Conclusion First

The era has arrived where a business system that costs 3 million yen to outsource can be created in-house for just 50,000 yen.

When you think about creating a business system and request a quote, you often receive an estimate of 3 million yen. For small and medium-sized enterprises in rural areas, this is a heavy burden. A few percent of their sales can vanish just like that. So they endure it. They manage with Excel. They become reliant on specific individuals. If someone leaves, it’s game over.

This structure is beginning to break down.

With JetBrains’ free coding agent model “Mellum,” local LLMs, and the evolution of coding agents like Claude Code and Cline, the situation where “business systems can be created without programmers” is becoming a reality.

The question is simple: What does the 3 million yen estimate consist of, and what can be eliminated?

Breaking Down the 3 Million Yen Outsourcing Estimate

First, let’s take a look at the breakdown of the 3 million yen outsourcing estimate. For a typical business application for small and medium-sized enterprises (such as inventory management, customer management, daily report management, etc.), it roughly looks like this:

Pay attention here. Coding itself accounts for only 30% of the total. The remaining 70% is the “cost of humans working for humans.” Meetings, minutes, specifications, test specifications, progress reports. The very structure of outsourcing generates costs.

In other words, simply “stopping outsourcing” can eliminate a large portion of the cost structure. The problem is, if we stop outsourcing, who will create it? The answer is coding agents.

Costs Eliminated and Costs That Cannot Be Eliminated by Coding Agents

Coding: 900,000 yen → Almost 0 yen

JetBrains’ “Mellum” is a lightweight model specialized for code completion. With around 4 billion parameters, it can run on a local PC. While it is not all-encompassing on its own, combining it with coding agents like Claude Code and Cline changes the game.

The mechanism of coding agents works like this: When you instruct them in natural language, “Create an inventory management app that can manage product names, quantities, and in/out dates,” the agent autonomously creates files, writes code, fixes errors, and outputs a functioning app.

I tried it out. When I instructed Claude Code to “Create an inventory management app with Next.js + Supabase,” an app with CRUD functionality was up and running in about 15 minutes. The API usage cost was around 200 yen. The human coding effort was just the 5 minutes it took to write the instruction.

Of course, additional adjustments are needed to bring it up to a usable level for business. However, the process that previously cost 900,000 yen now falls within the range of a few thousand to tens of thousands of yen. This is not just a minor difference; it represents a structural change.

Testing & Quality Control: 450,000 yen → A few thousand yen

Coding agents also automatically generate test code. If you instruct them, “Write unit tests for this app,” they will produce tests covering the main patterns. While not perfect, they are sufficient for the quality level required in business applications for small and medium-sized enterprises.

In outsourcing, you create test specifications, conduct manual tests, generate bug reports, make corrections, and retest. This back-and-forth costs 450,000 yen. With an agent, tests can run immediately after the code is written. If a bug is found, you can issue a correction instruction on the spot. The time required is just a few minutes, and the cost is a few hundred yen.

Requirements Definition & Design: 600,000 yen → Do it yourself (0 yen)

This is the turning point. In outsourcing, there is a cost associated with conveying “what we want to create” to others. This costs 600,000 yen, involving five meetings, minutes, specifications, and screen design documents.

In-house, this communication cost becomes zero. The people who know the business best are the ones doing it. “When stock comes in, it increases, and when it goes out, it decreases. At the end of the month, we just need a list of inventory” — for requirements like this, you can simply write it in a chat.

A 100-page specification document is not needed for business applications in small and medium-sized enterprises. What is necessary is the ability to write down “what we want to do” in three lines.

Project Management & Communication: 450,000 yen → 0 yen

Weekly meetings with the outsourcing partner, progress checks, negotiations for specification changes — all these costs arise from “having someone else do the work.” If you create it yourself, they are unnecessary.

Infrastructure: 300,000 yen → A few thousand yen per month

Supabase, Vercel, Cloudflare. All of them have free tiers. For a business application typical of small and medium-sized enterprises, it can be operated for 0 to 3,000 yen per month. The initial setup cost of 300,000 yen disappears.

Breakdown of the 50,000 Yen

So, what does the 50,000 yen for in-house production get spent on?

Even including the time cost of the person in charge, it totals 50,000 yen. This is one-sixtieth of outsourcing costs.

Answer to “But We Don’t Have Engineers”

This counterargument will inevitably arise. However, consider this:

What is required for coding agents is not programming skills. It is the ability to articulate “what you want to create” and the ability to notice discrepancies when interacting with something that works. This is a capability that the people on the ground who are running the business possess the most.

In fact, in a local manufacturing company we support (with 30 employees), the general affairs staff used Claude Code to create a daily report management app in three days. They had zero programming experience. All they did was instruct in chat, saying, “I want a screen like this” and “This display is strange.”

Not having engineers is no longer a valid reason for not having a system.

The Ease of Local LLMs Accelerates Development

The significance of lightweight models like JetBrains’ Mellum running locally is substantial. You can experiment as many times as you want without worrying about API usage fees.

When using external APIs, there is a psychological barrier of “if I fail, it will cost money.” With local LLMs, the only cost is electricity. The cost per inference is less than 0.01 yen. Even if you try 100 times a day, it costs just 1 yen. This “ease” dramatically increases the learning speed of non-engineers.

Code completion is processed locally with Mellum, while only complex design decisions use Claude’s API. This division further reduces API usage costs.

This is Not “DX”; It’s a Survival Strategy

Large corporations spend tens of millions of yen to commission systems from SIers and take years to implement them. Small and medium-sized enterprises do not need to imitate this.

Create it for 50,000 yen, and if it doesn’t work, throw it away and start over. This state of “trial and error costs being almost zero” becomes a weapon for small and medium-sized enterprises. Large corporations have approval processes and bureaucracies. Small and medium-sized enterprises have the ability to “think of something today and act on it tomorrow.”

Coding agents accelerate this speed even further.

So, What Should We Do?

1. First, try Claude Code or Cline. The free tier is sufficient. Choose one “most troublesome task in your business” and instruct, “Create an app to manage this.”
2. Show the working product to the field. It doesn’t have to be perfect. Changing the perception to “We can do something like this” is the first step.
3. Install JetBrains’ Mellum locally. Experience the speed and accuracy of code completion. Create an environment where you can experiment infinitely with zero API costs.
4. When you receive an outsourcing estimate, first try in-house production. If it doesn’t work after three days, then outsource it. However, in many cases, you will end up with something functional in three days.

The cost of 3 million yen becomes 50,000 yen. This structural change has already occurred. The difference lies in whether you know it or not, and whether you act on it or not.