xu Zuwei said that polishing the "yardstick" is where most effort is spent. "The establishment of standards must be combined with business reality and repeatedly polished. Each yardstick takes about 700-800 yuan to refine."

However, according to Paida's understanding, many times the bottleneck in companies' AI tool adoption is not purely at the technical level, but at the human level. Paida has previously interviewed many merchants, and they all said: Many employees resist AI tools.

On one hand, many employees fear AI will replace them, so they refuse to hand over data for the company to distill. On the other hand, many employees believe AI is not objective and fear that AI intervention will actually reduce work efficiency.

Xu Zuwei has naturally considered this issue, but his answer is simple: We use AI to improve efficiency, not to lay people off.

AI Is Only for Error Correction — It Cannot Replace Training

Xu Zuwei told Paida that he does not believe AI's results can 100% reflect an employee's true work level. "AI is just a reference. Human intervention is also important."

In Xu Zuwei's view, AI can define an employee's most basic abilities — such as response speed, average communication rounds, quotation information completeness, etc. These processes are relatively mechanical and repetitive, so AI can quantify them fairly easily.

But for more complex issues — such as application in specific scenarios, personalized needs, etc. — these test salespeople's familiarity with the business, and having AI judge these would be somewhat strained.

To address these specific pain points, Xu Zuwei used two methods as safeguards: Building corresponding Q&A libraries based on specific scenarios, covering as many full-process problem scenarios as possible, and verifying the match between chat records and the problem library on top of keywords; Setting up HRBP positions, where humans use system-generated results to conduct interviews, correcting AI-induced deviations during communication.

Under this approach, the advantages of this model become obvious — using AI to give salespeople a preliminary diagnosis, finding their weaknesses, then having HRBPs correct them, and then having HRBPs report to the boss. Communication efficiency is noticeably higher.

Of course, finding problems is just the beginning — training is the real goal. To better conduct business training, Xu Zuwei used AI to organize 60-70 training documents, containing almost all customer-facing scripts, covering all scenarios. Even if salespeople encounter cognitive blind spots, they can quickly match corresponding documents.

According to Xu Zuwei, under this training system, some newcomers get up to speed quickly. For example, one fresh graduate achieved a $15,000 order after only one and a half months on the job.

Regarding the AI trend, Xu Zuwei is also quite reflective. He mentioned a saying he heard during training at Alibaba: "Every business is worth redoing with AI."

Beyond this AI talent development system, he has also tried many other things with AI. For example, in May of this year, through brainstorming with ChatGPT, he explored a new category possibility. Through multiple rounds of dialogue with AI, he continuously enriched project details — such as keywords, main image details, etc. Once the product launched, it sold over 1,000 units.

Xu Zuwei believes AI is definitely the trend, but he always firmly believes that talent is the core of building the yardstick. Some people have captured new opportunities in the AI era, but many bosses have gotten lost in the "false prosperity" provided by AI.

He frankly stated: "Many bosses made money too easily in previous years. After market conditions worsened in the past two years, they became anxious. AI came out like a lifeline, but how many of AI's proposed solutions can actually be implemented at the company level? Many people haven't really figured it out."

Only Focusing on Cost Reduction Without Efficiency Gains — Business Definitely Won't Last

In the AI era, everyone talks about reducing costs and improving efficiency, but some bosses only focus on cost reduction, thinking every day about how to lay people off, without figuring out how to "improve efficiency."

But Xu Zuwei sees it differently: "I think cost reduction is fine — everyone has different starting points. But my purpose in using AI is not cost reduction, but efficiency improvement."

According to Xu Zuwei, his company has been using AI for a long time. Although costs haven't improved significantly, business performance has indeed grown considerably. "Since using AI, domestic GMV has grown about 30%, and export has directly doubled or tripled."

So Xu Zuwei believes there are many merchants now waving the AI banner, but figuring out how to apply the business value logic is the original purpose of using AI. "Someone once distilled Kazuo Inamori into a model to teach them how to manage a business. Can a company with thousands of people be the same as one with ten people? There's absolutely no reference value — it's pure韭菜割."

In his view, a company wanting to use AI really has only 0 and 1. Not integrating with the business is always "0" — AI is just a chat tool, providing only emotional value without commercial value. Only by landing AI, improving efficiency, and seeing growth can you achieve "1."

From an emotional level, he can well understand some merchants' anxiety: "E-commerce is getting harder and harder. Everyone is getting more anxious. The more anxious, the harder it is to calm down and think. AI is like an understanding expert, planning a grand blueprint for you, letting you see many things beyond your cognition. But how much can you actually implement? Very limited, actually."

Xu Zuwei believes that currently, operational materials that can be found online are very scarce and extremely fragmented. Relying solely on this information combined with a general large model can provide very little empowerment to a company.

Therefore, in his view, people will always be the important "yardstick." Only by having professionals define professional "yardsticks" and finding AI's application boundaries can AI become smarter the more it's used.

Putting Xu Zuwei's practices together, at least three conclusions can be drawn.

1) Before AI can be deployed, bosses must first teach AI. Take Xu Zuwei's company as an example — a salesperson isn't closing deals. Where exactly is the problem? The answer won't be automatically generated by a large model. Only after defining the standards and judgment criteria will AI know what to check. The vaguer the business, the more distorted the system's conclusions. The clearer the operational experience, the easier it is for AI to become a stable execution tool. Therefore, the first step for merchants introducing AI is not choosing the strongest model, but organizing their own business: which tasks happen repeatedly, which actions can be measured, which results can reverse-verify judgments.

2) AI's value is efficiency improvement first, cost reduction second. In Xu Zuwei's team, the AI system is not a "one-click managed" role. It only handles patrol, recording, and flagging issues. HRBPs handle interviews, bosses set rules, and salespeople still face customers. What Xu Zuwei did was simply break apart the work previously concentrated on supervisors, handing some repetitive, mechanical tasks to AI, while key steps like communication, training, and decision-making are still handled by humans. This also explains why his AI investment didn't immediately reduce costs: AI's first benefit is exposing employee problems earlier, shortening the cycle of discovering problems, coaching corrections, and verifying results.

3) In AI e-commerce competition, what ultimately matters is who understands the business better. General models can write copy, make images, and give suggestions. These capabilities will become increasingly common. What truly differentiates merchants is whether they can connect their own customer records, product data, operational processes, and business judgments into tools, letting AI participate in specific work. Being able to build a tool doesn't equal creating commercial value. Only when AI truly enters the organization and improves operations can the leap from "knowing how to use AI" to "using AI well" be completed.

Therefore, the merchants who benefit first from AI in the future may not be those who use the most models, but those who earliest clearly articulate their business, accumulate their experience, and are willing to continuously correct the system with results. After all, AI is just a tool. It won't conjure a mature management system out of thin air — it will only accelerate the transmission and execution of a company's existing experience.