Conclusion First: The Error Rate of Personalization is “100%”

An AI memory error rate of 95% — looking at this number alone, one might think it’s “useless.”

But wait a moment.

What is the error rate of the “work only that person knows” in your company?

What if that person leaves? What if they fall ill? The information recall rate is 0%. It’s not just an error rate. It’s total loss.

A 95% error versus a 100% total loss. Which is better? The answer is clear.

In this article, we will unravel the meaning behind the “95% error” of AI memory and demonstrate, with concrete numbers, how small and medium-sized enterprises can benefit from using “imperfect AI.”

—

What Does the “95% Error” Actually Mean?

First, we want to accurately understand this number.

The recently reported “AI memory error rate of 95%” refers to the accuracy of large language models (LLMs) in retaining and reproducing past conversational contexts over the long term. In other words, it’s about whether the AI can accurately remember what a user said three months ago, and the result was that it was inaccurate 95% of the time.

This does not mean that “95% of AI’s responses are incorrect.” Confusing this can lead to poor judgment.

The accuracy of AI processing information given “at this moment” is entirely different from the accuracy of “retrieving past memories.” Even humans struggle to accurately recall the details of a meeting from yesterday, let alone three months ago.

The important thing is that if we externalize the mechanism of memory, this problem nearly disappears.

—

Viewing the Costs of Personalization in Numbers

Let’s calculate how much cost personalization generates in small and medium-sized enterprises.

Case: A manufacturing company with 30 employees. Order management is handled by one veteran employee (with 18 years of service).

• Annual salary of the veteran employee: 4.5 million yen
• Recruitment cost if this employee resigns: 800,000 to 1.2 million yen
• Time for a new hire to reach the same level: at least 12 months
• Loss due to decreased operational efficiency during that time: 300,000 to 500,000 yen per month × 12 months = 3.6 to 6 million yen
• Risk of complaints and lost orders due to handover omissions: incalculable (but will definitely occur)

In total, just losing one veteran employee can result in a loss of 5 to 8 million yen.

Moreover, this is only in the case of a “resignation” event. In reality, problems begin before resignation. On the days that person is absent, calls come in that no one can answer. The quality of work fluctuates based on that person’s mood or health. Estimates cannot be provided without consulting that person.

This is the true cost of personalization. If we include unseen costs annually, it’s reasonable to assume that 200,000 to 300,000 yen is constantly lost per person.

—

Making Use of “5% Memory” with Imperfect AI

So, is an AI with a 95% memory error rate — meaning it “only accurately remembers 5%” — really unusable?

The answer is no. The way to use it is simply different.

Instead of relying on AI memory, we should design it to “feed the AI the information to reference every time.”

Specifically, this involves:

1. Textualizing operational manuals and procedures (existing ones are fine; they don’t have to be perfect)
2. Database-ing past response histories and emails
3. Each time we ask the AI a question, automatically searching for and providing relevant information (a system known as RAG)

By doing this, we do not depend on the AI’s “memory.” The AI only needs to “read the information provided at that moment and respond appropriately.” This significantly increases accuracy.

What is the cost of creating this system?

• Knowledge base construction (organizing and inputting internal documents): 300,000 to 800,000 yen
• RAG environment setup: 200,000 to 500,000 yen
• Monthly usage fee for AI tools: 20,000 to 50,000 yen (annual total of 240,000 to 600,000 yen)
• Operation and maintenance: 10,000 to 30,000 yen per month (annual total of 120,000 to 360,000 yen)

Total for the first year: 900,000 to 2.3 million yen. From the second year onward: 360,000 to 960,000 yen.

Compare this with the annual loss of 200,000 to 300,000 yen due to personalization. You will nearly break even in the first year, and from the second year onward, it will definitely be a profit.

—

The Biggest Enemy is the Assumption That “It Must Be Perfect to Be Usable”

When talking to small and medium-sized business owners, the most common reaction is this:

“AI makes mistakes, right? Our operations can’t afford any errors.”

I understand the sentiment. However, I urge you to think calmly.

Is the current system of “only that person knows” free of mistakes? Really?

• The veteran provided estimates using outdated pricing due to their assumptions.
• An order quantity was incorrectly communicated verbally.
• They assumed specifications were understood without confirmation because it was “the usual.”

Such mistakes occur regularly. They just don’t surface because they are recovered by the veteran’s experience. Personalization is also a mechanism that makes errors invisible.

The advantage of AI is that errors become visible. Outputs are recorded in text. They can be checked. They can be corrected. They can be improved.

The knowledge in the head of a personalized veteran cannot be checked. It cannot be corrected. And one day, it suddenly disappears.

It’s clear which is the “manageable risk.”

—

Practice: Where to Start First

Let’s move from abstract discussions to three concrete actions.

Step 1: Record “That Person’s” Work for One Week

First, choose one task that is personalized. Have the person responsible record their work content in voice memos for one week. A smartphone recording app is sufficient. Then, use an AI transcription tool (costing a few thousand yen per month) to convert it into text.

Just this will reveal the skeleton of “what judgments are being made and how.” It doesn’t have to be a perfect procedure manual. The value lies in being able to see the skeleton.

Step 2: Ask AI with “Reference Materials” Attached

Paste the transcribed work records into ChatGPT or Claude and ask, “Organize this work as if explaining it to a newcomer.” A plan costing 2,000 to 3,000 yen per month is sufficient.

Show the output to the responsible person and get feedback: “This part is wrong,” “This part is correct.” Just repeating this three times will create a decent prototype of a usable knowledge base.

Step 3: Establish a Rule of “Ask AI First, Then Ask a Person”

When newcomers or other employees have questions about the work, establish a rule: “First ask the AI knowledge base → If unclear, then ask the responsible person.”

Just this will reduce inquiries to the responsible person by 30% to 50% in practice. The responsible person’s time will free up. They can use that time to further organize knowledge. A positive cycle will begin.

—

Now is the Time for Small and Medium-Sized Enterprises to Act

Large companies can cover for personalization due to their organizational depth. If there are ten people, someone will know.

Small and medium-sized enterprises are different. If one person leaves, it’s game over.

That’s why it makes sense to externalize knowledge with AI, even if it’s imperfect. It’s now possible for small and medium-sized enterprises to achieve knowledge management that large companies spend millions on, for just a few hundred thousand yen, with the help of AI.

With an AI memory error rate of 95%, there’s no need to be afraid of this number.

What’s truly frightening is the state of “everything ends if that person leaves,” which is still being neglected today.

There’s no need to wait for perfect AI. Even with 5% memory, if structured properly, it can overwhelmingly defeat 0% personalization.

I encourage you to start with just one task, beginning this week.