Conclusion

Let’s get straight to the point: the “fair price” of business systems is about to change dramatically.

A monthly fee of 300,000 yen for SaaS is a common story.

Customer management, inventory management, invoice processing—when small and medium-sized enterprises (SMEs) try to run their operations properly, they often end up spending 300,000 to 500,000 yen a month on 3 to 4 SaaS solutions. That’s 3.6 million to 6 million yen a year. For a company with 20 employees, this is equivalent to the salary of one staff member.

However, we have entered a phase where “building AI in-house can reduce costs to 50,000 yen a month,” which has been the case for the last year or two. The costs of cloud LLM API and some infrastructure have indeed decreased.

But the next wave is going to be even more disruptive.

Business systems that used to cost 3 million yen a month will soon operate for less than 5,000 yen a month.

This may sound exaggerated, but with three technological trends moving simultaneously, this figure is no longer a fantasy. The incremental updates of MCP servers, the scaling of development through synthetic tools, and the practical capabilities of local LLMs are converging. When these three elements align, the very concept of the “price” of business systems will be shattered.

Let’s examine each one in turn.

—

1. Incremental Updates of MCP Servers—Eliminating “API Integration Costs”

A significant portion of the costs associated with business systems actually comes from “integration.”

Connecting accounting software to CRM, linking CRM to inventory management, and integrating inventory management with e-commerce sites. The development and maintenance of these API integrations justify the high monthly fees charged by SaaS vendors. If you try to build it yourself, you need to retest the entire system every time an external API changes, and hiring a development company can cost anywhere from hundreds of thousands to millions of yen per instance. This reliance on SaaS has driven up costs for SMEs.

MCP (Model Context Protocol) is a standard protocol that allows LLMs to call external tools and APIs. In essence, it’s a system that enables “AI to autonomously call APIs.” This concept has been gaining traction since late 2024.

What’s noteworthy is that the mechanism for “incremental updates” has begun to be integrated into this framework. The approach called DeltaMCP allows for updates to be made only to the parts that have changed, without needing to rebuild the entire MCP server when external API specifications change.

What does this mean?

Traditionally, the maintenance costs for API integration have been around 1 million to 2 million yen a year, but this could approach nearly zero. By feeding the MCP server the API specification (like OpenAPI Spec), it can automatically detect changes and update only the relevant tool definitions. The only task for humans is to “place the new API specification.”

The elimination of maintenance costs means that one of the reasons to continue paying monthly fees to SaaS vendors disappears.

—

2. Synthetic Tools—Reducing “Development Efforts” to One-Tenth

The biggest hurdle to creating a business system in-house is the development effort. Requirement definition, design, implementation, testing. If SMEs outsource this, they must be prepared to spend between 3 million and 10 million yen.

The framework of synthetic tools (Synthetic Tools) is set to change this. This approach, represented by SynthTools, uses LLMs to simulate the “business environment itself” and define and validate tasks based on that simulation.

Specifically, it works like this:

1. Environment Generation: When you describe, “We are a manufacturing company with 15 employees, and we have a flow from order to production to shipping to invoicing,” the LLM automatically generates that business environment for simulation.
2. Tool Synthesis: The LLM automatically defines and implements the necessary tools (like order registration, inventory checks, invoice issuance) for that environment.
3. Task Validation: It runs actual business scenarios to automatically test whether the tools function correctly.

Traditionally, just defining requirements would take two weeks, and implementation could take 1 to 2 months, but now it can be done in a matter of hours to days. Moreover, you’re not bogged down by the complexities of actual APIs. You create something that works in a synthetic environment first, and once validated, you switch to the production API. This process of “rapidly creating something that works first” fundamentally changes development costs.

If the development period shrinks from 2 months to 3 days, what happens to the outsourcing cost of 3 million yen? The labor cost for your staff instructing the LLM would only be a few tens of thousands of yen.

—

3. Local LLMs—Making “Running Costs” Just Electricity Fees

“But isn’t there a cost for LLM API usage?”

That was true. Running a GPT-4 class API constantly for business systems could cost tens of thousands to over 100,000 yen a month. This accounted for a significant portion of the “in-house AI at 50,000 yen” breakdown.

However, the performance of local LLMs is rapidly entering the practical realm.

As of 2025, models like Llama 3.1, Qwen 2.5, and Phi-4 are capable of handling business-level tasks with 8B to 70B parameters. Advances in quantization technology mean that even a 70B model can run on consumer-grade GPUs (like RTX 4090 or RTX 5090). Optimizations in inference engines like vLLM and llama.cpp have also brought response times to practical speeds.

Let’s calculate the hardware costs:

• Inference PC with RTX 4090: Approximately 400,000 to 500,000 yen (around 300,000 yen for used models)
• Electricity costs: 2,000 to 3,000 yen per month
• Maintenance: Almost zero (just replace parts if something breaks)

An initial investment of 500,000 yen and running costs of 3,000 yen per month. After 3 years, the total monthly cost would be about 17,000 yen. After 5 years, it would drop to about 11,000 yen per month. API costs would be zero, and data wouldn’t leave the company.

A company that was paying 300,000 yen a month for SaaS could now operate an equivalent or better system for less than 5,000 yen a month.

—

What Happens When These Three Elements Align

Now we get to the main point. Each of these three trends is interesting on its own, but when combined, they create structural changes.

Local LLMs bring running costs close to zero,
Synthetic Tools reduce development costs to less than one-tenth,
Incremental Updates of MCP Servers bring maintenance costs down to nearly zero.

The cost structure of business systems is “development costs + operational costs + maintenance costs.” All three of these will drop dramatically at the same time.

Traditional structure: Development 3 million + Monthly operation 300,000 + Annual maintenance 1 million = 6.6 million in the first year, 4.6 million per year thereafter.

Future structure: Development 50,000 (a few days of in-house staff) + Monthly operation 5,000 yen (electricity costs) + Maintenance almost zero = 110,000 in the first year, 60,000 per year thereafter.

Annual costs drop from 4.6 million to 60,000. That’s about 1/77.

Of course, this is an ideal scenario. In reality, adjusting LLM accuracy and customizing for specific business needs will take time. However, there is no doubt that we are moving towards a two-digit reduction.

—

This Will Create a “Reversal Structure” for SMEs

Consider this:

Large corporations have already signed annual contracts with SaaS vendors, invested tens of millions in customizations, and have dedicated IT departments for maintenance. This “heavy structure” will become a hindrance in a scenario where costs drop dramatically. Contractual obligations, compatibility with existing systems, and internal adjustments—larger companies tend to move slower.

On the other hand, SMEs are more agile.

Is the existing system inadequate? That’s actually an opportunity. The less you have to lose, the easier it is to switch to a new structure. No IT department? We are entering an era where just one person who can instruct an LLM is sufficient.

The structure of “losing because you couldn’t invest in IT” is transforming into a structure where “you can compete on the same level because IT costs are negligible.”

This is similar to how the cloud saved “SMEs that couldn’t afford servers.” However, the impact this time is much greater. The cloud reduced “infrastructure costs.” This time, the “cost of the system itself” is disappearing.

—

So, What Should You Do Starting Today?

There are three things to do.

1. Take stock of your “monthly SaaS expenses.”

How much are you currently paying each month? For what SaaS? Just listing this out will raise the question, “Is this amount really necessary?” Many SMEs continue to pay monthly fees out of inertia.

2. Experiment with one local LLM.

Install Ollama on a used gaming PC and run the 8B model of Llama 3.1. It takes about 30 minutes and costs less than 50,000 yen. Gaining the experience of “Is this really possible locally?” will be the starting point for everything.

3. Set up one MCP server.

Tools like Claude Desktop and Cursor are already MCP-compatible. Try connecting the APIs of the services you use to the LLM via the MCP server. Experiencing “AI autonomously fetching and processing data” will make you realize how wasteful it was to click through the SaaS management screens.

All of these can be started today. The cost is just a few tens of thousands of yen. Even if you fail, it won’t hurt.

—

Finally

“From 3 million to 5,000 yen” is not just a story for today. But it’s also not a story for three years from now.

The essential technologies are already starting to come together. What’s missing is “real-world examples of combining them and making them work.” And those examples will likely be created not by large corporations, but by agile SMEs.

No one will teach you if you just wait. Start experimenting. Start making things work. Whether you’re on the side where costs are collapsing or the side where they’re being collapsed. That turning point is already right in front of you.