The Fear of Losing a Veteran

If you are a small or medium enterprise owner, you have likely thought at least once:

“Who will manage this task if that person leaves?”

The decision-making criteria, processes, and know-how that reside solely in the mind of a veteran employee earning 400,000 yen a month are not documented in manuals. Even if handover documents are created, they convey only about 30% of the original knowledge. This is the essence of “individual dependency” and the greatest risk for small and medium enterprises.

In response to this issue, an intriguing experiment has provided an interesting answer. The Wired article “I Built AI Clones of My Coworkers. Things Got Weird” reported on the attempt to create AI clones of colleagues.

To put it succinctly, the mechanism for transferring a veteran’s knowledge to AI is no longer a “thing of the future.” Moreover, the cost has dropped to around 50,000 yen per month.

What Was Done—An Experiment in Copying Colleagues to AI

The experiment introduced in the Wired article is straightforward. It involved training AI to learn the speech patterns, expertise, and decision-making quirks of specific colleagues, creating an AI that responds in a manner characteristic of that person.

The crucial point is that this is not just a simple chatbot. While a generic FAQ bot merely “returns what is written in the manual,” an AI clone attempts to reproduce the implicit knowledge and context that are not documented in manuals. In other words, it aims to duplicate the tacit knowledge that had been individualized.

The article reports that one clone possessed knowledge of improvisational theater, while another had the decision-making logic for specific tasks, reflecting individual characteristics. While not perfect, the transition from “I don’t know unless I ask that person” to “I can get a general idea by asking the AI” has been made.

This difference is monumental for small and medium enterprises.

The Cost Structure Has Changed—From 3 Million to 50,000 Yen

Traditionally, the methods for retaining a veteran’s knowledge within an organization have been limited.

• Manual Creation: Hiring external consultants can cost between 1 to 3 million yen. Moreover, the moment it is completed, it becomes outdated.
• On-the-Job Training (OJT): This can take six months to a year. During this time, the veteran’s productivity is halved.
• Knowledge Management System Implementation: Initial costs can reach several hundred thousand yen, plus monthly fees of tens of thousands. This system is designed for large corporations and is too cumbersome for a company with around ten employees.

In contrast, what does the cost structure for AI clones look like?

• LLM API Usage Fees: For a GPT-4 class, it ranges from 10,000 to 30,000 yen per month (depending on usage).
• Building RAG (Retrieval-Augmented Generation): Storing internal documents and chat histories in a vector database. Using cloud services, this can cost a few thousand to 10,000 yen per month.
• Initial Prompt Design and Tuning: If done in-house, it only incurs labor costs. Outsourcing can be done for several tens of thousands of yen.

In rough terms, the running cost is under 50,000 yen per month. If a veteran’s monthly salary is 400,000 yen, it means that 70-80% of their knowledge can be operational 24/7 for just 50,000 yen.

Moreover, AI does not quit. It does not fall ill. There are no days when it is in a bad mood.

The Notion That “It Must Be Perfect to Use” Is Wrong

A common counterargument arises here: “AI makes mistakes,” or “There’s no way to replicate a veteran’s judgment 100%.”

That is correct. Achieving 100% is impossible. However, consider this:

If a veteran leaves and a newcomer takes over, what would the reproduction rate be? Realistically, it would be around 30-50% at best. Moreover, it would take over six months to become proficient.

If an AI clone can respond with 70% accuracy from day one, that is better than a newcomer six months in. Additionally, as feedback is provided, accuracy continues to improve. Unlike human handovers, there is no risk of “forgetting what was taught.”

Seeking perfection and doing nothing is far less valuable for small and medium enterprises than operating with 70% accuracy starting today.

How to Get Started—Three Steps

“I get that it’s interesting. But how do we implement this in our company?” Let’s discuss that.

Step 1: Accumulate the Veteran’s “Decision Logs”

The first step is to convert the veteran’s decision-making into text data. There’s no need for an elaborate system.

• If you use Slack or Chatwork for daily reports, that history can serve as learning data.
• Email histories of customer interactions.
• Records of past estimation judgments.
• Recording for 15 minutes once a week on “why this decision was made” and converting it into text.

After three months of accumulation, you will have a sufficient amount of data.

Step 2: Create a “Personal AI” Using RAG

Store the collected data in a vector database and connect it to an LLM. This is the structure known as RAG (Retrieval-Augmented Generation).

It may sound technically challenging, but today, tools like Dify, Flowise, or GPTs allow for construction without coding or with minimal coding. Even without engineers, a staff member with some IT knowledge can shape it within one to two weeks.

Step 3: Use and Develop in the Field

An AI clone is not a one-and-done creation. It needs to be used in the field, with feedback such as “this answer is incorrect” or “this is how to judge in this case” to improve accuracy.

This is crucial. It is necessary to start while the veteran is still present. If you wait until they leave, there will be no one to provide feedback. Countermeasures against individual dependency should be implemented before a crisis arises, not after it has occurred.

A Structure Where Small and Medium Enterprises Can Win

This concept is actually difficult for large corporations to replicate. Why?

Large companies often try to create a “unified knowledge base for the entire company,” spending six months on requirements definition, three months on vendor selection, and one year on implementation, only to find it ultimately unused—this pattern has been repeated. Decision-making is slow and disconnected from the realities on the ground.

On the other hand, small and medium enterprises with 10 to 50 employees can say, “Let’s try it starting next week.” The proximity between veterans and newcomers allows for a faster feedback loop. Instead of a company-wide rollout, they can begin with a small experiment to “clone the knowledge of one veteran.”

Being small becomes a weapon here.

What’s Truly Scary Is Not the “Evolution of Technology”

The technology for AI clones will undoubtedly improve in accuracy in the future. With multi-modal capabilities, it will be able to learn not just from text but also from images and videos. Responses via voice will also become more natural.

However, what is truly frightening is not the evolution of technology itself.

It is the fact that “the neighboring company is already doing it.”

Competitors in the same region and industry are replicating veteran knowledge with AI, creating a system that can respond 24/7 for just 50,000 yen. Meanwhile, your company is still fearful of veteran departures, creating handover documents on paper.

This gap will become an irretrievable difference in just six months.

Conclusion: What You Should Do Today

The ultimate form of eliminating individual dependency is to “replicate the veteran’s mind in AI.” And the cost has already dropped to 50,000 yen per month.

The one thing you should do today is to choose the most individualized task in your company and start accumulating the decision logs of that veteran.

There is no need to design a perfect system. Just organizing Slack histories will suffice. Simply start weekly 15-minute interviews using a recording app.

It will be too late to realize after a veteran leaves that “that knowledge is nowhere to be found.” Systematization should begin while there is still time.