From Tacit to Fragmented: When Knowledge Stops Behaving

Retirement is not just an HR event. In many organizations, it is a data-loss event with a farewell cake.

A veteran maintenance worker leaves. A senior nurse changes hospitals. A plant supervisor retires after thirty years of noticing small abnormalities before anyone else sees them. The company still has manuals, checklists, inspection records, training videos, and perhaps a cheerful knowledge portal that everyone praises and nobody searches. What disappears is harder to name: the half-formed judgment, the workplace memory, the sense that “this noise is different from last month’s noise.”

The old knowledge-management answer was simple: capture the expert’s tacit knowledge and convert it into explicit knowledge. Interview the veteran. Write the manual. Build the database. Train the junior employee. Elegant, administratively comforting, and often wrong in the usual corporate way: clean on the slide, fragile in the field.

Naoshi Uchihira’s paper on the GenAI SECI model argues for a more realistic route.¹ The point is not that generative AI finally turns every expert intuition into a perfect manual. That would be a lovely fairy tale, and enterprise software has already sold enough of those. The paper’s central move is different: generative AI may make it useful to work with incomplete, partial, messy knowledge fragments before they become polished documentation.

That distinction matters. If the old model treated tacit knowledge as something waiting to be fully formalized, the GenAI SECI model treats a large part of workplace knowledge as something that can be partially expressed, digitally accumulated, loosely structured, and then returned to humans for reflection. The machine does not become the wise elder. It becomes a very fast, very strange assistant to the process by which humans learn from each other.

The old SECI problem was not the arrows; it was the missing middle

The paper begins from the SECI model, the familiar knowledge-creation framework associated with Nonaka and Takeuchi. SECI divides organizational knowledge into tacit and explicit knowledge, then explains knowledge creation through four conversion modes: socialization, externalization, combination, and internalization.

The framework remains useful because it gives organizations a vocabulary for learning. Its weakness, at least in digital knowledge-management practice, is that it makes externalization look cleaner than it is.

In real workplaces, much of what people know is not neatly divided into “impossible to express” and “ready for the manual.” There is a wide, annoying middle. A worker may not be able to articulate a complete rule, but can describe what they noticed in a particular situation. A technician may not produce a general theory, but can attach a photo, a voice note, and a sensor reading to an incident. A nurse may not formalize a protocol change, but can explain why one patient pattern felt risky.

Traditional systems struggled with this middle layer. They wanted full formalization: categories, rules, structured inputs, carefully maintained databases. The cost of preparing knowledge was high, and retrieval was often disappointing. The knowledge portal became a museum of stale documents. Visitors were rare.

The paper’s contribution is to make this middle layer explicit.

Latent knowledge is the part that can speak, but only in fragments

Uchihira separates workplace knowledge, or Gen-Ba knowledge, into three continuous layers: explicit knowledge, latent knowledge, and tacit knowledge in the narrow sense.

This is the paper’s most useful conceptual adjustment. The practical target is not pure tacit knowledge, because pure tacit knowledge is precisely the part that resists codification. The target is latent knowledge: the part that workers may not normally articulate, but can express partially when they are in context, asked the right question, or shown the right trace of the work.

That is where “Digital Fragmented Knowledge” enters.

Digital Fragmented Knowledge is not just another name for documents. It refers to explicit and latent knowledge that has been accumulated in cyberspace in some digital form: text, voice, photos, video, sensor data, and other fragments. These fragments may not form a complete manual. They may not even be clean. Their value is that they preserve situated traces of workplace judgment.

The paper is careful here in a way many AI discussions are not. It does not pretend that everything tacit becomes explicit once a large language model arrives with a confident tone. Instead, it says that generative AI makes incomplete fragments more usable than they were before.

That is a smaller claim. It is also a better one.

The mechanism: capture fragments before forcing them to become manuals

The GenAI SECI model keeps the four SECI processes, but changes the material moving through them. In the original model, the dream was often to externalize tacit knowledge into explicit knowledge, combine it, and then internalize it again. In the GenAI SECI model, the organization can externalize knowledge only partially, accumulate it as Digital Fragmented Knowledge, loosely combine it, and use it to support human internalization.

A simplified mechanism looks like this:

Physical workplace experience
        ↓
Partial externalization into fragments
        ↓
Digital Fragmented Knowledge
        ↓
Loose structuring and linking
        ↓
AI-supported recommendation
        ↓
Human reflection, discussion, and practice
        ↓
Amplified workplace knowledge

Generative AI appears in three places.

The word “loose” is doing serious work. The paper does not propose a return to old expert systems, where knowledge had to be registered as formal logic or IF-THEN rules. It proposes a system that can tolerate partial structure. This is exactly where generative AI has practical value: not because it magically understands the factory, hospital, farm, or service counter, but because it can help cluster, summarize, link, and retrieve imperfect traces.

That does not eliminate human interpretation. It changes where human effort is spent. Instead of spending all the effort on cleaning knowledge before storage, the organization spends more effort on reflection after retrieval.

This is less tidy. It is also how work actually works.

The Digital Knowledge Twin is an architecture, not a victory lap

The paper then presents a concrete architecture: the Digital Knowledge Twin System. This is important because the model would otherwise risk becoming another conceptual diagram that looks sensible until someone asks who will build it on Monday morning.

The system has three connected processes.

First, semi-automatic aggregation supports externalization. The architecture extends earlier Smart Voice Messaging System work, which collected skilled workers’ awareness through voice messages and photos. The new direction is to add voice, photos, physical sensor data, and other workplace traces, then use generative AI to consolidate them interactively into knowledge fragments.

Second, loose Gen-Ba knowledge combination links fragments with work documents such as manuals and operation procedures. The paper discusses extending procedure-based and purpose-based knowledge graph work that had been explored in caregiving and maintenance inspection contexts. The goal is not to create one perfect ontology of the workplace, because that way lies committee meetings and despair. The goal is to create enough structure that fragments can be used in internalization.

Third, automatic recommendation supports internalization. The system recommends relevant fragments for workshops involving skilled and less experienced participants. These workshops matter because, in this model, knowledge is not “transferred” merely when AI presents a result. It becomes practical when humans connect the fragment to lived experience.

The paper’s figures and comparison table serve different purposes. Treating them all as “evidence” would overstate the paper; treating them as mere decoration would miss the architecture.

This distinction is essential for a business reader. The paper is not offering a benchmark table where Model A beats Model B by 7.3 percentage points. It is offering a conceptual model plus an early architecture. That makes the contribution strategic, not statistical.

For managers, the right question is therefore not “What is the measured ROI?” The paper does not give that. The better question is: “Does this architecture describe a real bottleneck in our organization, and can we test it in a bounded workflow?”

AI is not the new expert; it is the assistant to symbol grounding

The paper positions GenAI SECI against two recent model families: GRAI and AKI.

GRAI treats generative AI as a new actor in the knowledge creation process, adding human–machine dimensions to the SECI interactions. AKI goes further by introducing Artificial Knowledge as a knowledge type generated by AI and refined through dialogue with humans. Both approaches respond to a real pressure: generative AI is too active to be treated like a normal database.

GenAI SECI makes a different editorial choice. It does not promote AI into a new organizational colleague. No desk, no badge, no mysterious job title like “Synthetic Knowledge Partner.” Mercifully.

Instead, it treats generative AI as auxiliary means. The new knowledge category is not “what AI knows.” It is Digital Fragmented Knowledge: digitally accumulated fragments of human workplace knowledge, including explicit and latent knowledge. AI helps collect, organize, and recommend those fragments, but it does not perform the final act of understanding.

That final act is symbol grounding.

In the paper’s framing, AI-generated or AI-organized content is not grounded in workplace meaning by itself. A fragment becomes practical knowledge when a human connects it to experience, action, and context. The paper links this to reflective observation in Kolb’s experiential learning model: people make sense of experience by reflecting on it, comparing it with other traces, and integrating it into their own working judgment.²

This is why workshops are not a soft add-on. They are part of the mechanism. A recommendation engine may surface a fragment, but the workshop gives skilled and less experienced participants a place to argue with it, explain it, recognize it, reject it, and eventually internalize it.

That is the correction to the common misconception. GenAI tacit-knowledge management is not about fully converting expert intuition into clean manuals. Nor is it about letting AI create organizational knowledge autonomously. It is about using incomplete fragments to support human-led reflection.

The distinction sounds philosophical until the first deployment fails because someone replaced the workshop with a chatbot and called it transformation.

The business value is thinner knowledge loss, not magical expertise transfer

For businesses, the strongest relevance is in settings where expertise is local, embodied, and at risk of disappearing. The paper explicitly discusses Gen-Ba contexts such as manufacturing, maintenance and inspection, medical and nursing care, agriculture, sales, and customer service. These are not industries where knowledge lives only in policy documents. They are environments where judgment often appears in small situated observations.

The business pathway is practical:

This is where the ROI conversation should become less theatrical. The value is not that the company buys an AI knowledge system and suddenly every junior employee becomes a veteran. The value is that the organization may reduce the loss rate of field memory. It may shorten the distance between incident and learning. It may lower the cost of preparing useful training material. It may make expert review sessions more concrete.

Those are testable business hypotheses. They are not proven by this paper. A serious pilot would measure things like onboarding time, repeated incident rates, time spent preparing workshop material, quality of recommendations, expert review burden, and whether less experienced staff make better situational judgments after exposure to fragments.

A less serious pilot would announce “AI-powered tacit knowledge transformation” and then upload PDFs into a chatbot. This is cheaper, faster, and almost certainly less interesting.

The boundary: fragments are useful only if the learning loop survives

The paper is clear that the Digital Knowledge Twin System is still under development. It also states that full integration and evaluation of generative AI remain future tasks. Near the conclusion, it calls for empirical validation across diverse workplace settings and further conceptual development of Digital Fragmented Knowledge.

Those limitations are not footnotes to be politely ignored. They define how the paper should be used.

First, this is not a universal proof that generative AI improves knowledge transfer. It is a model and architecture proposal. The correct business response is a constrained pilot, not a full-platform procurement ritual.

Second, Digital Fragmented Knowledge depends on capture quality. A fragment without context can mislead. A sensor trace without the worker’s situated observation may become just another dashboard squiggle. A voice note summarized too aggressively may lose the very nuance that made it valuable.

Third, loose structure must remain useful structure. Knowledge graphs can help connect fragments to procedures and purposes, but they can also become another ontology project where the diagram becomes more loved than the work. This is a known organizational disease; symptoms include taxonomy meetings and a sudden lack of field staff in the room.

Fourth, the internalization layer is fragile. If workshops disappear, the model loses its grounding mechanism. AI recommendation alone can surface fragments, but it cannot guarantee that humans understand them, trust them, or know how to act on them. The model’s human-centered claim only holds if the organization actually protects time for human reflection.

Finally, governance will matter in real deployments. The paper discusses voice, photos, videos, physical sensors, and human sensors. In a workplace, these are not neutral data exhaust. They involve worker consent, surveillance boundaries, data ownership, and the politics of who gets labeled “expert.” The model does not resolve those issues, but any business implementation will have to.

What this paper changes in the AI knowledge-management conversation

The useful shift in this paper is not that it says “use generative AI for knowledge management.” Everyone has already said that, usually while pointing at a chatbot.

The useful shift is more precise: stop treating incomplete knowledge as failed documentation.

A photo, a voice note, a sensor trace, a short comment after an abnormal event—these are not poor substitutes for a manual. They are a different layer of organizational memory. Generative AI makes that layer more workable because it can aggregate, summarize, connect, and recommend fragments at lower cost than previous systems. But the knowledge still has to click into place inside human practice.

That gives the GenAI SECI model its disciplined modesty. It does not ask us to believe that AI has become an organizational knower. It asks us to redesign knowledge systems around the fact that much workplace knowledge arrives fragmented, contextual, and unfinished.

For companies facing expert retirement, operational turnover, or repeated loss of field judgment, that may be enough to justify attention. Not because the model is already proven across industries. It is not. But because it points to a better pilot question:

Can we capture the fragments that usually disappear, connect them loosely enough to be found, and use them in human learning before the knowledge walks out of the building?

That is less glamorous than “AI replaces expertise.” It is also less silly.

And in knowledge management, less silly is already progress.

Cognaptus: Automate the Present, Incubate the Future.