How many people do you need to run a company?
✍️ This article is reposted from “Vista Newsletter” No.151, originally published in iamvista.substack.com (opens in a new tab). If you like this kind of content, welcome to free subscription to “Vista Newsletter” (opens in a new tab) and receive my observations and methods first-hand every week.
Hello, hello! I have been on typhoon leave for two days, and the wind and rain outside have not stopped until today. I hope you and your family are safe.
I haven’t been very active these past two days, but one thing makes me relieved: I have finally finished writing the preface to my new book “AI Unmanned Company”. A recommendation for this book written by a good friend also happened to arrive, so I handed both manuscripts to the editor.
When I was writing the preface, I kept thinking about a question: How many people are needed to run a company?
In this e-newsletter, I want to talk to you about this issue.
A boss asked me a question
A while ago, the boss of a design studio told me that as long as one more person is hired, there will be tens of thousands more fixed costs every month, not counting labor and health insurance, education and training, and the risk of whether that person will be able to work for a long time. But if we don’t hire people, the cases at hand can only be pushed out, and the income will also be lost.
He asked me: What should I do?
I believe this is not his only problem. Small and medium-sized enterprises in Taiwan account for more than 98% of all enterprises. The declining birthrate and aging population have turned labor shortage into an irreversible structure. The road to joining is becoming increasingly difficult: expensive, slow and uncertain.
The answer I gave him was the title of this book: Unmanned Company.
No one means no employees
This term is easily misunderstood, so I spent a lot of time explaining it in the book.
The “no one” in an unmanned company means that there are no employees, no one else to pay your salary, manage attendance, and worry about whether you can do it for a long time. The only person in the company is you, who designs the entire system, directs an AI team, and is ultimately responsible for all results.
To be honest, this is how I live my life myself. Every morning when I turn on my computer, the first thing I do is say good morning to Sebastian, my AI butler. It answered me early and woke up the entire team of virtual secretaries: one to check the mailbox, one to check the calendar, and one to take stock of the to-do and project progress. By the time I made the coffee and sat down, the things I needed to do today were sorted out in front of me. When I finish work in the evening, I say good night, and it will review it with me and write down what I did today and what I want to do tomorrow.
During this typhoon holiday, this system gave me an unexpected reminder. When I feel unmotivated, the system still operates as usual: Fan posts are sent out as scheduled, information is updated as usual, and review records still remind me of things that I have not yet paid for what I promised others. What a one-person company fears most is the risk of shutdown. If you get sick, take a vacation, or are in a bad mood, the company will shut down. The part that continues to operate in your absence is the insurance you buy for yourself.
Many people use AI, but in fact they only use it on the first layer.
In the book, I divide AI capabilities into three levels.
- AI Assistant: You give it a question and it passively waits for you to ask. You are still a person who works with your hands.
- AI workflow: You give it a set of rules, and it executes them repeatedly according to the rules. You are the one who designs the process.
- AI agent: You give it a goal, and it will take steps and choose tools on its own, make its own decisions when encountering situations, and report back to you afterwards. You are only responsible for judgment.
Many people say that they are using AI, but in fact, most of them stop at the first level. Crossing from assistant to workflow is the threshold where most people get stuck. Treating AI as a question-and-answer tool is like treating an object that can be a colleague as a talking dictionary.
Therefore, the questions that operators should ask have changed. In the past, I asked “How many people do I still lack?”, but now I have to ask “How many systems do I still lack?”. The ability to build systems used to be the domain of large companies, but now it can be done by one person plus an AI team.
Only you can do the last 30%
At this point, you may be worried: What is left of that person?
The principle I use in the book is called “73/70 man-machine”. Anything that can be repeated, has clear input and output, has precedents to refer to, and can be corrected if done wrong, belongs to the 70% and can be slowly handed over to AI. The 30% left for you is three things:
- Judgment: whether this thing should be done, which one to do first, and when to stop.
- Taste: your definition of good.
- Trust: Others are willing to say “I’ll leave this matter to you, I’ll rest assured.”
This ratio is not a golden number calculated by actuarial calculations. It may be sixty-four or eighty-two. The point is, that 30% will become more and more expensive.
This week I happened to read a long article by Dan Koe titled “The influencer is dead. The Age of Experience is here. (opens in a new tab)”, which talks about the same thing. He divides creators on the community into three types: Internet celebrities treat the community as a job and pursue attention; personal brands treat the community as a monetization funnel and pursue conversion rates; craftsmen use the community to support a set of works and care about things that can outlast algorithms.
He has a saying that I agree with: The last competitive advantage worth paying for in the AI era is a point of view. AI can generate 1,500 words of beautiful text in ten seconds. Information is no longer rare. What cannot really be copied is your own perspective on things.
I think this and “human-machine seventy-three” are like two sides of the same coin. When AI reduces the cost of execution, the more valuable your judgment, taste and opinions become. AI replaces your hands and magnifies your eyes.
I have organized my experience into five key points. Friends who are interested can take a look:
Three myths about unmanned companies
In the process of writing a book, I heard three reactions most often, and I want to share them with you.
The first one is the sweetest and the most dangerous: fully automatic, making money while lying down. The truth is that you have to build the factory solidly in the early stage, and the system design is quite hard, but the sweat will accumulate. Unmanned companies have to suffer first and then gain, and there is no way to make money while lying comfortably.
The second one hides fear: AI will replace one’s own profession. My view is cautiously optimistic. Those who will be replaced are those who only execute but do not develop judgment and taste.
The third pity is that only technical people can afford it. The people who are most suitable to take this path are actually the people with the most domain knowledge, the most taste, and the most trustworthy assets. These three things are not in the program code.
Having said that, this is why I have been teaching people from different backgrounds to make things with AI for the past two years. This week I wrote an article about “Vibe Coding in 2027 (opens in a new tab)”, which mentioned a phenomenon that I have seen repeatedly in class: it is almost normal for students to make small movable web pages in the first class; but when I ask “Are you still using that thing?” three weeks later, less than 10% of the people raise their hands. There will be more and more people who can make things. What is truly rare are those who have finished making things and use them every day.
A little exercise: Should this be left to AI?
If you also want to start handing over the work at hand, you can try the five-question sieve in Chapter 5 of the book. Take something you do every week and ask yourself these five questions:
- Does this happen repeatedly?
- Does it have clear inputs and outputs?
- Are there any examples that I can refer to?
- If it does something wrong, will the cost be high?
- Could you put a manual checkpoint in the middle?
As long as three of the five questions are passed, it is worth designing it into a workflow that is handed over to AI. Don’t throw away the few questions you failed, that’s where you want to set your checkpoints.
For example, if you organize the verbatim draft of your speech into notes and almost all five questions are answered, you can hand it over immediately and let AI produce the first draft and you finalize it. But “the core proposition that determines this speech”, most of the answers are no. It belongs to the 30% that only you can do. Don’t pretend to others first.
I made this quiz into an interactive game on the new book page. There is also a work map next to it. You can click on it to see how many departments you have shouldered:
👉 Look at the new book page of “AI Unmanned Company”
“AI Unmanned Company” is expected to be published in December. If you want to receive notification as soon as possible when it is listed, please click “Register for Listing Notification” on the new page.
Other updates this week
In addition to the book manuscript, there are a few things I have done in the past two weeks, which I can briefly report to you:
I have added a Paper Stuck Map (opens in a new tab) in the researcher base, which connects the 21 most common symptoms of postgraduate students to corresponding articles and tools. Finding answers from symptoms is much faster than looking from catalogs.
I also read the submission instructions for 26 Taiwanese humanities and society journals one by one and compiled a list of AI usage regulations. Whether the manuscript should be disclosed, where to write the statement, and whether the useless AI should be expressed, each journal has different opinions, so it is worth checking before submitting.
👉 List of AI regulations for 26 Taiwanese humanities and society journals (opens in a new tab)
In addition, I recorded a podcast last week to talk about a few things about AI applications. I will share it with everyone after the program is launched online.
If you want to do it, don’t just finish reading
The first phase of the seven-day challenge camp mentioned in the previous issue will start on 11/30, and registration will close on 11/28.
I divided making a small tool into seven days, with a task letter every day: pick a topic, explain the requirements in five columns, use it in a real day, find someone to try it out, and finally publish it and get the website. There is a LINE learning group where we work together, and there is an acceptance live broadcast on the seventh day. No background in programming is required. You can write every letter using the web version, app, and terminal.
If after reading this newsletter, you have something in mind that you want to hand over to AI, these seven days are a good opportunity to do it:
👉 Seven-day challenge camp: Make your first AI gadget in seven days (opens in a new tab)
Finally, I want to ask you a question: If your work was broken down into ten links, from finding customers, receiving demands, delivering, collecting payments, and after-sales, how many of them would you be responsible for by yourself now? Which one wants to hand it over first?
Let’s brainstorm together and come up with great ideas.
After the typhoon, the sky turns blue after the rain. See you next time!
Vista
📮 This article is reposted from “Vista Newsletter” No.151. The original text can be found at iamvista.substack.com/p/vista-no151 (opens in a new tab). If you like content like this, welcome to subscribe for free to the Vista Newsletter (opens in a new tab) and receive a helpful letter every week.
Next step
Bring this topic to your team
Vista Cheng speaks and runs workshops on AI workflows, writing and content systems for companies, schools and public agencies.
