Conclusion

Let’s get straight to the point: the cost of ‘creation’ is beginning to collapse.

A business system that used to cost 3 million yen to outsource is now operational for just 50,000 yen a month.

This is not a story about the future. In workplaces that have mastered AI coding tools, this is already a reality.

According to a report released by Anthropic, the company is using AI to accelerate the development of AI itself. A survey by the third-party organization METR reported that the speed of improvement in AI coding capabilities has reached about twice that of traditional methods. AI is speeding itself up. Do you understand what this means?

The decline in development costs may progress not linearly but exponentially from here on out.

For small and medium enterprises, this is not a threat. It’s an opportunity.

—

What’s Happening—Interpreting Three News Items

1. Anthropic: AI is Automating AI Development

The latest report from Anthropic indicates the fact that “AI’s coding capabilities continue to improve without human intervention.”

Until now, software development has been a contest of how many talented engineers one can employ. Large companies have hoarded dozens of engineers earning over 10 million yen a year, leaving small and medium enterprises with the choice of either paying those outsourcing costs or giving up altogether.

That structure is beginning to break down.

As the quality of code written by AI improves, the scope of what can only be done by human programmers narrows. While complex system design and requirements definition are still human jobs for now, routine business systems—customer management, inventory management, daily report aggregation, invoice issuance—are now at a level where they can be sufficiently created by AI.

2. Google’s ‘Elephant-Goldfish Model’: The Shock of Specification-First

Google’s release of the EGM (Elephant-Goldfish Model) plugin has directly addressed the weaknesses of AI coding.

The traditional problem with AI coding is this: when instructed to “just write some code,” AI produces something functional. However, because it does not correctly understand the specifications, it leads to a hell of corrections later on with “this is not what I wanted.”

EGM has reversed this process. First, it requires writing the specifications. Then, after reviewing them, it writes the code.

Thousands of engineers within Google are already using it, and it has significantly reduced incidents of “development proceeding under incorrect assumptions.”

This has significant implications for small and medium enterprises. Why?

The biggest reason small and medium enterprises fail in outsourcing systems is “miscommunication of specifications.” The client conveys, “I want something like this” in vague terms, and the contractor interprets it as “it’s probably something like this” and creates it. When they see the delivered product, it turns out to be “not what I wanted.” Additional costs arise. This cycle can inflate an initial estimate of 1 million yen to 3 million yen.

By incorporating a specification-first approach like EGM into AI tools, “miscommunication of specifications” can be structurally reduced. Without outsourcing, someone who knows the internal operations can interact with AI to solidify the specifications and generate the code directly.

3. A 15-Year-Old Created an LLM Editing Tool: The Definition of ‘Creators’ Has Changed

A 15-year-old boy developed and released a tool that can edit and adjust the output of LLMs.

This news itself is small. However, what it suggests is significant.

The definition of ‘creators’ has changed.

In the past, creating software required graduating from a computer science university and accumulating several years of practical experience. Now, even a 15-year-old can create and publish tools. We have entered an era where one can complete a product by interacting with AI without specialized programming education.

This fundamentally invalidates the excuse of small and medium enterprises that “we cannot develop because we cannot hire engineers.”

—

So, How Low Will Costs Go?

Let’s take a closer look.

Of course, this is limited to “routine tasks.” Complex core systems or areas with strict security requirements will still require human engineers.

However, 80% of small and medium enterprises’ operations are “routine.” The costs in these areas could drop to one-tenth, or in some cases, one-hundredth. This fact alone is sufficient reason to take action.

—

However, There Are Pitfalls—The Issue of ‘Code Debt’

There are invisible risks associated with code written by AI: “code debt.”

AI is good at writing “working code” but still struggles with writing “maintainable code.” If you instruct it to “just create something” while the specifications are vague, a system may emerge that works but is internally chaotic.

Initially, this is not a problem. However, six months later, when a specification change is needed, no one understands the internals. When trying to make corrections, other parts break. Ultimately, it leads to a complete rewrite—this is the typical outcome of “code debt.”

The countermeasures are simple.

• Clearly define the specifications in Japanese before letting AI write (EGM’s approach)
• Have at least one person in-house who can read the generated code
• Clearly differentiate between “disposable” and “long-term operation.” Daily report aggregation scripts can be disposable. Customer management requires long-term operation, so invest in code quality.

There is no need to seek perfection. It is sufficient to draw the line on “where to allow debt and where not to.”

—

Three Things Small and Medium Enterprises Should Do Tomorrow

I won’t discuss abstract “DX promotion” topics. I will only write about what can be done starting tomorrow.

1. Automate One Existing Task with ChatGPT or Claude

“Monthly aggregation tasks,” “standard email creation,” “format conversion for estimates”—anything is fine. Try letting AI handle just one task. It will take about 30 minutes and cost 20 dollars a month. This alone will give you a sense of “what can be done with AI.”

2. Consider Whether the System You Plan to Outsource Can Be Created with AI

Next time you get an estimate for system outsourcing, simultaneously try to create the same thing using AI tools (Cursor, Replit, Claude, etc.). Even if you can’t do everything, if you can replace just part of the functionality with AI, the outsourcing costs will be reduced to less than half.

3. Cultivate One Person in-house Who Can Interact with AI

There’s no need to hire a programmer. Have the person who knows the business best learn how to use AI tools. How to write prompts, convey specifications, and check outputs. With just this, they can become a “creator.” If a 15-year-old can do it, there’s no reason your employees can’t.

—

The Real Competition Is Not Whether You Can Use AI

Finally, just one thing.

AI tools can be used by anyone. Both large and small enterprises can use the same ChatGPT for 20 dollars a month. This means that there is no difference in whether you can use AI.

The difference lies in who understands their business better than anyone else and can accurately convey that to AI.

Large companies have fewer people who grasp the overall picture of the business due to their size. Small and medium enterprises have owners who know the field. They can see the faces of customers. They have the business flow in their heads.

This “resolution of the field” will become the greatest weapon for small and medium enterprises in the AI era.

We are entering an era where a 3 million yen system can be created for 50,000 yen. The issue is not the technology. It’s whether you can articulate “what you need.”

Companies that can do that will start to win.