Conclusion

To put it simply, the decision of whether to hire a person or operate AI has begun to shift at the 100,000 yen per month line.

For small and medium-sized business owners, the cost of hiring a single employee is significant, ranging from 300,000 to 500,000 yen per month, including social insurance. The time spent on recruitment, training costs, and the risk of turnover all add to this burden. However, now there is a realistic option to “operate an AI agent for several tens of thousands of yen per month” for specific tasks.

This is not just about “improving operational efficiency with AI”; the cost structure itself has changed.

In this article, we will discuss three examples: the coding agent “GoatCode” that runs on 85MB, the self-testing AI “Devin,” and the full-duplex phone AI that extracts information while conversing simultaneously with humans. What these three examples demonstrate is the fact that tasks that were once entirely dependent on labor costs—such as “writing code,” “validating code,” and “answering phone calls”—are beginning to become significantly cheaper.

—

GoatCode Turns 85MB Binary into a ‘Coding Factory’

GoatCode is an open-source terminal AI agent. What stands out is its size: a mere 85MB. It can run on a single laptop without the need for a large cloud environment.

What can it do? Code generation, debugging, and refactoring. It can connect to over 180 LLM providers and automatically fall back to another provider when the quota of one API is exhausted. In other words, work does not stop. Human engineers will not encounter a situation where they say, “I hit the API limit, so that’s it for today.”

Let’s consider the impact on small and medium-sized enterprises in numerical terms:

• If you outsource a minor modification of a web application, it costs between 200,000 and 500,000 yen per project.
• Hiring a full-time engineer costs between 400,000 and 600,000 yen per month (including social insurance).
• Using GoatCode with a pay-per-use LLM API can cost only a few thousand to tens of thousands of yen per month.

Of course, whether the code produced by GoatCode can be deployed directly into production is another matter. Reviews are necessary. However, when the cost of “writing code” itself has decreased to one-hundredth, what happens? If there is one reviewer, that person’s productivity can increase tenfold.

A job that required a team of ten developers can now be handled by one reviewer and a few instances of GoatCode. This is not “efficiency improvement”; it is a “structural change”.

—

The Significance of Devin ‘Writing and Running Its Own Tests’

Devin is an AI software engineer developed by Cognition. Not only does it write code, but it also generates its own test code, executes it, and verifies the results. The cycle of a human engineer writing, testing, fixing, and retesting is autonomously managed by the AI.

What does this change?

In software development, the ratio of time spent writing code to time spent on testing and reviewing is generally estimated to be about 3:7. Writing takes 30%, while verification and correction take 70%. The fact that Devin can handle self-testing means that the AI takes over that 70%.

Consider this in the context of a small business where there is only one engineer. That engineer can focus on developing new features while Devin handles the testing. When a pull request comes in, the human only needs to verify the logic’s validity.

What about the monthly cost? Devin’s pricing structure is primarily based on usage, and for small to medium-sized development scales, it often falls within 50,000 to 100,000 yen per month. Compared to the 500,000 yen for a full-time engineer, this is less than one-fifth.

However, it is important not to misunderstand. Devin is not a tool that makes engineers unnecessary. It is a tool that multiplies the output of one engineer by five. Misunderstanding this can lead to failed implementations. A human decision-maker is always required. However, we have entered an era where most of the hands-on work can be entrusted to AI.

—

A World Where Phone Handling ‘Ends Automatically’—Full-Duplex Phone AI

The third example may have the most significant impact on small and medium-sized enterprises in rural areas: the AI transformation of phone handling.

Traditional voice AI was “half-duplex”. While the AI was speaking, it could not hear the human voice. When a human was speaking, the AI would remain silent. This led to unnatural conversations, causing people to realize, “Oh, this is AI,” and hang up.

Full-duplex phone AI is different. It can speak simultaneously with humans. It listens to the other person while interjecting, extracting necessary information, and responding. Conversations can flow in the same rhythm as those between humans.

Let’s consider the cost of phone handling for small and medium-sized enterprises:

• Hiring one administrative staff member to handle calls: 200,000 to 250,000 yen per month (part-time, including social insurance).
• Using a phone answering service: 30,000 to 100,000 yen per month (depending on the number of calls).
• Full-duplex phone AI: 1,000 to 5,000 yen per month (based on API pay-per-use).

Moreover, AI can handle calls 24/7. Nighttime inquiries, holiday reservation calls—customers that were lost due to “we’re closed” can now be captured by AI.

Reservation calls for restaurants, reception at clinics, inquiries about real estate properties—these “routine yet high-volume” phone tasks are the perfect domain for full-duplex AI.

In a case study from a local dental clinic that implemented this technology, it was reported that AI handled about 70% of the 200 reservation calls per month, reducing staff phone handling time from 40 hours to 12 hours per month. With the 28 hours saved, staff could focus on improving the quality of patient interactions.

—

What Changes at the ‘100,000 Yen’ Break-Even Point?

Let’s summarize the three examples discussed so far:

What is common among these is that the area where AI can take over human tasks is rapidly expanding below the line of 100,000 yen per month.

This is not just a story for large corporations. Rather, it is a structure where small and medium-sized enterprises are likely to benefit first. Why?

Large corporations have existing systems, organizations, and approval flows. Implementing AI requires formal approval, security reviews, and coordination with existing vendors. Their movements are slow.

A company with ten employees is different. If the president decides, “We will implement phone AI starting next week,” it can be done. There is no need for formal approval to install GoatCode on an engineer’s PC. The speed of decision-making directly translates to the speed of AI implementation. This is a structural advantage for small and medium-sized enterprises.

—

So, What Should We Do?

I do not want to conclude with just a suggestion to “implement AI.” Here are three specific actions you can take starting tomorrow:

1. Identify ‘Routine Tasks Costing Over 100,000 Yen Per Month’ in Your Company

Phone handling, data entry, report creation, code modification, inquiry responses. List tasks that cost over 100,000 yen per month in labor costs and have a certain degree of established patterns. These are the prime candidates for AI substitution.

2. Start Small with One Test

GoatCode is available for free. There are services to test phone AI starting at 10,000 yen per month. Instead of trying to replace all tasks at once, experiment with one task using AI for just one month. If it doesn’t work out, you can revert. The cost of failure is only a few tens of thousands of yen.

3. Separate ‘Decision-Makers’ from ‘Operational Mechanisms’

What you delegate to AI is “operations,” not “decisions.” AI writes code, but humans decide whether to deploy it. AI handles calls, but humans switch to handling complaints. As long as this distinction is maintained, the quality will not decline with AI implementation.

—

‘Hire or Operate’ is No Longer a Philosophical Debate

The fact that AI’s monthly cost has dropped below 100,000 yen means that there are areas where AI is cheaper than part-time labor costs.

This is not about “humans becoming unnecessary.” It is about “the nature of work that humans should do changing.” By handing routine tasks to AI, humans can focus on judgment, creativity, and building trust with customers. A company of ten can manage the work of fifty using AI. This is the optimal solution to the question of ‘hire or operate’ at this point in time.

While large corporations are going through bureaucratic processes, small and medium-sized enterprises can act. We have entered a phase where those who take action will win.