Conclusion First: Many Companies Are Correct to Say, ‘We Stopped Using AI Agents’

AI agents are a hot topic. They autonomously determine tasks, call tools, and automate complex operations—sounds great. However, more companies are now deliberately discarding AI agents and reverting to simpler systems.

Why? The answer is simple: They are not cost-effective.

This is not just a story about large corporations. In fact, it is small and medium-sized enterprises in regional areas that face fatal consequences if they make the wrong decision. For companies with monthly sales of a few million yen, the operational costs of AI can balloon to 300,000 or 500,000 yen per month. And has this dramatically changed their operations? Not really; they feel like things are ‘somehow automated.’ Companies that have realized this structure are quietly withdrawing.

The Invisible Cost of ‘Tool Usage Tax’

To put it simply, the mechanism of AI agents works like this: LLMs (large language models) determine ‘what to do next’ and automatically execute external tools—such as database searches, API calls, and file operations. The AI acts independently without human instructions.

The problem is that this ‘acting independently’ incurs costs.

Every time the AI calls a tool, additional tokens are consumed. There are prompts for decision-making, prompts to interpret tool results, and prompts to decide the next action. Completing a single task can require 5 to 10 API calls to the LLM.

This is referred to as ‘tool usage tax‘.

One study reported that the overhead of tool calls in agent-based systems accounts for 30-40% of total operational costs. In other words, out of a monthly AI operational cost of 1 million yen, 300,000 to 400,000 yen is spent on ‘AI thinking to think.’ Only the remaining 600,000 to 700,000 yen is actually used for processing business tasks.

What’s more troublesome is that this cost increases the more you use it. As tasks become more complex, the number of tool calls rises, and costs balloon exponentially rather than linearly. It’s not uncommon for a system that started at 100,000 yen per month to reach 500,000 yen after six months.

Case Study: Replacing a 3 Million Yen AI Agent with a 50,000 Yen Script

Let’s discuss a specific example.

A local manufacturing company (with about 50 employees) introduced an AI agent to automate its order processing. The plan was to automate the flow of reading incoming order emails, checking inventory, responding with delivery dates, and registering in the internal system.

The vendor’s initial estimate was 800,000 yen per month, claiming, ‘We will save the equivalent of two employees’ salaries.’

However, once they started actual operations, this is what happened:

• The AI could not recognize different formats of order forms, resulting in an error rate of 15%.
• Humans still had to be involved in error handling, so labor costs did not decrease.
• In cases requiring complex judgments, the AI would ‘get stuck,’ leading to an average latency of 8 seconds.
• Increased tool calls caused monthly costs to balloon from the initial 800,000 yen to 3 million yen.

In the end, what did this company do?

They standardized the order form into three patterns and processed them with a simple script using pattern matching. No AI was used—just if statements and CSV conversions. Latency dropped to 0.2 seconds, and the error rate fell below 2%. Monthly costs were reduced to just 50,000 yen for cloud server fees.

What went from 3 million yen is now just 50,000 yen—a reduction to 1/60th of the original cost.

Moreover, processing speed improved by 40 times, and the error rate decreased to 1/7th. In every metric, the ‘simple system’ triumphed.

Question: Is an AI Agent Really Necessary for That Task?

At this point, I want you to pause and think.

AI agents truly excel in tasks where ‘rules cannot be predetermined.’ Situations that change every time, vague criteria, and infinite patterns—these are areas where AI’s flexibility shines.

However, most operations in small and medium-sized enterprises are the opposite. Patterns are finite, and rules can be established. Most of the time, it’s simply that they haven’t been established yet.

Order processing, invoice issuance, inventory checks, daily report aggregation, attendance management—these tasks do not require ‘AI thinking.’ They can be completed by setting rules and writing scripts.

In other words, for many small and medium-sized enterprises, AI agents are ‘overkill.’

The ‘Just Right’ Option of Small Models

That said, I’m not saying to avoid using AI altogether.

What deserves attention is the rapid evolution of small open-weight models. Lightweight models with parameter counts between 1B and 8B are beginning to show remarkable performance in specific tasks.

Recent benchmark evaluations compared 16 models ranging from 0.27B to 32B. The results showed that open models in the 8B class achieved 85-90% of the scores of GPT-4o class in structured tool usage tasks. Operational costs can be reduced by about 80% compared to using the GPT-4o API.

This has significant implications.

Operations that were running on a GPT-4o-based agent for 500,000 yen per month could potentially be reduced to 50,000 to 100,000 yen by using a small model running on an in-house server with a simple pipeline. Moreover, data does not leave the premises. This is a benefit that cannot be ignored for local manufacturing and construction industries.

Framework for the Decision to ‘Quit’

Let’s specifically outline the criteria for deciding whether to continue or stop using AI agents.

Signs You Should Stop:

• The task can be classified into fewer than 10 patterns.
• Humans are still involved in handling errors.
• Monthly costs have exceeded twice the initial estimate.
• It’s not that ‘the operation cannot function without AI,’ but rather ‘even with AI, humans are still needed.’
• Response time is a bottleneck in operations.

Signs You Should Continue:

• The input data format varies each time, making it impossible to establish rules in advance.
• Natural language understanding is essential for decision-making (e.g., complaint classification, contract review).
• Automation has clearly reduced labor costs (proven by numbers).

The key point is to make decisions based on ‘numbers’ rather than ‘feelings.’ By measuring monthly costs, error rates, processing speeds, and human labor hours, the right time to stop will naturally become apparent.

The ‘Reversal Structure’ Possible for Small and Medium-Sized Enterprises

Large companies find it difficult to stop using AI agents. They invest hundreds of millions in implementation projects, and with internal politics involved, the cost of making the decision to ‘quit’ becomes even greater. This is the curse of sunk costs.

On the other hand, small and medium-sized enterprises are different.

If the president says, ‘Let’s stop this,’ they can switch over by next week. The speed of decision-making is overwhelmingly fast. While large corporations continue to pay hundreds of thousands of yen per month for excessive AI systems, small and medium-sized enterprises can achieve equal or better results with simpler systems.

This is not a joke; it’s a structural advantage.

The evolution of AI has reduced the costs of ‘advanced automation.’ But at the same time, the costs of ‘simple automation’ have decreased even more. By combining no-code tools, lightweight scripts, and small models, sufficient automation can be achieved for just a few tens of thousands of yen per month.

So, What Should You Do?

I have three things I want you to do.

1. Identify all current AI-related costs.
API usage fees, server costs, payments to vendors, and the human labor hours spent on handling AI errors. Add them all up. How much is it per month?

2. Consider whether that task can be written with if statements.
‘If A is B, then do C’—if it can be described at this level, you don’t need an AI agent. Identify the patterns and translate them into scripts. That’s all it takes.

3. Experiment with one business process first.
You don’t have to change everything at once. Choose the AI operation that incurs the highest costs and try replacing it with a simple system. Run it for a month and compare costs and results. The numbers will provide the answer.

AI agents are a remarkable technology. However, they are not necessary for every operation. Especially for small and medium-sized enterprises, the decision to ‘not use’ AI may be the greatest utilization of AI.

Technology is a means. The goal is to generate profit. By returning to that fundamental point, the answer becomes clear.