Are You Watching the Bills from AI Agents? — The Structure That Turns 50,000 Yen into 500,000 Yen and Immediate Actions You Can Take

What Was Meant to Be 50,000 Yen Turns into a Bill of 500,000 Yen

Small and medium-sized enterprises that have introduced AI agents are increasingly reporting the same concern.

“I thought I could manage it for around 50,000 yen a month, but the bill exceeded 500,000 yen.”

This is not a special case; it is a structurally occurring issue.

Why? Because AI agents, with just a single instruction from a human, can invoke APIs dozens of times behind the scenes. They query GPT-4o or Claude 3.5, call external tools, verify results, and query again. A single task can consume thousands to tens of thousands of tokens. If this happens hundreds of times a day, costs can balloon exponentially.

The essence of the problem is not “overuse”; it is the lack of visibility into expenses.

The Structure of “Autonomous Operation” of Agents That Generates Costs

Traditional SaaS had fixed monthly fees. Slack was 850 yen per month, Notion was 1,650 yen per month. Whether you used them or not, the cost remained the same. Therefore, budgeting was predictable.

AI agents are different. They are a bundle of pay-per-use charges.

Let’s take a closer look. OpenAI’s GPT-4o charges $2.50 for every million tokens of input and $10 for every million tokens of output (as of 2024). At first glance, it seems inexpensive. However, when the agent operates autonomously, the situation changes.

For example, a customer support agent might perform the following actions to handle a single inquiry:

1. Analyze the user’s question (500 tokens)
2. Search the knowledge base and retrieve three candidates (1,000 tokens each)
3. Compare and integrate the candidates to generate a response (2,000 tokens)
4. Self-check the quality of the response (1,500 tokens)
5. Regenerate if necessary (2,000 tokens)

This totals about 7,000 tokens per inquiry. If there are 3,000 inquiries in a month, that amounts to 21 million tokens. Calculating at GPT-4o’s output rate, this alone would cost $210 (around 30,000 yen) per month. When you add multiple agents for tasks like code reviews, data analysis, and email creation, surpassing 500,000 yen a month becomes easy.

Even more troublesome is the “loop”. If an agent is designed to automatically retry upon detecting an error, a single failed task can continue to hit the API 10 or 20 times. While no one is watching in the middle of the night, tokens keep burning. It’s like a water pipe with the meter running continuously.

A study by a research team at Carnegie Mellon University analyzed 465 AI agent-generated pull requests and found that agents with higher code quality had token costs that were 3.5% to 6.1% lower. In other words, agents that operate carelessly consume more money. Quality and cost are two sides of the same coin.

Why Are Small and Medium-Sized Enterprises at Greater Risk?

In large companies, dedicated MLOps engineers can monitor dashboards, set cost alerts, and optimize model switches.

Small and medium-sized enterprises lack such personnel.

A company president might introduce an agent because it seems convenient, delegate the setup to an external freelancer, and have on-site staff manage it somewhat haphazardly. Cost management? They only realize the issue when they see the credit card bill at the end of the month.

This structure is dangerous. The operational costs of AI agents are far greater than the initial implementation costs. If you use pay-per-use AI agents with the mindset accustomed to fixed monthly SaaS fees, you will definitely get burned.

The Option of “Keydris” — Managing the Agent’s Wallet

A straightforward solution to this problem is the expense management tool “Keydris”.

What Keydris does is simple and clear:

• Real-time Monitoring: Visualizes the token consumption and costs of each agent in real-time
• Budget Cap: Sets a monthly limit and automatically restricts agent activity if it approaches that limit
• Alert Notifications: Sends warnings when 70% and 90% of the budget is reached
• Cost Analysis: Breaks down which agents are spending how much on which tasks

In short, it’s an AI agent version of a household budget app.

If you set a budget of 50,000 yen, it will stop at that amount. You can easily see which agent is a money pit. It can also detect runaway loops in the middle of the night.

This may sound mundane, but it is a tool that solves a life-or-death issue for small and medium-sized enterprises. An unexpected bill of 500,000 yen can wipe out profits for a company with monthly sales of 5 million yen.

However, Considerations Before Implementing the Tool

Will everything be resolved by implementing Keydris? Not necessarily.

First, there are three things to confirm within your company.

1. Is that agent really necessary?

The belief that “introducing AI agents will streamline operations” is often an illusion. Out of 3,000 inquiries a month, how many should be handled by the agent? Throwing all of them at the agent versus assigning only 500 routine inquiries can result in a sixfold difference in costs.

2. Is the model selection appropriate?

It is not necessary to use GPT-4o for every task. For routine classification tasks, GPT-4o-mini is sufficient, and costs can be less than one-tenth. It costs $0.15 per million tokens of input. Simply questioning “Is the highest-performing model necessary for this task?” can often cut monthly costs by more than half.

3. Have you set a retry limit?

Is the agent designed to retry indefinitely in case of errors? It should stop after three failures and notify a human. This alone can prevent most of the runaway costs.

After organizing these three points, you can implement a monitoring tool like Keydris. The order of operations is important.

What Lies Beyond “Seeing Costs”

Let’s think a step further.

When the costs of AI agents are fully visible, what happens?

You will be able to determine with numbers whether it is worth having AI perform this task.

For example, if you are spending 80,000 yen a month on an agent for data entry tasks, but hiring a part-time worker would only cost 60,000 yen, then there’s no point in using AI. Conversely, if a 30,000 yen agent is handling an analysis task that would cost 300,000 yen for a human, then you should invest more.

By making costs visible, you can make decisions about “AI vs. human” based on numbers rather than intuition. This is the true value for small and medium-sized enterprises.

Large companies have the resources to “AI-ify everything for now.” Small and medium-sized enterprises do not. Therefore, the accuracy of the judgment on where to use AI and where to use humans becomes a competitive advantage. The foundation for that judgment is the visibility of costs.

What to Do This Week

I’ve written a lot, but ultimately there is one thing to do.

Accurately assess the bill from the AI agents you used last month.

OpenAI’s Usage dashboard, Anthropic’s Billing screen, AWS’s Bedrock billing. Open them all and calculate the total amount.

If you find that number to be “higher than expected,” it’s worth considering a tool like Keydris. If it’s “as expected,” you can continue with your current operations.

AI agents can be a weapon for small and medium-sized enterprises to compete with large companies. However, if you collapse under the maintenance costs of that weapon, it defeats the purpose.

Make the invisible costs visible. It may seem mundane, but this will become the basic operation of management in the era of AI agents.