Conclusion First: We No Longer Need Giant AI Models

“I want to implement AI. But the costs…”

When talking to owners of small and medium-sized enterprises (SMEs) in rural areas, 90% say this. Using a giant model like GPT-4 can cost tens of thousands of yen per month. Each API call incurs charges, and data is sent to external servers. Security is also a concern. Ultimately, it ends with, “It’s still too early for us.”

However, this assumption is beginning to crumble.

A retail chain in Iceland has reduced store losses by 80% using AI. In invoice classification, a small language model (SLM) achieved 96% accuracy with just 100 training data points. Neither case involved the use of a giant model with hundreds of millions of parameters.

The misconception that “AI must be a giant model” is the biggest barrier to AI adoption for SMEs. Today, I will introduce two case studies that break down this wall.

—

Case Study 1: Iceland Retail Chain Cuts Annual Store Losses by £800,000

Iceland Foods, a frozen food chain with about 1,000 stores in the UK, faced store losses (shrinkage) due to shoplifting, internal fraud, and erroneous orders. This is a challenge across the retail industry, with a British Retail Consortium (BRC) survey indicating that total store losses in the UK reach approximately £7.7 billion (about 1.4 trillion yen) annually.

Iceland implemented an AI-based anomaly detection system that analyzes footage from in-store cameras in real time to detect suspicious behavior patterns.

What stands out are the numbers from the results:

• Before Implementation: Annual loss per store was around £1 million.
• After Implementation: Losses were reduced by 80%, down to around £200,000.
• Difference: Annual loss reduction of approximately £800,000 (about 150 million yen) per store.

This is not a story about giant language models. A machine learning model specialized in image analysis simply determines whether a scene is abnormal or normal. Its versatility is low. However, when focused on specific tasks, even a small model can yield overwhelming results.

Another important point is that the nature of human work has changed. Previously, store staff manually monitored surveillance cameras. After eight hours of watching the monitor, a human could find only a handful of anomalies due to concentration limits. AI can monitor all cameras simultaneously, 24/7. Staff can now focus on “response” instead of “monitoring.”

They did not reduce the number of people; they changed how people are utilized. This is the essence.

—

Case Study 2: Invoice Classification Achieves 96% Accuracy with Small Model SBERT—No Need for GPT-4

The second case is more understated but may have a greater impact for SMEs.

Those in accounting will understand the task of classifying incoming invoices by account category. “This is for supplies,” “This is for outsourcing,” “This is for communication expenses”—simple yet voluminous, and mistakes can affect financial statements. It is also a breeding ground for personal knowledge. If a veteran accountant leaves, the classification rules may disappear.

A study was published applying the small language model “SBERT (Sentence-BERT)” to this classification task. The results are as follows:

• Overall Classification Accuracy: Approximately 96% (F1 Score)
• Accuracy for New Client Invoices: Over 90%
• Required Training Data: Only about 100 items per client
• Model Size: Fewer than several hundred million parameters (less than 1/1000th of GPT-4)

Moreover, applying INT8 quantization can double the inference speed and halve the required memory. In other words, it can run on a regular laptop.

Let’s summarize the cost structure:

Should one continue to pay 150,000 yen monthly to an external API, or create a system for several tens of thousands of yen and run it for a few thousand yen monthly? For SMEs, this difference is decisive.

Moreover, data does not leave the company. Invoices contain client names, amounts, and contract details. Is it worth the risk of sending this information to an external server every time? With a small model, everything can be handled on the company’s own PC or server. For rural SMEs, the reassurance of “not having to send data outside” is significant.

—

Why “Small AI” is the Key for SMEs

There is a common structure in these two case studies.

“If you narrow down what you need to do, a small model is sufficient.”

GPT-4 can do anything. It can write poetry and code. However, if the task is simply to classify invoices into ten categories, it’s like using a large truck to go to the neighborhood convenience store. A light truck is sufficient. In fact, a light truck is more maneuverable.

The operational challenges faced by SMEs are mostly of the type that involves “repeated judgments over a narrow range.”

• Invoice classification
• Sorting inquiry emails
• Detecting inventory anomalies
• Extracting keywords from daily reports
• Checking estimates

None of these require general intelligence. It is sufficient to accurately determine specific patterns. And that is where small models excel.

When we organize the cost structure, it looks like this:

• Giant Models: High accuracy, versatile, high cost, external data transmission, vendor dependency
• Small Models: Sufficient accuracy, specialized, low cost, in-house completion, controllable by the company

Clearly, SMEs should choose the latter.

—

So, What Should We Do?

Here are three specific actions:

1. First, Identify “Repetitive Judgment Tasks”

List the “classification,” “judgment,” and “sorting” tasks that are repeated daily or weekly within the company. These are candidates for AI implementation. A grand DX plan is unnecessary. Start with one task.

2. Do Not Start with Giant Models

Trying out ChatGPT isn’t a bad idea. However, operational deployment is a different story. The structure of accumulating monthly fees while continuously sending data outside is not suitable for SMEs. The option of operating a small model like SBERT in-house should be considered from the outset.

3. Start by Accumulating 100 Data Points

In the recent study on invoice classification, practical levels were achieved with just 100 training data points. This means that “not having enough data” is no longer an excuse. Start with 100 data points. Simply begin recording classification rules today.

—

True Competitive Advantage Lies in Having a System

Large companies invest millions in giant models and run grand AI projects. However, many of these remain at the PoC (proof of concept) stage and do not reach operational deployment.

On the other hand, SMEs can make decisions quickly. When a business owner says, “Let’s do it,” they can start moving by next week. With small models, the initial investment is also small. Even if they fail, the damage is minimal.

While large companies are in meetings, SMEs can finish implementation.

This is the structure for reversal.

The case of Iceland is a large chain, but the essence of the technology used is “specialized small AI.” The same approach can be applied in local retail stores, manufacturing inspection lines, and document processing for professionals.

I want to ask:

What are the “judgment tasks” that someone in your company repeats every day?
Is it really a job that must be done by humans?

If the answer is No, then it’s time for small AI. There’s no need to wait for giant models. With 100 data points and the decision to “let’s try it,” you could be operational by next month.