Hiring AI as a colleague, how is my business doing?

The term "digital employee," which once carried a touch of science fiction, is now becoming a real - world work mode.

In the past year or two, the capabilities of AI have been significantly enhanced. Especially at the beginning of this year, during the "lobster fever" triggered by OpenClaw, AI Agents could already take over browsers, read and write files, execute code, and call terminals. Many people truly felt for the first time that AI had started to "get its hands dirty" and do real work. Robin Li, the founder of Baidu, even proposed a new metric, using DAA (Daily Active Agents) to replace the long - used DAU, to better track how many Agents are working for humans and delivering results.

While the concept is gaining popularity, what we really want to know is how those who have actually put "digital employees" to use are faring?

「Focus One」had conversations with five individuals. A lawyer wrote 45 Skills to teach an Agent to "handle cases" for them; a cross - border e - commerce merchant laid off three employees and relied on four digital employees, each taking on "four roles"; an investment manager let an Agent screen BPs while sleeping... These digital employees, which don't require salaries or social security contributions and are available 24/7, with a "monthly salary" of only one or two thousand yuan, have truly emerged in the daily work of independent developers, small teams, small and medium - sized business owners, and workers.

Here are their stories.

01. Building AI Infrastructure Like Crazy, I Taught an Agent to Work for Me with 45 Skills

Yang Weixin | 33 years old, Suzhou, Lawyer at Cambridge Yihua Law Firm

The legal profession is naturally suitable for collaboration with AI. I started using Agent tools in April or May last year. At first, I used Cursor to develop some small tools, such as a browser plugin to export AI chat records from the web.

Gradually, I began to try applying it to legal services. At that time, the capabilities of the Agent were not yet fully developed, and its context window was limited. However, it could indeed find some key information from the evidence materials. For dozens or hundreds of pages of evidence and hundreds of contract clauses, it could directly help me locate the content I needed.

Later, the Agent gradually permeated every aspect of my work.

My work is divided into two parts: one is the core legal business, and the other is legal content writing and knowledge sharing. For this, I spent five months writing 45 Skills to enable the Agent to produce "legal language" according to my set requirements in case handling, research, and creation.

My day usually starts with "harvesting." I let the Agent automatically collect news and typical cases in the legal field from the previous day. At 10 a.m., I'll browse the articles it has sorted for me and decide on the direction for learning or creation for the day.

When handling cases, I let the Agent automatically download and archive the complaints and evidence materials sent by the court into the specified project folder. Then, I use the OCR (Optical Character Recognition) Skill to complete the format conversion of the materials. Next, I use the legal search Skill connected to a professional database to conduct a legal analysis around the focus of the dispute. Finally, the writing Skill outputs the first draft of the defense or the legal service plan.

Image source / pexels

The most crucial step is manual inspection. AI can produce the first draft, conduct legal searches and analyses, but it's difficult for it to make professional judgments like a lawyer. After reading the first draft, I'll give revision suggestions and let it make corrections.

The Agent tools I use more often include Claude Code and Codex from abroad, as well as WorkBuddy and QodeWork from China. The monthly expense for AI subscriptions is about one thousand yuan.

Now, with the help of the Agent and specialized Skills, I can complete my legal work in 20% of the time, and the remaining 80% is spent on "building infrastructure."

I turn my experience and preferences into individual Skills to make the Agent run in my way. The building process is very energy - consuming. However, an Agent can't directly meet your requirements. Its usability depends on how much of your own "context" you've instilled into it.

In addition to using it myself, I also share my experience at local bar associations. When giving lectures, I can clearly feel that there are significant differences in people's understanding. Many people want to get closer, but they're also a bit resistant in their mindset. This resistance comes partly from professional pride and partly from the fear of being replaced.

Many people don't know how to use Agents or are quite resistant, essentially because they haven't found a good scenario. Without a scenario, there's no motivation, and without motivation, a positive feedback loop can never be formed.

So, for those who want to use Agents, my advice is to first find a very small, specific work scenario or a process where you want to improve efficiency, and turn your methodology into a Skill. Don't try to make AI do everything right from the start. Just focus on one point, invest in it, and solve it. Once you've solved it, the positive feedback will drive you to explore the next scenario. By connecting multiple Skills, you can get an Agent that can solve specific problems.

I believe that in the next one to two years, the most important ability is to structurally break down your work and continuously precipitate your personal "context" to train an Agent that can truly help you.

02. Three Employees Left, and Four AI Employees Stepped In

Lei Zi | 32 years old, Chengdu, Cross - border E - commerce Practitioner

I'm engaged in cross - border e - commerce in Chengdu, selling smart home accessories, with the main market in North America.

This time last year, I had three employees: a graphic designer, an operator, and a customer service representative. The total salary and social security for the three of them were nearly 50,000 yuan per month. With the office rent, the annual fixed cost was approaching 700,000 yuan. In good months, the monthly profit was 70,000 - 80,000 yuan, and in bad months, it was 30,000 - 40,000 yuan.

In May last year, the operator left. I thought I'd use AI as a temporary replacement while recruiting. At that time, Claude was already popular, so I gave it a try. Claude quickly wrote a complete set of English copy for the A+ page, from product selling points to scenario descriptions, all very authentic. I still used the original graphic design materials, but the quality of the copy was better than before.

At that moment, I realized that I didn't need to recruit anymore.

By the beginning of this year, the graphic designer went back to his hometown to get married, and the customer service representative quit because of the low salary. Now, my team consists of me and four digital employees: Claude is in charge of copywriting, Fin handles customer service, Cursor takes care of coding, and there's an automated system monitoring advertising placement. The price data of competitors is imported from Keepa, and the data is automatically summarized on the dashboard every day.

Image source / pexels

Every morning at 9 a.m., I turn on my computer and first look at the data dashboard I built myself, which runs five automated tasks: monitoring competitor prices, creating drafts for new product listings, handling customer emails from the past 12 hours, analyzing advertising placement data, and scheduling social media posts.

The after - sales Agent uses Intercom's Fin, which is connected to my Shopify after - sales service. I synchronize the order data from Amazon through Zapier. Fin can handle common problems on both platforms, such as logistics inquiries, returns and exchanges, and usage instructions. More complex issues are then forwarded to me.

The biggest change is in advertising placement. In the past, it was common for me to stay up until 2 a.m. monitoring data. Now, I've written the rules into the workflow: automatically reduce the budget if the ROAS is below 2.5 for 48 consecutive hours, increase the investment if it's above 4, and automatically change the image if the click - through rate of the ad material drops for three consecutive days. I let the system run according to these rules. During the pre - sale period of Black Friday last year, I ran six groups of ads simultaneously by myself.

In terms of coding, I had some programming basics before. The landing page of the store, AB test pages, and data scraping scripts were all developed through conversations with Cursor. Last month, I wanted to create a tool to track competitor prices. I found that Amazon's anti - scraping measures were very strict. Later, I used Keepa's API to get the price data and wrote a dashboard for classification and anomaly alerts by myself. In the past, I would have had to spend 20,000 yuan to outsource the task and wait for two weeks. This time, I did it myself and finished it in two days by chatting with Cursor.

My current monthly expense for AI subscriptions and API usage is about 4,000 yuan, which is significantly lower than the previous cost of hiring employees.

However, AI can also make mistakes. In March this year, I asked Claude to write a batch of Facebook ad copy and was lazy and didn't review each one. It wrote a set of copy promoting a "50% limited - time discount," but I wasn't offering any discounts at that time. I didn't notice until the ad had been running for two days. After that incident, I must review all content intended for customers.

Over the past year, many people have asked me for suggestions on using digital employees.

First, don't aim for full automation right from the start. Find one or two time - consuming processes, such as customer service, copywriting, or data organization, and let AI handle them. Second, be as detailed as possible when assigning tasks to AI.

Another point is that your judgment ability can't degrade after using AI. You need to be able to tell if the customer service response is correct, if the ad copy is off - track, and if the conclusion of the data analysis is logical.

I now work six to seven hours a day, half of my previous working time. I use the extra time to learn about the supply chain, visit factories, and have video conversations with overseas customers. Sometimes I think that if the operator hadn't left last year, I might still be that small business owner worrying about the profit statement. To be honest, for most cross - border e - commerce businesses, the endgame is one person and one computer.

03. After Ten Years in Investment, I Finally Have an Agent to Work Overtime for Me

Ray | 36 years old, Beijing, Partner of a RMB - denominated Fund GP

I've been in primary - market investment for nearly ten years. In the early days, I focused on aerospace and new materials. In the past two years, almost all my projects have been related to AI.

The pace of RMB - denominated funds is already tight. Due to work requirements, I fly across the Pacific more times per month than within China. Because of the time difference, if I want to be available at all times during working hours, both I and my project team need to be "on standby 24/7."

Once, I met with seven companies in one day. It was already 1 a.m. when I got back to the hotel, and the IC (Investment Committee) was going to discuss a semiconductor equipment company the next day. I hadn't even started writing the memo. So, I gave the company name to an Agent running on my computer and told it to "thoroughly research this company according to the IC framework," then I turned off the computer and went to sleep.

The next morning when I woke up, I read its "work log" and found that it had done a lot of things overnight, including opening the browser by itself, scraping all the public patents of the target company in the past three years, accessing my own database, and crawling all the public interviews and LinkedIn posts of the founder in the past six months.

At that moment, I felt for the first time that I didn't have to stay up until 3 a.m. myself.

Image source / pexels

Later, I used more and more "AI digital employees." While I'm sleeping, AI mainly does three things for me: summarizing roadshows, conducting corporate research, and managing my schedule.

It automatically transcribes meeting recordings. In the research process, I use the Claude plugin in Obsidian in combination with Claude Code to scrape prospectuses, annual reports, public news, and check corporate lawsuits, and then integrate this content into my personal knowledge base. In terms of scheduling, it can help me figure out when I'm in which time zone, whether I can take on another meeting, and whether I need to cancel one.

Half a year ago, I built a multi - Agent BP filter for myself.

During the peak project period, I can receive three or four hundred BPs per week. It's impossible to go through them all manually, and the cost of missing good projects is very high. So, I wrote down my BP - screening methodology in plain language one by one and let AI build the filter for me.

I encountered many pitfalls along the way. At first, the judgment criteria were set too strictly, and several good projects were rejected by it. The current version is more "lenient." AI only conducts the initial screening and grading. I'll review Class A projects myself, let another Agent write a short review of about a hundred words for Class B projects before deciding whether to meet, and archive but not delete Class C projects.

My monthly expense on AI fluctuates greatly. In months with less usage, it's four or five hundred dollars, and it can reach over 2,000 dollars during intensive due - diligence periods. On average, it's about 1,000 dollars per month.

I feel that my current "digital employees" can help me solve over 80% of my work. If calculated based on my income, the cost of "hiring" them accounts for less than 10%.

Digital employees can't make investment decisions for me, but they allow me to review twice as many projects, sleep two more hours a day, and bring a more detailed memo to the investment committee. For an investor, this is already a huge benefit.

04. ChatGPT as the Advisor, Cursor as the Engineer, and I as the Boss

@Freddy's Grand Universe, 29 years old, Hefei + Chengdu, Founder of a Travel and Corporate Business Travel Service Brand

I'm the founder of KoraScale, a travel and corporate business travel service brand. Currently, my team has exactly 10 members.

However, if we include my "digital employees" such as ChatGPT, Cursor, and Codex, the team size has actually exceeded that. After using them for a year, the most important thing I want to say is that AI is not here to take your job; it can really help you get things done.

Previously, I thought AI could at most help me write a few code snippets. But when I started letting Cursor read the entire project, analyze the page structure, modify database queries, and even directly deploy the results online, I realized that AI is no longer just a Q&A machine; it's an executor that can enter the work site.

This experience has broken through my ability ceiling. I'm not a professional programmer; I'm an industrial engineer