“Are you still continuing with ‘I can’t place orders without that person’?”

A wholesale business with 30 employees. There is only one person in charge of ordering. When that person is absent, inventory goes awry. If they leave, operations come to a halt. This structure is familiar to many small and medium-sized enterprises in rural areas.

This week, all the components needed to eliminate this “personalization” have come together. A benchmark paper on inventory management using AI agents, a large-scale evaluation of MCP, and an implementation example running on 325 lines of Python. Although these news items appeared separately, combining them reveals a new landscape. Automated inventory ordering has become a reality within the budget constraints of small and medium-sized enterprises.

To put it simply, a system that can automatically cycle through the loop of “demand forecasting → order decision → order execution” can now be assembled for a running cost of less than 50,000 yen per month. The era of purchasing a 3 million yen packaged system is over.

—

How Close Have AI Agents Come to Being “Veterans in Inventory Management”?

First, let’s establish the facts. The paper titled “AI Agents for Inventory Control” published on arXiv is intriguing. It benchmarks over 1,000 inventory scenarios (demand fluctuation patterns, variations in lead times, seasonality, etc.) and compares AI agents based on large language models (LLMs) with traditional operations research (OR) algorithms.

There are three key points:

1. In stable demand patterns, OR algorithms still perform strongly. This is obvious. Areas where mathematical optimization excels should be left to mathematical optimization.
2. In situations where demand fluctuates irregularly, LLM agents have outperformed OR in some cases. In scenarios where “last month’s patterns don’t apply,” the ability to read context and change decisions is the strength of LLMs.
3. A hybrid structure of OR and LLM produced the most stable results. Rule-based during normal times, with LLM intervention during anomalies. This design is the closest to practical application.

Consider this: The reason inventory management becomes personalized in rural wholesale businesses is that “it can usually be managed by patterns, but experienced judgment is needed during seasonal changes or irregular large orders.” The fact that LLMs can now handle this “exception processing” is the significance of this paper.

The “gut feeling” that existed in the minds of veteran staff was actually a form of “comparison with past similar patterns.” LLMs excel at exactly that.

—

MCP Has Changed the Cost of “Connections”

Next is MCP (Model Context Protocol). This is a standard protocol proposed by Anthropic that allows AI agents to call external tools. It may sound complicated in katakana, but what it does is simple. It standardizes how AI can perform external operations like “reading Google Sheets,” “calling APIs for ordering systems,” and “sending notifications via Slack” in a unified manner.

The arXiv paper titled “MCP-Atlas” reports large-scale benchmarks using actual MCP server groups. There are some noteworthy numbers:

• Number of MCP servers evaluated: More than 50
• Success rate of AI agents in “correctly selecting and calling tools”: Varies by model, but over 80% for top models
• The success rate for “multi-step” tasks across multiple tools did not drop significantly.

What does this mean? Traditionally, connecting AI with business systems required expensive development known as “integration.” If you relied on a system integrator, it could cost hundreds of thousands of yen. With MCP standardizing this process, the cost of connections has dramatically decreased.

For small and medium-sized enterprises, this is critically important. No matter how much the performance of AI models improves, it means nothing if they cannot connect to their own inventory database or ordering system. MCP solves this “connection issue” for a server cost of just a few thousand yen per month.

—

What Can Be Achieved with 325 Lines of Python?

So, how exactly do you set this up? This is the most crucial part.

Imagine a daily necessities wholesale business with 30 employees. The number of SKUs (stock-keeping units) is 500, and there are 10 suppliers. Currently, one person in charge of ordering creates purchase orders every morning while looking at Excel. This amounts to about 300,000 to 350,000 yen in labor costs per month.

Configuration

Processing Flow (Contents of the 325 Lines of Python)

1. Automatically starts at 6 AM every morning. Retrieves inventory data from the spreadsheet (via MCP).
2. Executes demand forecasting for each SKU. Generates forecasts for the next 14 days based on the shipping data from the past 90 days.
3. Extracts SKUs that fall below safety stock levels. Up to this point, it is rule-based (OR algorithm).
4. Delegates judgment to LLM for outliers. For SKUs that deviate from patterns, such as “200% increase in shipments compared to last month,” it provides context (seasonality, past similar events, supplier information) to the Claude API to suggest order quantities.
5. Posts the order draft to Slack. The person in charge just needs to press the “approve” button.
6. Once approved, automatically sends the order email. The email sending tool connected via MCP executes this.

These six steps fit within 325 lines. Why? Because MCP abstracts tool connections, significantly reducing the boilerplate code for API calls.

Cost Comparison

From 300,000 yen a month to 15,000 yen a month. This results in an annual savings of about 3.4 million yen. The initial investment can be recouped within six months.

Moreover, this is not a discussion about “firing the order manager.” The order manager’s routine that took two hours every morning will be reduced to just 15 minutes of approval work. With the freed-up time, they can focus on tasks that only humans can do, such as negotiating prices with suppliers or developing new sourcing channels.

—

A Structure Where Small and Medium-Sized Enterprises Can Win

This is the most important point I want to convey.

Large corporations pay tens of millions of yen annually for inventory management modules from SAP or Oracle. Customization costs even more millions. Implementation takes one to two years. Then comes the absurdity of “aligning business with the system.”

Small and medium-sized enterprises are different. With 500 SKUs and 10 suppliers, 325 lines of Python are sufficient. They can change the code to fit their business. Thanks to MCP, it’s also easy to swap out the tools later. If they want to switch from Google Sheets to kintone, they just need to replace the MCP server.

Being small becomes a weapon. Decision-making is swift. If something doesn’t work, they can change it the following week. While large corporations are going through bureaucratic processes, small and medium-sized enterprises can execute three rounds of trial and error.

This is not limited to inventory management. Invoice processing, estimate creation, daily report aggregation—any business where “only that person knows how to do it” can be systematized in the same way. With the emergence of MCP as a “standard for connections,” the disparate components of automation that were previously scattered can now be assembled under a common design philosophy.

—

So, What Should Be Done?

You can start in three steps.

1. Conduct an Inventory (1 Day)
List up the “personalized tasks” in your company. The criterion is, “Does it stop if that person takes a week off?” If it does, that is the top priority for automation.

2. Assemble with Minimum Configuration (1-2 Weeks)
Don’t try to do everything at once with all SKUs. Start with the top 50 best-selling items and set up the flow of demand forecasting → order draft generation → Slack notification. If you can’t write Python, hire a local freelance engineer for 300,000 to 500,000 yen. Thanks to MCP, development that used to cost 1 million yen is now less than half that.

3. Run for Two Weeks and Analyze the Numbers (2 Weeks)
Only track three metrics: forecasting accuracy, stockout rate, and order workload. If the numbers improve, expand to all items. If not, adjust the forecasting logic. Since it’s just changing the code, it won’t incur costs like “customization fees of 1 million yen” as with packaged software.

You can obtain “automated ordering” in as little as one month.

—

Summary: The Components Are Ready. Now, Just Assemble Them.

When you line up the three news items that came out this week, it looks like this:

• AI Agent Paper → It has been demonstrated that inventory management decisions can be entrusted to AI in 1,000 scenarios.
• MCP-Atlas Paper → It has been confirmed that the standard protocol for connecting AI with business tools has reached a practical level.
• 325 Lines of Python Implementation → It has been shown that the combination of the above two can operate within the budget constraints of small and medium-sized enterprises.

All the components needed to eliminate personalization have come together.

The question is whether this will end with a simple acknowledgment of “that’s interesting” or if you will start assembling it next week. For small and medium-sized enterprises in rural areas, the democratization of technology is not a “future that will come someday.” It is already here. It’s just a matter of whether you reach out for it.