The paper in brief
Experienced employees routinely use knowledge they would struggle to write down. They adjust a sequence because something sounds wrong. They recognize an unusual condition before it becomes a fault. They know when the standard instruction is sufficient and when the situation calls for a different response.
In Retaining Tacit Knowledge Through Dialogue: The AI Moderator in Work Processes, researchers at Fraunhofer IFF explore whether an AI-based dialogue system can help make more of that experience visible. Their prototype, KIMO—short for the German KI-Moderator—supports employees during complex work while prompting them to explain what they are doing and why.
The important idea is not simply to put a chatbot in front of a document repository. KIMO is designed to participate at the moment knowledge is being applied: during the work itself, when context, action, and judgment are present together.
The paper is available through the ECKM proceedings and DOI record.
Why experienced people leave gaps in documents
Tacit knowledge is bound to experience and context. An expert may know how to respond without consciously separating every observation, inference, and choice involved. Asked later to “document the process,” they are likely to describe the visible steps and leave out the cues and reasoning that have become automatic.
That is why a repository can contain accurate instructions and still fail a newcomer during a non-routine situation. The missing knowledge is often not another fact. It is the relationship between the facts: what the employee noticed, how they interpreted it, and why that interpretation changed the action.
The Fraunhofer IFF paper treats transfer as more than transmitting information. It is a process of making meaning. Context, application situations, and decision processes have to travel with the instruction if another person is expected to use it well.
From retrospective documentation to dialogue in the work
KIMO was developed around crane commissioning, a complex and safety-critical industrial setting. Operators need both hands for the task, so the prototype uses spoken, hands-free interaction. It provides step-by-step guidance from a validated, domain-specific knowledge base and can capture the user's spoken account while work is under way.
The approach combines several methods:
- Think-aloud practice: experienced workers verbalize actions and decisions while performing the task.
- Narrative prompts: open questions encourage stories and reflection that can expose implicit heuristics.
- Instructional scaffolding: guidance is delivered in manageable steps, with the user confirming progress.
- Targeted follow-up: the dialogue asks for procedural knowledge—what happens—and conditional knowledge—why, when, and under what circumstances.
- Structured review: the system proposes additions or revisions to the knowledge base, which a person must confirm or reject.
This combination is the paper's most useful contribution. AI is not presented as a machine that somehow extracts a person's entire expertise. It is a moderator: it creates a structured occasion for reflection, asks the next question, and helps turn dialogue into candidate knowledge that the organization can review.
The “why” appears in conversation
A blank page asks an expert to remember what a successor does not know. Dialogue can do something more active. It can respond to an answer, notice that a condition has not been explained, and ask for the reasoning underneath an action.
That makes conversation especially valuable for knowledge that is:
- triggered by a specific situation;
- expressed through examples and stories;
- distributed across small adjustments rather than one formal rule;
- difficult to recall away from the work environment; or
- so familiar to the expert that it no longer feels worth mentioning.
The paper also makes an important design distinction between retrieving existing explicit knowledge and eliciting new experiential knowledge. A conventional retrieval system can find an instruction that already exists. It cannot, by retrieval alone, reveal the undocumented exception an expert notices while carrying out the task.
Early evidence, with clear limits
The researchers conducted an exploratory evaluation in two phases: nine students tested general interaction and usability, followed by two crane-commissioning experts who assessed the system in its intended domain. Participants found the dialogue-based guidance supportive, described the process as logical and comprehensible, and reported that they still felt in command of the work.
Those findings are promising, not conclusive. The sample was small, the evaluation focused on one industrial use case, and it assessed perceptions rather than objective measures such as error rates, task completion time, cognitive load, or long-term knowledge retention.
The study also surfaced practical constraints. Industrial noise challenges speech recognition. A voice-only system cannot independently see whether an action was carried out correctly. Knowledge captured from dialogue still needs expert validation before it becomes part of the organizational record. And over-reliance on assistance can create a longer-term risk: people may complete tasks without developing the problem-solving competence the expert once held.
These limitations are not side notes. They define what responsible AI-supported knowledge transfer should look like: grounded in approved information, transparent about its boundaries, embedded in the real work process, and subject to human review.
Why this matters in Germany
The demographic context makes the research particularly relevant for German employers. Citing 2025 workforce data, the paper notes that roughly a quarter of Germany's workforce will reach retirement age within the next ten to fifteen years. At the same time, automation is reducing some of the routine opportunities through which less-experienced employees traditionally learned by doing.
Organizations therefore face two pressures at once: experienced workers are approaching retirement, and future workers may have fewer everyday chances to build the same experience. Waiting for a final-week handover document is unlikely to recover knowledge that took years of situated practice to develop.
The stronger response is to make knowledge capture part of work: create repeated opportunities to explain decisions, preserve examples and exceptions, and validate what is captured before the organization depends on it.
The Flamekeeper perspective
The paper supports three principles behind Flamekeeper's approach.
First, conversation is a better starting point than an empty template. Role-specific prompts and follow-up questions can help experienced employees articulate context they would not volunteer in a generic form.
Second, the first answer is rarely the complete answer. A description of an action may still lack its trigger, rationale, exception, owner, or escalation path. Reviewing the developing handover for those gaps creates a second chance to surface tacit knowledge.
Third, AI should structure and challenge the record, not become its unquestioned authority. The departing employee and manager remain responsible for confirming whether the captured knowledge is accurate, appropriate, and safe to reuse.
Flamekeeper operates in employee handovers rather than live crane commissioning, and it should not be read as a replication of KIMO. The shared insight is more fundamental: AI is most useful when it helps people say what experience has taught them—not when it pretends experience can be replaced.