A New Era Where a Company Can Operate on the Salary of Just One Employee

To get straight to the point: by combining five AI agents, you can manage the workload equivalent to five employees for a monthly cost of 80,000 to 120,000 yen. This is less than the salary of a single part-time worker.

This isn’t a distant future scenario. There are already examples overseas where solo entrepreneurs are successfully managing businesses with a dozen AI agents. Blake Lapierre, for instance, utilizes around 17 AI agents to run his business with zero employees. Customer support, marketing, data analysis, content creation—tasks that once seemed impossible without human involvement are now being handled entirely by systems.

So, what does this mean for small and medium-sized enterprises in rural areas?

In regions where “finding employees” has become the biggest management challenge, the option to run operations through systems instead of relying on human resources has now become realistically accessible at a manageable cost.

Breaking Down the Costs: What Can Be Achieved for How Much?

No need for abstract discussions. Let’s look at concrete numbers.

We’ll estimate the monthly costs for a five-agent system, breaking it down by task. This includes the API usage fees for LLMs (such as GPT-4o, Claude, etc.) plus the costs of associated tools.

Subtotal: 29,000–51,000 yen per month

Adding server costs (cloud) of about 3,000–5,000 yen and monitoring/workflow management tools (like n8n, Make, etc.) at 5,000–10,000 yen, we can roughly estimate—

Total Monthly Cost: 40,000–70,000 yen. Even with generous estimates, it’s 80,000 to 120,000 yen.

For comparison, hiring one administrative staff member in rural areas would cost around 250,000 to 300,000 yen per month, and for five employees, it would be 1,250,000 to 1,500,000 yen. The cost of a five-agent system is 1/15 to 1/20 of the labor costs.

This difference is not just a slight saving; it represents a structural change.

The Design Philosophy of “Running on Systems”: The Significance of 17 Agents

The reason Blake Lapierre uses 17 agents is that he focuses on one agent per task, rather than relying on a single all-purpose AI.

This should be intuitively clear to those who manage operations. A person who claims to “do everything” usually cannot do anything deeply. The same applies to AI; agents with narrowly defined roles tend to have higher accuracy and a smaller impact when they fail.

Thus, the design principles of AI agent management are as follows:

1. One Agent = One Task (Narrow and Deep)
2. Automate handoffs between agents using workflow tools
3. Humans only review at designated “review gates”

The third point is crucial. If the CEO has to check every output, it ultimately reverts back to the “individualization of the CEO.” Therefore, it’s essential to design checkpoints (review gates) where humans verify only critical outputs, allowing everything else to flow automatically.

For example, using a CLI-based task manager like “Tkmd,” one can review a list of tasks processed by AI agents and pick out only anomalies for judgment. The CEO’s workload can be compressed to about 30 minutes to an hour a day.

Zero individualization does not mean “no humans”; it means “humans only need to make judgments.”

Identifying Weak Points: The Vulnerabilities of AI Agent Management

It’s not useful to only highlight the positives. Let’s honestly discuss the potential pitfalls.

1. Customer Interactions Requiring Complex Contextual Judgments

Handling complaints, nuanced negotiations, and emotional customers—this is where AI struggles the most. While 80% of standard inquiries can be managed by AI, the remaining 20% must be handled by humans. Ignoring this 20% can lead directly to customer attrition.

2. Miscommunication Between Agents

As the number of agents increases to five or ten, errors can occur in data handoffs. The most concerning issue is a “silent failure” where customer information meant for the CRM does not reach the marketing agent. This can result in incorrect emails being sent to customers without anyone noticing.

Countermeasure: Implement a “health check” that cross-checks logs of all agents weekly.

3. Risks of Changes in LLM Specifications or Price Increases

The API prices of OpenAI and Anthropic have fluctuated significantly over the past two years. While there is a trend toward lower prices, the sudden discontinuation of specific models or changes in specifications can halt workflows unexpectedly.

Countermeasure: Avoid dependency on a single LLM. Design workflows to operate on two or more models for major processes.

4. The Issue of “If the CEO Falls Ill, Everything Stops”

Even with zero individualization, if the sole decision-maker at the review gate is the CEO, the company will come to a halt the moment the CEO is hospitalized. This poses the same risk as traditional one-person management.

Countermeasure: Document the criteria for review gates so that external partners can step in during emergencies.

What This Means for Small and Medium-Sized Enterprises in Rural Areas

I want to ask the CEOs who think, “This doesn’t concern us.”

How many administrative staff members do you currently have? What percentage of their work consists of “routine repetitive tasks”? Likely, around 60-70% is repetitive. If you could automate that portion for 50,000 yen a month, what could you do with the saved labor costs?

I want to clarify that this is not a call to “cut staff.”

The real challenge for small and medium-sized enterprises in rural areas is the inability to hire. Even when job postings are made, no one applies. Even if they do, they don’t stay long. That’s why it’s necessary to think about filling the gaps with systems where hiring is not possible.

Moreover, tasks that are systematized with AI agents automatically generate manuals. The workflow itself serves as a procedural document. When people leave, the know-how does not disappear. This is the true value of achieving “zero individualization.”

So, What Should Be Done?

There’s no need to set up 17 agents all at once. Start with just one.

Choose the most repetitive and boring task. Delegate that to an AI agent.

For example, the initial sorting of inquiry emails received daily. Or creating reports that format monthly sales data into a standard template. This can be started with the ChatGPT API and Google Apps Script for less than 1,000 yen a month.

Once you have one task running, add a second. Then a third. Before you know it, you’ll find yourself in a situation where, “Oh, this is running smoothly with just one person.”

The important thing is to not aim for perfection from the start. AI makes mistakes, but so do humans. AI’s mistakes leave logs that can be improved upon. Human errors are individualized and cannot be replicated. It’s clear which is stronger as a “system.”

With a 50,000 yen experiment, there’s a possibility of saving 250,000 yen in labor costs. Is there any reason not to test this ROI?