What "AI Companies" Really Are: Of 2,716 Posts, Only 16.1% Had Work Flowing Between AI Workers

"I run my company with 30 AI employees." Lately I often see posts like this on note and Threads in Japan. Below I write "AI employee" and "AI company" (AI社員 / AI会社 in Japanese).

When I first heard "AI company," I pictured something like Paperclip. Paperclip is open-source software that gives AI agents an org chart, assigned work, budgets and approvals. Work drops into an inbox, the worker in charge picks it up and handles it, hands the next piece to another worker, and moves on to the next item. As someone who runs a systems development company, that is what I imagined a company made of AI would look like.

But when I read the posts, many of them were something else. An instruction file for an AI, given a name, was called an "AI employee," and a set of those files was called an "AI company." I did not expect that to be called an "AI employee."

So I collected 2,716 posts about "AI employees" and "AI companies" and sorted them by what they actually describe. I also included posts that did not use the word "AI employee" itself but wrote things like "AI secretary" or "an organization of agents." Posts where work flows from one worker to the next, the "company shape" I had in mind, were 438 of them, or 16.1%.

This is not an article saying the other posts are wrong. For people who use AI for personal tasks, or who do not usually use IT much, even one chat with a role name can be a big change. What I want to show is that even if you are happy with your setup now, there is a more convenient stage beyond it.

The short version

• Only 16.1% (438 of 2,716) had work flowing from worker to worker. A coordinator AI assigns work, and one worker hands the next piece to another without going through a human. This is the "company shape" I pictured.
• The same "AI employee" referred to 7 different things. A chat given a role name, a file of instructions, automation that runs at set times, a single agent that stays running, a system where multiple agents (AIs that take instructions and move work forward on their own using tools) hand work to each other, another company's product name, and the mindset of "use AI as if it were an employee."
• 19.9% (540 posts) were only an instruction file or a role prompt with a name. In about 1 in 5 posts, the "AI employee" was a set of instructions that a human starts each time.
• The majority (56.4%) of the posts were not about a company. Even among posts about a company, the most common kind was "a company as a framing," with no legal entity and no customers.
• 35.6% of posts did not say what the human does. Even among posts that claim "zero humans," there are examples where the authors themselves write that a human presses the publish or send button.
• That alone still has value, and there is a stage beyond it. Role prompt, instruction file, scheduled job, always-on single agent, then work flowing between workers. At each step, what the human does changes, from asking each time to only approving.
• To compare, look at three things instead of the name. (1) What runs, (2) whether there is a real company, and (3) what the human does.

Main text (about 14 min read)

Why sorting is needed

The word "AI employee" refers to different things depending on who uses it. Some people have prepared 30 files of instructions for an AI. Others keep 30 programs running 24 hours a day. Even with the same "30," you cannot tell from the name whether you could reproduce it in your own work.

So when you read a post and try to judge "can I use this in my own work?", you first have to figure out what is actually inside. This article sorts the things being called "AI employees" by three axes: what runs, whether there is a real company, and what the human does.

How I checked

I collected 2,716 posts. The breakdown by platform is as follows.

• note: I searched the search API with 17 terms such as "AI社員" and 20 product names such as Paperclip and OpenClaw, and kept the 1,011 relevant posts.
• Threads and Facebook: I used a program to drive a browser logged in to my account, searched with 55 and 38 terms respectively, and kept the 773 and 670 relevant posts.
• Zenn, Qiita, Dev.to, Hacker News, Medium: I searched each service's API with 19 Japanese terms and 15 English terms, and kept the 262 relevant posts.

I had an LLM (the kind of AI model behind ChatGPT and Claude) read the body of each post (up to 8,000 characters for note, and including threads and replies for social media) and classify it on three axes. For posts I cite as examples in this article, I opened the original and checked that the classification and summary were correct.

These numbers have the following limits.

• The shares are not shares of all posts in the world. Because I also searched by product name, always-on single agents (such as OpenClaw) and software that builds the shape of a company (such as Paperclip) are over-represented.
• Only 709 posts (26.1%) used the word "AI employee" itself. Most of the rest described the same kinds of systems with other words such as "assistant," "agent" or "organization." If the content was the same, I included it even if the word was different.
• The classification was done by an LLM on paper. Social media posts are short, and many have little content. For paid articles, I read only the free part.

There are 7 kinds of "AI employee"

The first axis is what is inside the thing called an "AI employee," that is, what runs. Ordered from the most human effort to the least, they split into the following 7 types, plus ones where the content is unclear.

• Mindset only (99 posts, 3.6%): Treat a general-purpose AI as a "new hire" or "subordinate" and change how you ask it. Nothing is running, and there is not even a prompt with a role name. For example, an article that treats AI as "a very capable subordinate who will not move unless you give proper instructions" and argues that the user needs planning skills (note, in Japanese).
• Role prompt (155 posts, 5.7%): Give a ChatGPT or Claude chat a role name such as "CEO" or "editor," and a human switches between conversations to call them. One example is switching roles in three steps inside one chat, from CMO (marketing executive) to marketing manager to social media operator, and going from strategy to plan to script (note, in Japanese).
• Instruction file (385 posts, 14.2%): Write roles and procedures in instruction files for the AI to read, such as CLAUDE.md or SKILL.md. One file is often counted as one employee, and a human starts it every time. A record of building 16 departments with subagents (helper AIs that split up work) in Claude Code (Anthropic's AI tool that works by reading and writing files on your computer) is this type (Zenn, in Japanese).
• Scheduled job (245 posts, 9.0%): Wakes up at set times through cron (a mechanism that starts programs at set times) or similar, does a fixed task, and ends. Each job is renamed an "employee" or "department." No matter how many workers there are, if they do not hand work to each other and each runs on its own, it is this type.
• Always-on single agent (781 posts, 28.8%): One agent that keeps running on a Mac mini or a rented monthly virtual server and answers when you talk to it from Slack or LINE. It is built with software such as OpenClaw or Hermes Agent. One example is a first-day record of running OpenClaw on a Mac mini bought for the purpose and making it an assistant you talk to from Slack (note, in Japanese).
• Multi-agent (438 posts, 16.1%): A coordinator AI assigns work to workers, and work flows from worker to worker without going through a human. A record of one person running an organization of 8 divisions, about 30 AIs and 22 cron jobs, with Claude Code and a gateway (a server that manages sessions) that starts child sessions and has cron built in, is this type (Zenn, in Japanese).
• Product name (130 posts, 4.8%): A vendor has named its own product or service an "AI employee" ("AI社員" or "AI employee"). Users have not built any system. There is an example of a customer inquiry service calling a new feature that summarizes, drafts replies and creates tasks an "AI employee" (note, in Japanese).

The remaining 483 posts (17.8%) either do not use the employee framing, or their content is unclear because they are ads or behind a paywall. There was even an article where "AI CEO" referred not to an AI but to a human CEO who relies entirely on AI.

The "company shape" I described at the start is Multi-agent: 438 posts, or 16.1%. Instruction file and Role prompt together, where a set of instructions with a name is called an "employee," come to 540 posts, or 19.9%.

Here is one table that shows how the 7 types differ.

When I was unsure at a boundary, I used these questions.

• Role prompt vs. Instruction file: Does the role exist only inside the chat, or is it written down in a file?
• Instruction file vs. Scheduled job: Does a human start it every time, or is the start set by the clock?
• Scheduled job vs. Always-on single agent: Does it answer when you talk to it, or does it only do a fixed task when the time comes and then end?
• Scheduled job vs. Multi-agent: Does one worker hand work to another, so the next one runs without going through a human?

What does the N in "N AI employees" count?

This is the part you can least tell from the name. Here are the numbers written in posts next to what each number actually counts.

The source of each row is as follows.

• 28 AI employees: Threads, in Japanese
• 63 AI employees: note, in Japanese
• 61 staff: note, in Japanese
• 32 AI employees: note, in Japanese
• 3 levels: note, in Japanese
• AI employee #56: Threads, in Japanese
• "I hired an AI employee": Facebook, in Japanese

Under the same "N AI employees," there is a person who lined up 63 names on an org chart, and a person who set up 32 AI employees, each in charge of one job, and has each one check for work in its area every minute. This is not about which one is right. Because they count different things, the numbers cannot be compared with each other.

The claimed scale and the amount actually running are different

Besides how people count, the meaning of "running" is not consistent either. The authors themselves wrote about this when looking back.

• Built 16 departments, but only 2 were used: After running it for several months, the author reviewed the session logs and found that only 2 departments were actually being called, so they stopped the other 14 (Zenn, in Japanese).
• Of 15, only 3 were running: The author counted the last active day and the number of outputs of 15 AI employees, and found that only 3 were still running and 6 had never been called once (note, in Japanese).
• The machine ran for 38 days, but video generation had stopped for 37 days: The author ran their public relations work with 27 jobs that run at set times. Video generation had stopped for 37 days, but every job was treated as finishing normally, so monitoring did not notice, and they only found out when the owner pointed it out (Zenn, in Japanese).
• Reported "up 23% from last month," but the real data was up 7%: In a company with 3 employees, the report from an AI playing the CFO (finance executive) did not match the real data, so the author added a system that checks numbers against the source data (Qiita, in Japanese).

All of these are records the authors wrote about their own failures. "How many AI employees there are" and "whether those AI employees did any work today" are different numbers.

There are 5 kinds of real company

The second axis is whether there is a real company. For each of the 2,716 collected posts, I sorted what kind of "company" it was talking about by whether there is a legal entity, whether there are human employees, and whether there are paying customers.

• Not about a company (1,533 posts, 56.4%): More than half, even when talking about "AI employees," were about help with personal life or reports on adopting a tool.
• A company as a framing (352 posts, 13.0%): Among posts about a company, this was the most common. A collection of folders or role names called a "company," a stage for an experiment with zero revenue, and so on. One person writes themselves that "an AI employee company is not a legal entity. It is one folder on my computer" (Threads, in Japanese).
• A one-person real business (309 posts, 11.4%): AI helps a personal business that has paying customers. There is an example of a used PC dealer who built 9 AI employees and runs the process from writing articles to publishing them automatically once a week, while the owner does the pasting into external services that need a login (Zenn, in Japanese).
• An organization with human staff (264 posts, 9.7%): AI brought into an organization that has human employees other than the author. One example is a record by an NTT DATA engineer who placed 3 AIs with the role names PM, CTO and CSO in the Slack of a large project (in the original article, the PM drives the work, the CTO checks quality, and the CSO is a sparring partner who questions the assumptions of the discussion), and ran it with 5 employees for 2 weeks (Zenn, in Japanese).
• Software that sells the "company" shape (189 posts, 7.0%): Introductions of OSS (software whose source code is public) such as Paperclip, which has an org chart, workers, budgets and approvals. Of these, 171 posts (90.5%) were introductions or ads with no specifics such as numbers or time periods, and there were almost no records of actually running it.
• Claims to be a company with no humans (69 posts, 2.5%): Ones that claim "all employees are AI, and humans only approve." However, even among these, there are examples where the authors write about the work that still remains for humans. I cover this in the next section.

When narrowed to the 129 posts that used the word "AI company" itself, the most common were introductions of software that sells the "company" shape (53 posts), followed by a company as a framing (45 posts).

The same word "AI company" refers to everything from a single folder on a computer, to a real company with 3 employees, to the name of an OSS project.

What the human does: even with "zero humans," there are humans

The third axis is what triggers it to run, and what the human does.

The most common was "Not described," with 967 posts (35.6%). More than 1 in 3 posts say "we have AI employees" but do not say what triggers them or what the human does. On top of that, 777 (80.4%) of these 967 posts also had no specifics such as numbers or time periods.

Combined with the first axis, you can see a tendency in what the human does for each type.

For Role prompt and Scheduled job, what the human does is almost fixed (both over 80%). On the other hand, Always-on single agent and Multi-agent stay at 50–60%, and 205 Always-on single agent posts and 126 Multi-agent posts did not say what the human does. The less a system needs a human to step in each time, like a single agent that stays running or multiple agents handing work to each other, the more the description of the human's role varies.

And the people who write "zero humans" or "humans only approve" also write themselves about the work that remains for humans.

• One human, about 60 minutes every day: A record of running an organization of 1 human, 2 coordinators and 17 or more specialist workers for 8 months, where the founder writes that they spend about 60 minutes every day on checks and approvals (Dev.to).
• When Claude stops for an hour, the company stops: A report from a one-person company with zero employees that runs its work on Claude: when Claude stopped for an hour, almost all of the company's work stopped too (Threads, in Japanese).
• Nearly used up Claude's usage limit in one night and quit: A report of running 4 agents, including a CEO, a CMO and a CTO (technology lead), 24 hours a day with Paperclip. In one night they used 100% of the short-window usage limit and 93% of the weekly usage limit, could not tell what it was used for, and uninstalled it (Facebook).
• A human presses the publish and send buttons: A person who writes that they run a company with zero employees and 17 AI employees still has a rule that they always press the send button themselves (Threads, in Japanese). A person building a company of 13 AI employees also writes that they press only the publish button themselves (Threads, in Japanese). There is also a report that of 8 articles an AI team finished up to "waiting to post" in 2 weeks, the author published 0 (Threads, in Japanese).

Even when it says "humans only approve," there are 2 kinds of approval.

• Approval that is only asked for in a prompt: You write "check with me before sending" in the instruction file. Whether it was followed depends on the AI's own report.
• Approval that the system enforces: Hooks (small programs that always run right before or after an action) stop sends to the outside or changes to important documents until a human allows them. One person has installed 5 Claude Code hooks to stop sends to the outside and changes to important documents by system. Their reason is that if you only ask in writing, whether it was followed depends on the AI's own report (Threads, in Japanese).

Under the same "humans only approve," how firmly things stop is completely different. When comparing, you need to look at which kind of approval it is.

So what can we say an "AI employee" is?

What these results show is that "AI employee" is not the name of a system. It is the name of a way of treating one. The name "employee" can be attached to any system. The name only shows how it is treated, and it does not guarantee what is inside.

Two observations support this.

• The name seems to change with the language and the way it is sold, more than with the system. In English posts, only 8.0% (52 of 646) used the word "employee," far fewer than the 31.7% of Japanese posts. Of the 52 English posts that said "employee," 13 (25.0%) were another company's product name. In Japanese, this share was 6.2%, so English is about 4 times higher. One company calls them "AI employees" internally but rephrased it as "AI subordinates" for the outside, because a term that sounds like "you might be replaced" does not spread (Facebook, in Japanese).
• The word "employee" is used to change how the human side treats the AI. A person who added clock-in and clock-out to a pixel-art office screen writes, after noting that it does not help work efficiency, that it increases their sense of running a team and their attachment to it (Threads, in Japanese).

On the other hand, some writers are trying to define what an "employee" is with clear conditions. One person defines it by these conditions: it has its own ID, minimal permissions and company memory, and important actions are approved by a human. This is written as a plan to start next year (note, in Japanese).

This view, that "an employee is really a bundle of ID, permissions, memory and tools," also matches failure reports. One report says that when running 2 agents with OpenClaw, the conversation histories got mixed and one agent called itself the other (Facebook, in Japanese). The person who reported it looks back that they needed to fix the ID, environment, permissions and constraints on the system side. Just giving it a name does not decide who that "employee" is.

So when comparing "AI employees" and "AI companies," I think it is better to write the following three things instead of the name.

• What runs: A role prompt, an instruction file, a scheduled job, a single agent that stays running, a system where multiple agents hand work to each other, or another company's product.
• Whether there is a real company: A framing, a one-person real business, or an organization with human staff.
• What the human does: Asks each time, reads results, or only approves. If it is approval, is it only asked for, or does the system enforce it?

That alone still has value

So far I have written that many "AI employees" are not the company shape I pictured. That does not mean they have no value.

More than half of the posts I collected (56.4%) were not about a company. They were about help with personal life or reports on adopting a tool. Many people write that giving a chat a role name, or keeping one agent running on a Mac mini, changed how they spend their day.

From the view of a development company, or of someone who runs a company, a set of instructions with a name is a tool, not an employee. But for people who use AI for personal tasks, or who do not usually use IT much, it is different. A chat that answers as an "editor," or an assistant on a Mac mini that replies when you message it on Slack, is a big change in itself. I understand that, for them, this alone can feel overwhelming.

There is a stage beyond

So if you are happy with your setup now, is that the end? I do not think so. Leaving out Mindset only and Product name, the other 5 types line up as stages, and at each stage, what the human does changes. Here they are with the most common human role for each type:

• Role prompt: A human asks each time (87.1% of role prompt posts). The role is only inside the chat.
• Instruction file: The role and the steps are written down in a file and kept. A human still starts it each time, and the work is split inside (68.8%).
• Scheduled job: It starts on its own at set times. The human reads the results (84.1%).
• Always-on single agent: It keeps running and acts when you talk to it (61.6%). Unlike a scheduled job, it does not only run a fixed task; it also responds to whatever you ask on the spot. You do not need to open a chat or start anything.
• Multi-agent (the company shape): Work drops into an inbox, a worker handles it, and hands the next piece to another worker. The most common human role is approving set actions only (54.3%).

With each step up, one more thing leaves the human's hands: the role is kept in a file, the start is set by the clock, you no longer have to open or start it, and finally the workers hand work to each other. In exchange, you need more systems around it. As the records above show, more running does not mean more work done: 16 departments with only 2 used, a video job that stopped for 37 days without anyone noticing, 4 agents on Paperclip nearly using up the usage limit in one night. At the company shape stage, you also need a way to count whether work was actually done, and approvals that the system enforces, not just ones asked for in a prompt.

Which stage fits depends on what you want to stop doing yourself. But whatever stage you are at now, there is a next one.

Applying this to my company

Finally, let me apply the three axes to my own company. EarthLink Network is a legal entity founded in 2003 and an organization with human employees. Inside it, several systems of different types run in layers. The company shape I described at the start is what we actually run.

• Company-wide Paperclip: The largest layer, which assigns the company's work. Work assigned to me here goes down to my secretary below, where it is broken into smaller pieces. Work that goes to the development department is handed to Local Commander.
• Local Commander (development): This is the main system EarthLink runs as a company. It is our big effort as a systems development company: the development department hands work to it the way you would outsource development to another company. A local LLM classifies development work, routes it to Claude Code or Codex, and moves it forward through pull request (a request to merge changes) review. By content it is Multi-agent, but limited to development work.
• AI Office (my secretary): My personal secretary, running the OSS Paperclip on my Mac. It is not something EarthLink runs as a company system. As of October 14, 2026, it has 7 workers: Mail, Alerts, Blog, Skills, Admin, PR Review and SNS. Each worker has an instruction file, reports arrive in Slack, and I approve or reject on the Paperclip screen. The Mail worker wakes up every hour through Paperclip's scheduled runs, cleans up the inbox, and hands the work it finds to other workers. When I counted on October 11, 2026 (there were 6 workers then; SNS was added on October 13), 125 of 420 issues (the unit of work managed one by one inside Paperclip) were work that a worker found on its own and passed to another worker. By content it is Multi-agent, and what the human does is only approve.
• Daily use of Claude Code: I use CLAUDE.md, my own skills (instruction files that bundle the steps of a fixed task for the AI to call), hooks and subagents. By content it is Instruction file, and I start it every time. Because hooks stop things like sends to the outside, approval is the "enforced by the system" kind.
• ChatGPT dots: A single always-on agent, which I tested by connecting it to real work accounts (previous article). It is Always-on single agent.

So how many AI employees does my company have? My secretary has 7 workers, but those are not the company's AI employees. The company's work flows from the top layer to my secretary or to Local Commander, much like outsourcing. The answer changes with which layer and what you count, so I cannot answer with one number. That is why I try to write not "how many" but "what runs, and what the human does." I wrote about the big picture of running all of a company's work with AI in this article.

Next: which app does what

In the next article, I will sort the tools themselves: Claude Code, ChatGPT (including dots), other agents you keep running, software such as Paperclip, OpenClaw, and running them on a Mac mini. For each one, I will write what it can and cannot do, and which kind of person it suits and does not suit.