ERPに年間300万円払っている中小企業に聞きたい。それ、まだ必要ですか？

The cost of using LLM APIs has dropped to one-hundredth of its previous price. The performance level of GPT-3.5, which cost 30,000 yen per month at the beginning of 2023, is now available for just 300 yen a month by late 2024. In terms of token pricing, even GPT-4 Turbo has seen a reduction to about one-tenth in just a year. If you run open-source models like Llama 3 or Mistral locally, the inference cost approaches nearly zero.

However, few business owners of SMEs are considering this “collapse of cost” in the context of core systems.

Yet, structurally speaking, the answer is clear. ERP is quietly beginning to die.

ERPの正体は「高い翻訳機」だった

What exactly is ERP? It is a system that centralizes the management of disparate business data such as inventory management, order processing, sales management, and financial accounting. When SMEs implement systems like SAP Business One or OBIC’s OBC series, the initial costs range from 2 to 5 million yen, with annual maintenance fees of 500,000 to 1.5 million yen. Over five years, this totals between 7 to 15 million yen.

For a manufacturing company with 30 employees, this is a heavy burden. It’s burdensome, yet they have paid it out of necessity. Why?

Because human language and system language are different.

When you ask, “Show me the clients with poor gross profit margins from last month,” traditional systems do not respond. You need someone who can write SQL or knows where to click on a dedicated screen. ERP was a “high-cost translation machine” that translated business language into system language.

LLMs eliminate this translation cost.

月300円のLLMが壊す「3つの前提」

There are three assumptions under which ERP has thrived. The collapse of LLM costs simultaneously destroys these three assumptions.

前提1：データの一元管理には専用システムが必要

Traditionally, invoices were managed in Excel, inventory in dedicated software, and customer information in spreadsheets—ERP was necessary to integrate these.

What about now? LLM agents can read across CSVs, PDFs, Google spreadsheets, and kintone, providing answers in natural language. If you ask, “List the parts whose purchase price increased by more than 10% in September,” it pulls data from multiple sources and presents it in a table.

The integration is now done by AI, not by a system. It doesn’t matter where the data is stored. At this point, half of the justification for ERP disappears.

前提2：業務ロジックはコードで書くしかない

“Automatically issue invoices for month-end closing with payment at the end of the following month” or “Notify via Slack when inventory falls below safety stock”—these rules previously had to be implemented as programs. Each customization cost between 300,000 to 1 million yen. For SMEs, even a slight change in rules could lead to significant outsourcing costs.

With LLM agents, you can describe rules in natural language. If you write, “If there are unpaid invoices at month-end, send a reminder email to the responsible person,” it will execute. This is not just no-code; it’s “Japanese code.” A world where customization costs drop from 1 million yen to virtually zero is on the horizon.

前提3：システム運用には専門人材が必要

ERP maintenance and operation required in-house IT staff or external vendors. Hiring a system administrator in an SME could cost 5 million yen annually. Outsourcing could range from 100,000 to 300,000 yen per month.

With LLM-based systems, you can ask, “What happened?” when an error occurs. You can also issue configuration changes in natural language. The operational knowledge that was previously dependent on specific individuals is now absorbed by AI.

In other words, much of the value that ERP provided can now be replaced by LLMs costing between 300 yen and a few thousand yen a month.

ローカルLLMという選択肢——データを外に出さずに済む

“But we can’t send our data to OpenAI.”

This concern is valid. Client information, cost data, employee information—core data for SMEs is a bundle of confidential information.

This is where local LLMs come into play. A model like Llama 3 8B can run on a laptop with 16GB of memory. Using Ollama, setup takes just 10 minutes. The inference cost is only the electricity for the PC, amounting to just a few dozen yen a month.

While its performance may lag behind GPT-4, it is more than sufficient for routine tasks such as “aggregating billing data to create monthly reports” or “extracting part numbers and quantities from order PDFs.” For heavier tasks requiring significant judgment, you can offload them to cloud-based GPT-4, while routine processing is handled locally. This routing—designing the system to allocate models based on task weight—provides a practical solution for SMEs to utilize AI.

エッジ推論のトレードオフを正直に書く

Local LLMs are not without their limitations. There are three trade-offs.

1. There is a ceiling on accuracy. With an 8B parameter model, complex financial analysis or multi-step reasoning can be challenging. There is a risk of producing incorrect aggregations. A flow where important figures are always checked by a human is necessary.

2. Response speed heavily depends on hardware. On PCs without GPUs, generating long texts can take over 30 seconds. They are not suitable for tasks that require real-time responses. Simply adding a used RTX 3060 (costing 30,000 to 40,000 yen) can increase perceived speed by five times, but whether this investment decision can be made is another matter.

3. Someone is needed to create operational frameworks. While the tools have become easier, designing “which model to apply to which task” and “how to design prompts” requires human input initially. Whether to rely on external help or train someone internally is a decision to be made. While a consultant at 50,000 yen a month can suffice, it won’t be zero.

With these considerations in mind, we need to confront the question: “Will you still continue to pay 3 million yen annually for ERP?”

具体的な移行シナリオ：従業員20人の卸売業の場合

Let’s consider actual numbers.

Current situation (using ERP):

• ERP license + maintenance: 1.8 million yen annually
• Customization outsourcing: 600,000 yen annually
• Labor cost of system administrator (part-time): equivalent to 1 million yen annually
• Total: 3.4 million yen annually

If transitioning to LLM-based systems:

• Cloud LLM API (like GPT-4): 3,000 to 5,000 yen monthly → 60,000 yen annually
• Local LLM PC (only in the first year): 150,000 yen
• Data integration tools (like n8n, Make): 5,000 yen monthly → 60,000 yen annually
• External consultant (monthly design review): 50,000 yen → 600,000 yen annually
• Total: 870,000 yen in the first year, 720,000 yen from the second year onward

This results in a cost reduction of over 2.5 million yen annually. Over five years, this amounts to more than 12 million yen. For a company with 20 employees, this difference is significant enough to change management decisions.

ERPベンダーはどう動くか

Major ERP vendors are not sitting idly by. SAP has integrated an AI assistant called Joule, and Oracle has added generative AI features. However, these approaches merely add AI on top of existing ERP systems. The licensing structure remains unchanged. The annual fee model costing several million yen continues, with AI treated as an additional option.

For SMEs, this does not provide a fundamental solution. It merely adds high-cost features on top of an already expensive foundation.

On the other hand, a “combinatorial approach” that connects LLMs to lightweight platforms like kintone or Notion is emerging as a realistic solution for SMEs. The base platform costs only a few thousand yen a month, while the intelligence costs a few hundred yen a month. We are entering an era where core ERP functions can be covered for less than 10,000 yen a month.

「次の10年」で起きること

Rather than predictions, let’s discuss what is visible from a structural perspective.

Within 1-2 years: The number of cases where LLM agents automate routine tasks (invoice processing, inventory alerts, monthly reports) will surge. Forward-thinking SMEs will begin to postpone their ERP contract renewals.

Within 3-5 years: Industry-specific LLM agent packages will emerge. Templates that combine prompts and data integration, such as “for manufacturing” or “for wholesale,” will circulate for less than 10,000 yen a month. The number of new ERP implementations will plummet.

Within 5-10 years: The very concept of core systems will dissolve. Data will exist in disparate locations, and AI will integrate and process it across the board. We will enter an era where we “teach AI our business” rather than “implementing systems.”

で、今日から何をすればいいのか

Just three things.

1. Check the ERP contract renewal date. Before the next renewal, experiment with one task that can be replaced by LLMs. Whether it’s automatically sorting invoices or generating monthly sales reports, it doesn’t matter.

2. Install Ollama and test it with your own data. It takes about 30 minutes. Load your sales CSV and ask, “Show me the top 10 sales from last month.” Gaining a sense of how it works is the first step.

3. Inventory whether you are really using the functions of your ERP. Many SMEs only utilize 20-30% of their ERP functions. Visualizing the costs associated with unused features can change decision-making.

The structure of continuing to pay several million yen annually for ERP is quietly but surely being dismantled by the collapse of LLM costs. The question is not “if it will happen” but “when it will happen.”

It’s not a race, but those who are slow will lose.