"Is there any way to preserve the knowledge locked inside that person's head?"—facing the retirement of veteran employees, many companies confront exactly this question.

According to the 2022 White Paper on Monozukuri (Manufacturing) for fiscal year 2021, the number of workers in the manufacturing industry has fallen by roughly 1.57 million over about 20 years, while the number of young workers has declined by 1.21 million over the same period. The risk of losing the knowledge and expertise (tacit knowledge) held by veterans is rising not only in manufacturing but across every industry. For an overview of the skills-transfer challenges in each industry and the AI-driven patterns for addressing them, see How to Use AI for Skills Transfer.

Traditionally, passing on tacit knowledge has relied on OJT (on-the-job training) and manual creation. However, in the Ministry of Health, Labour and Welfare's "FY2024 Basic Survey on Human Resources Development," among establishments that reported some kind of problem with human resource development (79.9% of all establishments), 59.5% answered that they "lack the personnel to provide instruction." With insufficient resources on the teaching side, an approach that uses AI to convert tacit knowledge into explicit knowledge and accumulate it as an organizational asset is drawing attention.

This article first clarifies what tacit knowledge is, then explains, step by step, concrete methods for transferring tacit knowledge using AI.

The difference between tacit knowledge and explicit knowledge

Tacit knowledge refers to knowledge based on personal experience and intuition that is difficult to put into words. Explicit knowledge, by contrast, is knowledge expressed in documents, formulas, diagrams, and the like—knowledge in a form that anyone can reference. The concept of tacit knowledge was originally introduced by the philosopher Michael Polanyi; Ikujiro Nonaka, professor emeritus at Hitotsubashi University, applied it to management studies and systematized the conversion between tacit and explicit knowledge as the SECI model.

In the SECI model, organizational knowledge creation is said to cycle through four stages: "Socialization (sharing tacit knowledge) → Externalization (converting tacit knowledge into explicit knowledge) → Combination (combining explicit knowledge with other explicit knowledge) → Internalization (embodying explicit knowledge as tacit knowledge)." Turning tacit knowledge into AI is an attempt to accelerate the "Externalization" process among these with technology.

That said, not all tacit knowledge is suited to being turned into AI. Tacit knowledge comes in three types with different characteristics, each calling for a different approach to transfer.

The three types of tacit knowledge and their suitability for AI

Tacit knowledge can be divided, according to its nature, into three types: "procedural," "judgment-based," and "values-based" (for details, see our article on how to reduce key-person dependency). This article digs into each type's suitability for AI and the concrete methods involved.

Procedural tacit knowledge — Suitability for AI: High

This is knowledge about "how to do something." It includes work procedures, operational know-how, and the order and timing of tasks. This type is relatively easy to put into words, and the difficulty of turning it into AI is also low.

A common method is to record procedures as text or video, feed them into a RAG system, and make them searchable through an internal chatbot. You can build a mechanism that, when asked "What is the procedure for operating XX?", responds with the content of the relevant procedure manual along with its source.

In an initiative to pass on tacit knowledge that Lion and NTT Data launched in June 2024, expert know-how in the manufacturing process for powdered laundry detergent was documented as a "collection of key points" and incorporated into a generative-AI-powered "knowledge transfer AI system" (Source: NTT Data press release, June 3, 2024). By referencing this system, newly assigned members can efficiently make use of expert knowledge.

Judgment-based tacit knowledge — Suitability for AI: Moderate

This is knowledge about "how to make decisions." It includes decision-making criteria that vary by situation, rules for handling exceptions, and how to spot anomalous values. Rules of thumb such as "this figure is a little higher than usual, but in this season this range is normal" often cannot be clearly articulated even by the person themselves.

Turning this into AI inevitably requires a "knowledge extraction" process that first puts the decision criteria into words (explained in detail later in the "knowledge-extraction interviews" section). Once the extracted decision criteria are built into a RAG database, the AI can present reference information in similar decision-making situations.

However, because judgment-based tacit knowledge is highly context-dependent, the AI's answers will not be correct in every situation. The AI's response is no more than reference information indicating "this is how a veteran might decide," and the operation must be designed on the premise that the final decision is made by a human.

Values-based tacit knowledge — Suitability for AI: Limited

This is knowledge about "why we do things." It is tacit knowledge rooted in the values of an individual or organization, such as management philosophy, attitude toward work, commitment to quality, and how to engage with customers.

This type is the hardest to put into words, and a complete conversion into explicit knowledge is difficult. That said, there is an approach in which the statements and actions of an executive or expert are recorded and built into a conversational AI (a personal AI), making that person's "tendencies in thinking" available for reference (for details, see our article on what is a CEO bot?).

Turning values-based tacit knowledge into AI must be used with an understanding that it serves only a limited role—not "producing the correct answer" but "providing reference information on how that person would think."

How to run "knowledge-extraction interviews" that draw out tacit knowledge

In turning tacit knowledge into AI, the most important and labor-intensive step is "knowledge extraction"—drawing out the knowledge inside a veteran's head. Simply asking "Please tell me how you do it" will not surface the things the person does unconsciously.

By using the following five questioning techniques, you can improve the accuracy of tacit-knowledge extraction.

Technique 1: Start from a concrete situation

Rather than abstract questions like "What is difficult about this work?", start from a specific case: "In last week's XX project, at what point did you make what decision?" Prompting recall of a concrete situation makes it easier to verbalize the decision-making process that is performed unconsciously.

Technique 2: Repeat "why" three times

In response to an answer like "In this case I do XX," dig deeper by asking "Why do you do XX?" The first answer often stays at a surface-level explanation of the rule; only with the second and third "why" do the underlying decision criteria and rules of thumb emerge.

Technique 3: Ask about failures

By asking "Have you had experiences where things didn't go well?" or "Have you ever made a wrong decision in this situation?", you draw out the cautions and exception rules behind successes. Failure cases are an effective means of making visible the pitfalls that veterans unconsciously avoid.

Technique 4: Have them articulate the difference from beginners

By asking "When a newcomer does the same work, where are they likely to go wrong?", you bring to the surface knowledge that veterans take for granted but that is in fact not being shared.

Technique 5: Start from the five senses

In manufacturing and on-site work, there is tacit knowledge based on the five senses, such as "pay attention when the sound changes" or "judge by the temperature when you touch it." Deliberately posing questions about each of "appearance," "sound," "touch," "smell," and "timing" provides clues for drawing out the bodily knowledge that is said to be hard to verbalize.

As a guideline, interviews should run about two hours each, ideally conducted two or three times per target task. Record them, transcribe them into text, and use them as training data for the AI.

Four steps to transferring tacit knowledge using AI

Step 1: Identify the knowledge to be transferred

You do not need to turn every veteran's knowledge into AI. First, identify knowledge that has a high impact on operations and is held by only a limited number of people. A practical approach is to work backward from the question "Which task would cause the most trouble if this person left?"

Step 2: Extract and document the knowledge

Conduct the knowledge-extraction interviews described above and convert the tacit knowledge into text. At the same time, gather existing operational documents (procedure manuals, response histories, emails, chat logs). In the Lion and NTT Data case, the "collection of key points" was created by combining interviews with experts and workshops among employees across a wide range of positions.

Step 3: Build the AI system

Incorporate the text-based data into a RAG system and make it searchable as an internal chatbot. Technically, internal documents are stored in a vector database, and in response to an employee's question the AI searches for and cites relevant knowledge while generating an answer.

At this stage, it is important to make the AI always display the "name of the source document" with its answers. By clearly indicating the source, users can verify the reliability of the AI's answers for themselves.

Step 4: Verify and improve

Have the veterans themselves check the answers of the AI system you built, and verify the accuracy of the responses. By reflecting feedback such as "the AI answered this way, but in reality there is an exception in this case," you progressively raise the accuracy of the answers.

Continuously running this verify-and-improve cycle is the key to maintaining the quality of a tacit-knowledge-transfer AI. It is not a build-it-once-and-done effort; you need an operational structure that keeps updating the data in line with changes in the work.

Tacit knowledge that is hard to turn into AI, and how to handle it

AI cannot perfectly reproduce all tacit knowledge. In particular, the following kinds of knowledge are areas that current AI technology struggles to cover.

Knowledge rooted in physical movement. Skills learned with the body, such as "how much force to apply with the hands" or "fine adjustment of a tool's angle," are difficult to convert into text. In this area, video manuals and visualization through sensor technology are effective, but complete reproduction is genuinely difficult. Approaches that combine sensor data with AI to analyze tacit knowledge at the level of physical movement—such as "Takumi AI" offered by Mitsubishi Research Institute—are also beginning to emerge, but the hurdles to adoption remain high.

Judgments that depend heavily on context and situation. Sophisticated judgments in which multiple contexts are intertwined—such as "it's better not to phrase it this way with this customer" or "this client is busy at this time of year, so take a different approach"—are an area that is difficult for AI to reproduce. For this type of tacit knowledge, transferring it directly through a multi-person system (cross-training) or paired work is more realistic than relying on AI.

Cases where the person is not even aware that they hold tacit knowledge. Work that a veteran feels they are "just doing normally" may in fact contain sophisticated tacit knowledge. In this case it is hard to draw out even through knowledge-extraction interviews, and an effective approach is to observe the actual work from alongside and have a third party record the differences from a beginner.

Conclusion

Transferring tacit knowledge with AI is an important effort to protect an organization's knowledge assets amid advancing veteran retirements and labor shortages.

There are three keys to success. First, correctly classify the type of tacit knowledge (procedural, judgment-based, or values-based) and choose the method suited to each. Second, carefully conduct knowledge-extraction interviews to secure data of a quality the AI can learn from. And third, build an operational structure that does not over-trust the AI's answers and continuously runs the verify-and-improve cycle.

Not all tacit knowledge can be turned into AI, but by centering on procedural tacit knowledge and gradually expanding the scope of AI adoption, you can steadily reduce the "we can't figure it out without that person" situation.

At Teraverse, we provide generative AI solution development tailored to the type of tacit knowledge—from search AI for internal knowledge, to chatbots that reproduce veterans' decision criteria, to conversational AI that carries on an executive's philosophy. If you are struggling with transferring tacit knowledge, please feel free to contact us.