You Cannot Upload What Your Experts Never Wrote Down

Scope note: This essay considers observed work, structured expert interviews, corrections to AI output, workflow knowledge graphs, and expert-learning models as complementary ways to capture tacit knowledge. It does not claim that an AI system can reproduce a person, embodied skill, or complete professional judgment.

An expert leaves. The organization keeps the manuals, the process map, the training deck, and the folder called “final.”

The knowledge still walks out.

What disappears is rarely the official sequence. It is the pause before step six. The exception that looks harmless. The instrument reading that passes the system check and fails the expert’s eye. The reason one customer, sample, machine, or market needs a different route.

That knowledge was never waiting in a document for an AI model to retrieve. It lived in attention, correction, memory, and practiced judgment. A new group of preprints asks whether AI can help capture it. Their combined answer is promising and properly inconvenient: no single capture method is enough.

The document is the floor, not the expert

Tacit knowledge is knowledge people use without fully stating it. The term can sound mystical. The examples are ordinary.

A laboratory scientist knows which “successful” automated run is scientifically invalid. An engineer notices a vibration that the threshold allows. A product manager changes the order of a conversation because the customer’s first answer altered the risk.

The procedure records the normal route. Expertise manages the boundary around it.

Expert Mind, proposed by Diego Ezequiel Cervera for the energy sector, combines interviews, multimodal capture, retrieval, and language models to preserve knowledge from departing specialists. The architecture is experimental, not proof that an expert can be copied. Its premise is sound: text alone cannot carry knowledge expressed through demonstrations, diagrams, stories, physical settings, and exception cases.

The archive needs more than another upload control.

Ask about failure, not only procedure

The strongest applied paper in this set comes from pharmaceutical research.

In Federated Semantic Knowledge Graphs for Laboratory Workflows, Luis Schachner and his coauthors describe a system deployed in Genentech’s Biochemical and Cellular Pharmacology department. An AI interview agent asks experts structured questions about decisions, confidence, exceptions, and failure conditions. The system converts those answers into connected graphs covering program milestones, assay procedures, and physical laboratory infrastructure.

The important result is not that the graph can repeat a protocol. Existing systems already store protocols.

The combined graph exposed what the authors call automation-masked silent failures: cases where the execution log reports success while the scientific result is no longer valid. That relationship was missing from the protocol, the machine log, and the existing ontology when each source stood alone.

The expert did not merely supply a fact. The expert supplied the condition under which another system’s fact should not be trusted.

Watching work captures action, not reason

Interviews have limits. People forget routine steps. They rationalize decisions after the fact. They omit small actions precisely because those actions feel obvious.

cotomi Act, by Masafumi Oyamada and his coauthors, takes the opposite route. A browser agent watches a person perform work and progressively turns the observed behavior into shared task boards and wiki records that both the user and agent can edit.

That approach can capture sequence, repetition, tool choice, and visible correction without asking the worker to narrate every movement. It may reveal the actual procedure rather than the official one.

It cannot, by observation alone, know why the person hesitated, what alternative they rejected, or which invisible condition changed the decision. A recorded click is evidence of action. It is not yet evidence of judgment.

This is the central split in tacit-knowledge capture. Observation finds what people do. Elicitation asks what they notice. Neither should impersonate the other.

Corrections are compressed expertise

A third route begins after the AI produces something wrong.

In Context-Mediated Domain Adaptation, Anton Wolter and his coauthors treat expert edits to AI-generated material as implicit specifications. When an expert changes terminology, structure, emphasis, or relationships, the system uses the correction to shape later reasoning instead of treating it as a one-time cleanup.

This is attractive because experts often find it easier to correct a concrete artifact than to describe every rule in advance. The wrong draft creates a surface against which judgment becomes visible.

But an edit still needs interpretation. A person may change a phrase for accuracy, policy, tone, audience, or taste. The system should not turn every revision into a universal rule. A correction becomes reusable knowledge only after its scope is known.

The smallest edit may contain deep expertise. It may also contain a preference. Capture requires restraint.

Formal structure makes the knowledge testable

Tacit Knowledge Extraction via Logic Augmented Generation and Active Inference tries to move captured expertise into a form that machines can query and validate. Lorenzo Lamazzi and his coauthors combine language models with explicit logical and ontological structures. The goal is not just to produce a fluent summary. It is to represent assumptions, constraints, decisions, and relationships so another system can test them.

AI Expert Twin, by Annie Yuan and her coauthors, adds another necessary layer. Their framework models actions and concepts alongside values, uncertainty, and trade-offs. That matters because expert judgment is not always a hidden rule waiting to be extracted. Sometimes two valid goals conflict, and the expert knows which loss the situation can bear.

These are frameworks rather than evidence that a complete expert representation exists. Their value is in refusing the easy reduction. Expertise is not a large bag of tips.

Capture should remain a relationship

My view is that organizations should stop treating tacit-knowledge capture as an extraction project. Extraction implies that the knowledge sits inside a person like ore and becomes complete once removed.

A better system would combine five modes:

1. Documents establish the official procedure and vocabulary.
2. Observation records actual sequence, tools, and visible corrections.
3. Structured interviews surface exceptions, weak signals, and reasons.
4. Artifact review turns expert corrections into candidate rules with explicit scope.
5. Formal representation makes the resulting claims traceable, testable, and revisable.

The expert must be able to inspect the record, lower its confidence, restrict its audience, correct its meaning, and withdraw it. Provenance should name the person and the transformation without pretending that the graph has become the person.

Preserve the judgment, not the ghost

AI may make tacit-knowledge work much more practical. It can watch patiently, ask consistent questions, connect fragments, and preserve the source beside the claim. Those are real gains.

The danger is not only a bad answer. It is a false sense of completion.

An expert model can preserve a route through past decisions. It cannot guarantee the next decision under conditions nobody has seen. The organization still needs living practice, apprenticeship, review, and people who can refuse the stored rule.

You cannot upload what the expert never wrote down. You can build a careful process that helps the expert make part of it visible.

Part of it. That limit belongs in the system.