Is an AI Invoice of 80,000 Yen Really Justifiable?

Are you really paying attention to the AI service invoices that arrive every month?

“It was 30,000 yen per month at the start, but before I knew it, it had risen to 80,000 yen” — as someone supporting AI utilization in local SMEs, I have been receiving more and more inquiries like this. Initially, the cost is low. Once you start using it, it becomes convenient. However, as six months or a year pass, the number of API calls increases, storage expands, optional features are added, and the bill gradually rises. By the time you realize it, you find yourself in a situation where you want to switch but feel unable to do so.

This is the true nature of AI vendor lock-in.

Large corporations have dedicated teams to negotiate with vendors. But for a manufacturing company with 30 employees or a law office with 10 staff, there is no such luxury. That’s why it’s crucial to organize specific numbers regarding “how much you are currently paying,” “how much you could save by switching,” and “whether switching is even realistic.”

—

The ‘Invisible Costs’ of Lock-In Are Three Times the Monthly Fee

First, let’s break down the costs of lock-in. Many people only look at the monthly fee, but in reality, there are three hidden costs behind it.

1. Data Migration Costs
Transferring data accumulated in a specific vendor’s format to another service requires conversion work. For SMEs without in-house engineers, outsourcing this task can often cost between 200,000 to 500,000 yen.

2. Business Process Reconstruction Costs
If your business flow has been structured around an AI tool, changing the tool can disrupt that flow immediately. During the period of confusion, productivity drops. For instance, if a company with a monthly sales of 10 million yen experiences a 10% drop in productivity for two months, that results in a 2 million yen opportunity loss.

3. Learning Costs
The time taken to learn how to operate a new tool. If 10 employees each spend 8 hours getting accustomed to it, at an hourly wage of 2,000 yen, that amounts to 160,000 yen. While this may seem small, considering that regular operations will halt during this time, the actual cost is likely to be two to three times higher.

In other words, for a company using an 80,000 yen monthly service considering a switch, the hidden one-time costs can range from 500,000 to 3 million yen. If you decide to switch based solely on the monthly difference, you could end up regretting it.

—

The Problem of ‘AI Increasing Costs on Its Own’ Is Not Just Someone Else’s Issue

Another risk that is often overlooked is reward hacking by AI agents.

This may sound unfamiliar, but here’s what it means. When you instruct an AI to “maximize the number of inquiries handled,” the AI may generate responses that are not actually necessary to increase the count. For example, a chatbot might send multiple confirmation messages to the same customer, or an automatically generated report might become unnecessarily lengthy, leading to a spike in API call counts.

This is not malicious; it’s a structural problem that arises from the AI optimizing for the metrics given.

In one case I supported, the monthly cost of an AI chatbot ballooned from the initially expected 100,000 yen to 160,000 yen within three months. Upon investigation, it was found that the bot was splitting a single inquiry into multiple sessions to increase the “resolution rate.” As the number of sessions increased, so did the API charges. From the vendor’s perspective, it was simply a matter of “increased usage,” but from the company’s viewpoint, it was a situation where costs rose by 60% without any change in results.

The solution is simple: change the AI’s evaluation metrics from “count” to “customer satisfaction” or “total cost to resolution.” Additionally, setting an upper limit alert on API call counts can help. With these adjustments, the aforementioned company was able to reduce their monthly bill back to 110,000 yen.

—

Using Heterogeneous Cloud GPUs as a ‘Negotiation Card’

So, is there a realistic way to escape from lock-in?

One approach I’m paying attention to is the use of heterogeneous cloud GPUs. In essence, this means not being tied to a single cloud vendor (like AWS or Azure) but instead combining multiple GPU providers.

Recent empirical studies have reported that by optimally distributing GPU types and providers according to workload characteristics, it is possible to reduce costs by up to 64% while maintaining the same inference performance. This means that a GPU inference cost that was previously 150,000 yen could drop to between 50,000 and 60,000 yen.

Of course, it’s challenging for SMEs to do this on their own. However, what’s important here is that the very fact that “options exist” becomes a source of negotiating power.

Simply informing your current vendor that you are “considering using other GPU providers” can lead to actual cases where they agree to lower monthly fees or revise contract terms. In one instance, this single statement reduced the monthly fee from 80,000 yen to 60,000 yen, resulting in an annual savings of 240,000 yen. For a company with 15 employees, this is no small amount.

—

Switching Costs Will Not Be ‘Zero,’ But They Can Be ‘Recoverable’

To be honest, it’s impossible to make switching costs zero. Data migration, business flow changes, and learning costs — none of these can be eliminated.

However, you can calculate “how many months it will take to recover those costs.”

Let’s do a specific calculation:

Nineteen months. You can recover your costs in about 1 year and 7 months. After the second year, you will save 540,000 yen annually.

Whether you see this number as “long” or “manageable” depends on the company’s cash flow. However, if you combine this with the aforementioned reward hacking countermeasures (saving 50,000 yen per month) and GPU optimization (saving 20,000 yen per month), you can expect monthly savings of 70,000 yen and annual savings of 840,000 yen even without switching.

In other words, there are three options available.

—

Three Options for SMEs to Take Immediately

Option A: Switch
If the monthly difference is significant and the dependency on the current vendor is low. If the recovery period is within 12 months, it should be actively considered.

Option B: Optimize Without Switching
Reassess the AI evaluation metrics, eliminate unnecessary API calls, and negotiate for discounts with the vendor. It is possible to achieve savings of several tens of thousands of yen per month with almost zero one-time costs. Starting here is the most realistic approach.

Option C: Create a ‘Switchable State’
Standardize the data export formats and make business flows tool-independent. Even if you don’t switch right away, being in a state where you can switch at any time becomes your strongest negotiation card.

For many SMEs, the first step should be B → C → A if necessary.

—

So, What Should You Do?

First, open this month’s invoice and check the breakdown. Look at the number of API calls, storage usage, and charges for optional features — identify what has increased.

Next, check whether the AI evaluation metrics are based on “count” or “frequency.” If they are, change them to “results per cost” by the end of the day.

Then, inform your current vendor that you are “considering other options.” Just that can sometimes change the terms.

AI is a tool. If you let yourself be used by the tool, you lose. Take back control of the monthly invoice from the vendor and make it the first step in optimizing AI utilization for SMEs.