Conclusion First: Customer Support Labor Costs Are No Longer Just a “Numbers Game”

Brazil’s digital bank Nubank is automating a significant portion of its customer support with AI agents, all while serving over 100 million users. Customer satisfaction has not only remained stable but in some areas, the NPS (Net Promoter Score) has improved by 37 points.

The impact of this story lies not in the fact that “large companies are doing amazing things with AI,” but rather in the potential for “this structure to become accessible to small and medium-sized enterprises at a cost of just 30,000 yen per month.”

Many small and medium-sized enterprises (SMEs) spend between 500,000 to 1,000,000 yen on customer support. Hiring just two or three people can lead to those costs. What happens when that cost drops to 30,000 yen? This article will explore that structural change with concrete examples and figures.

—

Nubank Proves the Fact That “AI Support is Viable”

Nubank’s AI support system is built on four pillars:

1. Evaluation Pipeline — A mechanism to quantitatively measure the quality of AI responses and rapidly iterate improvements.
2. Context Engineering — A design that accurately conveys the user’s context (transaction history, past inquiries, etc.) to the AI.
3. Continuous Accumulation of Training Data — Ongoing improvements to the model using actual inquiry data.
4. Online Measurement — Real-time quality monitoring in a production environment.

The key point is that this is not just about “implementing AI and calling it a day”; it is the very mechanism for rapidly cycling through PDCA (Plan-Do-Check-Act).

Let’s look at some specific results:

• Inquiries about card delivery: In a large-scale A/B test, the NPS for the group handled by AI improved by 37 points.
• Debt management, credit limits, card management, product explanations: Customer satisfaction improved across all five different areas.

Why does AI lead to higher satisfaction? The answer is simple: No waiting. 24/7 availability. Consistent responses. The decline in quality that often occurs in the evening when human operators become tired does not happen with AI.

I can hear some saying, “But Nubank is a large company, right?” That’s true. However, the cost of bringing this structure to SMEs is dramatically decreasing right now.

—

Breakdown of 30,000 Yen — Is AI Support Really Feasible for SMEs?

Let’s estimate the costs for SMEs to build a similar AI customer support system.

Initial Setup Costs (One-time)

This is cheaper than hiring one part-time employee for six months.

Monthly Running Costs

A monthly cost of 30,000 yen is a very realistic figure.

On the other hand, hiring one human operator costs at least 200,000 – 250,000 yen. With two operators, it would be 400,000 – 500,000 yen. That’s a cost difference of over 10 times.

Moreover, AI operates 24/7, providing immediate responses to late-night inquiries. This is a scenario where SMEs can achieve what large companies have accomplished through three-shift systems for just 30,000 yen a month.

—

The Interesting Alternative of Edge Devices

There are now methods that do not rely on cloud APIs. This involves running LLMs directly on edge devices.

Recent performance comparison data is intriguing.

The combination of Raspberry Pi 5 + Hailo-10H is particularly noteworthy. It consistently delivers 6.9 tokens/second with a mere 2W of power consumption. The total cost for the device and NPU is around 20,000 yen, with electricity costs at just a few dozen yen per month.

What does this mean?

It means that even small and medium-sized enterprises that do not want to pay for API usage or share data externally can have the brain of AI support in-house.

Currently, this is limited to small models (7B – 13B parameters), but it is sufficiently practical for handling FAQ-based inquiries. With an initial investment of 20,000 yen and almost zero running costs, this option itself is evidence of a structural change.

—

Where the True “Reversal Structure” Lies

Now we get to the main point.

The decrease in costs is merely a means. The real question is “What value increases as a result of lower costs?”

1. “Response Quality” is Liberated from Being Person-Dependent

Support in SMEs is often in a state where it cannot function without a specific person. If a veteran employee leaves, the quality collapses. By consolidating knowledge into an AI agent, response quality becomes independent of individuals. This not only raises the quality but also reduces business risk.

2. SMEs Can Compete with Large Companies on “Response Speed”

Call centers at large companies often have long wait times. Everyone has experienced being passed around endlessly by IVR (Interactive Voice Response). If SMEs can provide immediate responses through AI chat, they can outperform large companies in customer experience.

3. “Saved Costs” Can Be Used for Offensive Investments

If a support cost of 400,000 yen drops to 30,000 yen, approximately 4.4 million yen is saved annually. That money can be redirected to new product development or marketing. Cost savings from defense can transform into resources for offensive investments. This is the true reversal structure for SMEs.

4. “Customer Feedback” is Accumulated as Data

All inquiries handled by AI are logged. Information such as “what complaints are most common” and “which products are seeing an increase in inquiries” is automatically datafied. Information that previously existed only in operators’ heads becomes an asset for management decisions.

—

So, What Should We Do?

If SMEs want to start AI customer support, the steps are simple.

Step 1: Inventory Inquiries (1 Week)
Collect and categorize 100 past inquiries. 80% of them should be similar questions.

Step 2: Build a Knowledge Base (1-2 Weeks)
Gather answers to those “frequently asked questions” from veteran employees and text them out. Perfection is not necessary.

Step 3: Build the AI Agent (1-2 Weeks)
Use ChatGPT API, Claude API, or no-code tools like Dify or Botpress to create a chatbot. Just feeding the knowledge base into the prompts will get the first version running.

Step 4: Test on a Small Scale (1 Month)
There’s no need to assign all inquiries to AI right away. Start with just “initial responses outside of business hours.” This alone will significantly reduce the response load for the next morning.

Step 5: Measure and Improve (Ongoing)
Add questions that the AI could not answer to the knowledge base. Doing this once a month will steadily improve accuracy.

Initial investment: 200,000 – 550,000 yen. Monthly cost: below 30,000 yen. Just try it out; if it doesn’t work, you can stop. There’s little to lose.

—

Summary — The “Difference in Support Systems” is No Longer Just a Difference in Company Size

What Nubank’s case shows is not just the technical fact that “AI support is viable.” We have entered an era where the quality of customer support is determined not by the size or number of employees of a company, but by the “design capability of the system.”

24/7 support for 30,000 yen. An NPS improvement of 37 points. With edge devices, initial costs of 20,000 yen and almost zero running costs.

When you see these numbers, will you think “this has nothing to do with us” or will you think “let’s first inventory 100 inquiries”? That difference will become the difference in competitiveness a year from now.

There’s no need to imitate large companies. SMEs have their own ways of competing. Start small and win with the system. AI agents are becoming the most cost-effective weapon for that.