AI has taken control of the mouse. Do you understand what this means?

Anthropic has released an API called “computer use.” AI can move the mouse, type on the keyboard, and click buttons on the screen. Even without a program API, it can look at the screen and operate just like a human.

This is a significant event for small and medium-sized enterprises in rural areas.

Why? Because I have heard the phrase, “Our core system is too old to integrate with APIs,” hundreds of times. Even when trying to implement RPA, it costs several million yen annually. In the end, a veteran office worker is handling everything manually. If that person leaves, the business comes to a halt. This is the reality for small and medium-sized enterprises across Japan.

This reality has the potential to change drastically.

The Structure Where the Work of ‘Excel Macro Craftsmen’ Becomes 30,000 Yen a Month

Let’s discuss specifics.

In a small manufacturing company (with 30 employees), there is a task where they receive order data via email, manually input it into their core system, transcribe it into an Excel management sheet, calculate delivery dates, and create a reply email. This task takes an average of 2 hours a day, totaling about 40 to 60 hours a month, handled by one veteran office worker.

Calculating at an hourly wage of 1,500 yen, that amounts to 60,000 to 90,000 yen a month. However, the real cost is not just that. If that person takes a day off, the work stops. If they leave, it takes three months for someone else to take over. Recruitment costs range from 500,000 to 1,000,000 yen. The real danger of personalization is the “invisible risk cost.”

Using Anthropic’s computer use, AI can perform this entire series of tasks through screen operation. It doesn’t matter if there is no API for the core system. It can look at the screen just like a human, click on input fields, and enter data.

What about API usage fees? Currently, estimating based on the cost of Claude 3.5 Sonnet, it would be around several dozen to several hundred yen per screen operation task. Even if you run 500 routine operations a month, the API costs would amount to about 20,000 to 30,000 yen a month. Even if you include infrastructure costs and maintenance for prompt adjustments, it would still be around 30,000 to 50,000 yen a month.

The labor cost of 90,000 yen a month could be reduced to 30,000 yen. That’s a savings of 720,000 yen annually. But I repeat, the true value is not just in the numbers. “The business does not stop even if that person leaves”—can you put a price on this peace of mind?

Why ‘Screen Operation AI’ is a Game-Changing Weapon for Small and Medium-Sized Enterprises

Here, I want to discuss the structural aspects.

Until now, there have been three main options for business automation:

The key point is that it can automate even systems without APIs.

Large companies use Salesforce or SAP, integrate via APIs, and have dedicated IT departments for maintenance. Small and medium-sized enterprises do not have that luxury. Sales management software from 20 years ago is still in use, and no one touches the source code. And that’s fine. Screen operation AI can automate those “untouchable systems” as they are.

This is a weapon that works precisely because it’s aimed at small and medium-sized enterprises. Large companies have already automated through API integration. It is precisely because small and medium-sized enterprises cannot integrate with APIs that the benefits of screen operation AI are overwhelmingly significant. The impact of cost reduction is incomparable to that of large companies.

Multi-Agent Systems—From ‘One AI’ to ‘AI Teams’

Let’s take it a step further.

If a screen operation AI can automate a single task, multiple AI agents working together can automate the “entire business flow.” Recent research (such as EntCollabBench) has shown that multi-agent systems with specialized roles achieve significantly higher accuracy than single agents.

I want you to visualize this concretely.

• Order Agent: Extracts order information from emails and inputs it into the core system
• Inventory Check Agent: Opens the inventory management screen and checks stock levels
• Delivery Calculation Agent: Calculates delivery dates based on production schedules and inventory
• Reply Agent: Automatically sends delivery response emails to customers
• Approval Agent: If there are anomalies (such as large orders), it checks with a human

The “flow of judgment” that used to exist in the mind of a single veteran office worker is externalized as a collaboration of agents. Personalization structurally disappears.

Moreover, what each agent “does” is defined by prompts (natural language instructions). It’s not programming; it can be written in Japanese. This means that the person who knows the business best can design it themselves.

The era of paying several million yen to SIers for requirement definitions is coming to an end.

Risks and Limitations That Should Be Viewed Calmly

However, this does not mean that everything will be replaced immediately. I will write honestly.

Accuracy Issues. Currently, the computer use may make mistakes with complex screen transitions or dynamically changing UI elements. Anthropic’s official documentation clearly states that it is still in the “experimental stage (beta).” If you plan to implement it in real business, it is essential to include human checkpoints in the design.

Security Issues. Having AI operate the screen means giving the AI access to login information and system access rights. Small and medium-sized enterprises often have vague security policies. This should be organized before implementation.

Selecting ‘What to Automate.’ Trying to automate everything will lead to failure. Start with tasks that are “done daily, have established procedures, and are troublesome if mistakes occur.” If there is one routine task taking more than 10 hours a month, that’s a good starting point for experimentation.

So, What Should We Do?

I will say just three things.

1. Create a ‘Personalized Task List’ Immediately.
Identify all tasks in the company that can only be done by “that person.” These are candidates for introducing screen operation AI.

2. Experiment on a Small Scale.
Anthropic’s computer use is available as an API. Start by testing it on one routine task. The initial investment is 50,000 to 200,000 yen. Even if it fails, it’s not a fatal amount.

3. Aim for ‘Systematization.’
Cost reduction is a result, not a goal. The goal is to create a state where “anyone can achieve the same results.” That is the survival strategy for small and medium-sized enterprises.

This is not ‘Automation’ but ‘Structural Change’

Finally, I want to write the most important point.

What Anthropic’s computer use has shown is not just a technical discussion about “automating systems without APIs.” It’s a structural discussion about “when the cost of automation dramatically decreases, the source of a company’s competitive advantage changes.”

Until now, the strength of small and medium-sized enterprises has been “the experience and intuition of veterans.” However, that also means the risk of “if that person is gone, it’s over.”

The combination of screen operation AI and multi-agent systems transforms “experience and intuition” into a “reproducible system.” For just 30,000 to 50,000 yen a month.

The era is coming when small and medium-sized enterprises can do what large companies have done for tens of millions of yen for just a few tens of thousands of yen. If you have a reason not to use this, I would like to know.

Just try one thing first. The answer is on the other side of the screen.