Conclusion

Let’s get straight to the point: the risk of “not using AI” has now surpassed the “runaway risk.”

GPT-5.6 deleted all files. The U.S. Senate has introduced a bill to regulate agent AI. Discussions are intensifying around confining AI agents within containers.

Hearing this might make you think, “AI is indeed scary; it’s still too early for us.”

But hold on a moment.

What’s truly frightening isn’t that AI agents might run amok, but rather that while competitors are fully utilizing AI agents, we are doing nothing.

In this article, we will compare runaway costs and inaction costs on a monthly basis. Based on that, we will specifically design a “minimal safe configuration” that small businesses can implement for under 50,000 yen per month.

—

The GPT-5.6 File Deletion Incident—What Happened and How Much Was Lost?

First, let’s clarify the facts. An AI agent using GPT-5.6 misinterpreted instructions and deleted all files, including those that were not supposed to be worked on. The reported damage amounts to approximately 3 million yen for one company, including recovery efforts, business downtime, and customer relations.

3 million yen. This figure appears fatal for small businesses.

However, concluding that “AI is dangerous” based solely on this number is akin to saying “don’t drive because there are traffic accidents.”

The issue isn’t whether an accident occurred, but whether the system was configured in a way that allowed for accidents to happen.

In this case, the agent was granted full access rights to the file system, and no backups were made. It’s like getting on a highway in a car without brakes or seatbelts.

A runaway incident didn’t occur; the system was left in a configuration that allowed for a runaway.

—

Runaway Costs vs. Inaction Costs—A Monthly Comparison

Now, let’s get to the main topic: comparing the numbers.

Runaway Costs (In the event of an accident using AI)

A 5% accident rate is estimated on the high side. With proper authority management and containerization, this probability can drop below 1%. This means the risk is at the level of a few thousand yen per month.

Inaction Costs (If operations remain unchanged without using AI)

In small businesses, tasks such as creating estimates, responding to inquiries, data entry, and report generation can easily free up 500,000 yen worth of human resources by implementing AI agents. Conversely, not using AI results in a monthly loss of over 500,000 yen.

Runaway costs: 10,000 to 40,000 yen. Inaction costs: 500,000 to 1,300,000 yen.

There’s a difference of over ten times. It’s clear which is more frightening.

—

The U.S. Senate’s Agent AI Regulation Bill—What Changes for Small Businesses?

The key points of the agent AI regulation bill introduced in the U.S. Senate are threefold:

1. Mandatory logging of agent actions—keep a record of what was done.
2. Mandatory human approval flow—important operations must be confirmed by a human.
3. Clarification of responsibility in case of damages—determine who will take responsibility.

This is actually a tailwind for small businesses.

Why? Large corporations have complex system integrations, making the costs of implementing logging and approval flows high. In contrast, small businesses have simpler systems. The narrower the operational scope of the agent, the lower the costs of regulatory compliance.

The stricter the regulations become, the more advantageous a “small and simple configuration” becomes. This is a structure where small businesses can win.

If the bill passes, similar regulations are likely to emerge in Japan as well. By creating a “regulatory-compliant configuration” in advance, the costs of scrambling later will be zero.

—

The Insurance Option—Can You Cover a 3 Million Yen Accident for 50,000 Yen a Month?

Cyber insurance related to AI agents is starting to emerge. There are several plans available for 30,000 to 50,000 yen per month that cover damages from data breaches and operational errors up to 5 million yen.

With a 50,000 yen premium, you can cover a 3 million yen runaway accident. Against an annual premium of 600,000 yen, one accident can be compensated with 3 million yen. Even if an accident occurs once every two years, it’s a worthwhile investment.

However, insurance is only a last resort. Just because you have insurance doesn’t mean you can forgo a safe configuration. Insurance companies will not provide coverage to companies that do not implement basic security measures.

Security (technical measures) × Regulation (compliance with rules) × Insurance (financial backup). Only when this triangle is aligned can you safely let AI agents “run wild.”

—

The “Minimal Safe Configuration” for Small Businesses—A Practical Setup for Under 50,000 Yen a Month

Now that the reasoning is clear, what should you actually do?

Here’s a minimal safe configuration that can be set up for under 50,000 yen a month.

1. Limit Permissions (Cost: 0 yen)

Set the default permission for AI agents to “read-only.” If writing or deletion is necessary, specify only one target folder.

The incident with GPT-5.6 happened because full access was granted. By simply limiting permissions, the potential damage from a runaway can be dramatically reduced. Since this is just a configuration change, the cost is zero.

2. Isolate with Containers (Cost: 0 to 5,000 yen/month)

Isolate the execution environment of the AI agent using Docker. No matter what the agent does, it won’t affect the outside of the container.

Docker itself is free. Even when running in the cloud, it can be done for a few thousand yen a month. This alone eliminates the risk of the host system being affected.

3. Keep Action Logs (Cost: 0 to 3,000 yen/month)

Log all operations of the agent. Record everything it executes and when. Free logging management tools (like Grafana Loki) can adequately handle this.

This is also what the U.S. Senate’s regulatory bill is asking for. By addressing this in advance, you won’t have to scramble when regulations come.

4. Insert Human Approval for Important Operations (Cost: 0 yen)

For operations like “file deletion,” “external transmission,” and “bulk data updates,” send a confirmation to a human via Slack or email before proceeding. The agent will wait for approval.

Implementation can be done in about 30 minutes using no-code tools like n8n or Make (formerly Integromat).

5. Daily Backups (Cost: 500 to 2,000 yen/month)

Daily backups of the data accessed by the agent should be performed. Google Drive or S3 is sufficient. This can be managed for a few hundred yen a month.

Even if the agent deletes everything, you can revert to the state from yesterday. A 3 million yen loss can be reduced to just a few hours of recovery work.

Total Costs

Under 10,000 yen per month. Even adding the 50,000 yen for cyber insurance, the total is under 60,000 yen per month.

Compared to inaction costs of 500,000 to 1,300,000 yen, this is a negligible amount.

—

“First, Create One That Acts on Its Own”

Finally, I propose a concrete first step.

You don’t need to implement AI agents in all operations at once. Start with one task that can be “completed on its own.”

For example, generating daily sales reports, initial classification of inquiry emails, or data entry for invoices.

Delegate one of these routine tasks to an AI agent. Surround it with the minimal safe configuration mentioned above and let it run for a week.

When you arrive at the office in the morning, the report is already done. Emails are classified. Data is entered.

That experience of having something “completed on its own” will change your next decision.

It’s not about avoiding use because you’re afraid of a runaway. It’s about fully utilizing it in a configuration that prevents runaways. That’s the AI strategy for small businesses that generates over 500,000 yen in value for just 60,000 yen a month.

Which is higher, runaway costs or inaction costs? The answer is already clear.