The Org Chart of a Company Running Many AI Agents: Three Things the President Decides

More and more companies are placing many AI agents, one for each role. Until recently, a person asked for something, AI returned an answer, and when the conversation ended, the work ended too. Now agents keep running, pick up the work in their own area, and leave a report when they finish. AI is changing from a single handy tool into a group of staff members, each with their own post.

When the number of staff grows, a company needs an org chart. In this article I lay out, as my own decision criteria, how a company should rebuild itself when the people it hands work to become many agents, and what the president should decide. On October 20 I talked about how a company's speed is set by the round-trips of human decisions, and on October 21 about taking risks with reversible operations and parallel running. This piece goes one step further.

Two examples from overseas

The first is Paperclip, an open-source system for running a business with a team of AI agents. According to its public documentation, you place role-based agents under a company goal, give them roles, reporting lines and permissions, and view them as an org chart. You can set a monthly budget for each agent; when spending approaches 80% of the cap a warning appears, and at 100% that agent is paused automatically. The overall status can be checked on a single dashboard.

The second is ECC. According to the description of a short video introducing it, it was built by a single developer and uses 64 agents and 261 skills to automate architecture planning, writing code, security testing and test coverage. The numbers are only what the video description says. What matters is not the count, but the way the work of development is split into posts with someone in charge of each.

The waste: adding AI to a human-centered work design

What I find most wasteful is designing work around people and inserting AI somewhere in the middle. The current procedures, approval flow and screens stay as they are, and only the part a person used to do is replaced with AI. It looks like a natural way to proceed, but AI then keeps waiting inside a flow built for people, on people's schedule.

Imagine AI drafting a quotation in seconds, and then a person retyping that draft into another screen, routing a paper approval and only then sending it. Even if the AI part gets faster, the total time hardly changes. And because AI never learns the retyped result, it cannot use it for the next job. I believe companies should rebuild their organization and systems around resident AI agents.

Keeping the existing systems makes it pointless

I think introducing AI agents while keeping the existing systems makes little sense. Existing systems assume a person looks at the screen, a person clicks, and a person remembers. Information is scattered and you cannot tell which is the latest; permissions for who may do what are vague; what was done is not recorded, so next time nobody can continue from where things stopped. A person asks around and manages, but an agent cannot know what is not written into its environment.

At my company, we rebuilt our internal systems in-house as systems for AI agents. Rather than adding agents to the old setup, we replaced the foundation itself with one that agents can read, write and move in easily. As a result, agents now take the initiative on a great deal of work, and the company is doing very well.

The president decides three things

Once there are many agents, it is impossible for the president to step into every decision. If the president tries to see everything, the president becomes the biggest waiting time in the company. So the president decides not individual calls, but the frame in which decisions are made.

The first is the responsibility chart. Write on one sheet which agent handles what: its post, the scope it is trusted with, what it must not do, and who it reports to when stuck. Overlapping posts mean the same job gets done twice, and work that belongs to no one is left untouched. Start with about three broad posts and refine them while reading the records.

The second is the budget. Decide the cap for the whole company and for each agent, and what happens when a cap is exceeded. Start with modest caps and raise them as you see the records and results. It helps to think of a budget as the amount of trust you give an agent.

The third is where reports and records go. I consider this the most important. Have agents leave what they did, what they decided and what remains in a fixed place and format, and decide who looks at it and when. Records are not only for people; for agents they serve as memory and as a handover note. I myself place agents per company and per project, and have them keep work logs in our shared internal knowledge base. That way, without watching every task, I can read the records and see what is moving and what has stalled.

Now is the time to take risks

In one sentence, my criterion is to make up your mind that now is the time for risk-taking. Rebuilding the foundation takes effort, and some parts will not go well. Even so, I believe this is the period in which companies decide whether they will be weeded out or be the ones who do the weeding out. While you wait and see, the companies that have rebuilt keep accumulating records and getting faster.

After employees' working environment comes AI agents' working environment

Companies have long discussed how to make work easier for employees. That remains important. On top of that, I believe we should now pursue a good working environment for AI agents: a clear place to receive work, defined tools and permissions, the necessary information gathered in one place, and a place to record what was done so the next run can continue from there. It is time to build a company where agents can work well, with the same energy we put into building workplaces where people can work well.

What you can do this week

Write down three jobs you want to hand to AI, one per role. For each, fill in one line each for its post, its monthly cap and where its records go. Any blank you cannot fill is a place where agents will struggle to work. Then put an agent in charge of just one role and look at its records for a week. You will start to see overlaps and gaps in the chart, and get a feel for the budget.

▶ Watch the full video (Japanese)
https://www.youtube.com/watch?v=K6QwYZgWdT4