I tried to build an AI company, but it was faster to use existing AI routines
When browsing social media, I sometimes come across posts saying, "I built an AI-only company."
There's a CEO AI, a marketer AI, a writer AI, and a designer AI.
It's a system where AIs consult with each other to carry out work, while humans only provide instructions and approvals.
It sounds like a dream.
Since I'm also building my own personal AI assistant, I thought, "Couldn't I just go as far as building an AI company?"
Assign roles to multiple AIs, delegate tasks, and have them manage progress.
If I could automate everything from planning to production and review, it seems like I could handle many things in parallel.
However, when I broke down what I actually wanted to do in detail, I realized that a slightly different approach would likely be better.
In my case, it was easier to use existing AI routines than to process everything through a self-built AI company.
What I want to do is operate multiple media outlets efficiently.
One of the things I want to do using AI is the operation of web media.
Research keywords, create article outlines, write the body text, prepare images, check the content, and publish.
Furthermore, after publication, I want to improve the articles while monitoring rankings and traffic.
I want to do this not just for one article, but continuously across multiple sites.
Because of that, I initially thought I should create research, planning, writing, editing, and design roles within an AI company and have the AIs pass work to each other.
However, recently Gemini Spark has arrived, and there are also Codex and Claude Code.
If each of them has routines or background execution mechanisms, I reached the conclusion that I could just entrust most of the routine tasks I want to do directly to them.
Let Gemini handle keyword research and outline creation
Have Claude write the body text
Have Codex create eye-catching images and diagrams
Use spreadsheets to pass along progress and deliverables
If you decide on the rules for each site beforehand, there is no need to have the AIs consult with each other every time an article is created.
You can just prepare routines for each type of article and execute the relevant one.
What I was looking for was a way to make article production proceed steadily.
Thinking about it that way, for routine tasks, "creating a routine" was more suited to my goal than "creating an AI employee."
Building from scratch increases the amount of work outside of what you actually want to do
If I try to achieve the same thing with a self-made AI company, I also have to figure out how to pass data between AIs myself.
I also need to decide where to store data, which AI starts the next process, what to do if it stops in the middle, and what constitutes completion.
In the process of building my own AI assistant, I have already implemented memory, databases, scheduled execution, and integration with external services.
What I felt many times there was that building the peripheral parts to keep the AI running stably is harder than the part where the AI does the work.
Of course, there are times when building that part itself is interesting.
However, when I want to increase the number of articles for my media, if I start by developing the foundation of a huge AI company, the development will continue without a single article being added.
In my case, completing an AI company is not the ultimate goal.
The goal is to grow the media I operate, give shape to new projects, and advance tasks that I cannot handle alone.
I came to the conclusion that if there are already routines I can use, it is faster to use those.
Even so, I still need my own AI company
So, does that mean I am abandoning the concept of building an AI company itself? No, that is not the case.
In my case, the work I want to entrust to an AI company and the work I want to entrust to existing routines were different.
What I want to entrust to existing routines is work where the procedure can be fixed.
Creating article outlines, writing the main text, producing images, and formal checks can be processed repeatedly once the rules are set.
On the other hand, what I want to entrust to my own AI company is work like the following:
Deciding which of multiple media outlets to prioritize
Managing this month's goals and progress
Analyze changes in access and rankings
Think about the reasons for a lack of growth
Design new plans and one-off initiatives
Review deliverables from each routine across the board
Organize things when I have spread myself too thin
These cannot be processed using the same procedure every time.
It is necessary to make judgments by looking not only at the information of the site itself, but also at what I want to do now, the time and budget available, and the status of other projects.
For this part, a dedicated AI that continuously understands my situation is better suited than the standard routines of existing services.
The ideal AI company was not a 'collection of workers'
The AI company I initially thought of was one where AI employees handled each task directly.
However, it looks like it will end up being a bit different.
I will leave routine work and mass-production tasks to existing routines like Gemini, ChatGPT, and Claude.
My own personal AI company will manage their progress and results collectively. It will also handle tasks that are difficult to routinize, such as strategy design, analysis, priority judgment, and complex initiatives.
The image is to use existing AI routines as the operational force, while giving my own personal AIs roles closer to management or executive leadership.
Come to think of it, human companies do not develop all the systems they use for their operations in-house either.
They use existing services for chat, accounting, attendance, customer management, etc., and perform their own unique strategies and judgments on top of that.
Perhaps the AI company I want to build could have been in the same form.
As a result of trying to build an AI company, I did not need to prepare every AI employee from scratch.
What I needed in my case was not a group of AI employees to process routine work, but a management team of my own that could manage the whole thing while selectively using existing AI routines.