How Can Experts Make Their Knowledge Available to AI?

Model knowledge is not expert knowledge

A general language model learns broad patterns from large training datasets. It may know common terminology and produce plausible explanations, but that knowledge is not the same as the developed judgement of a particular doctor, engineer, lawyer, teacher, musician or researcher.

Professional expertise includes ways of framing a problem, deciding which details matter, recognising exceptions, weighing evidence and knowing when a conclusion is unsafe. Some of this is written in articles or notes. Much of it remains tacit: it appears in examples, corrections, case discussions and the expert’s reactions to unusual situations.

Start by capturing the structure of judgement

Uploading a folder of documents may make text searchable, but it does not automatically reveal how the expert connects ideas. A useful process combines existing material with deliberate knowledge elicitation.

• Collect the expert’s authoritative documents, examples, frameworks and source references.
• Identify recurring questions and the distinctions the expert uses to answer them.
• Ask the expert to explain difficult cases, trade-offs, failure modes and exceptions.
• Record terminology and keep important concepts stable across the knowledge base.
• Separate enduring principles from information that changes with time or jurisdiction.

The objective is not to remove the expert’s complexity. It is to make that complexity navigable enough for an AI system to retrieve the right material at the right point.

Structure knowledge around use, not file type

Documents are containers. An AI knowledge system needs smaller units with relationships and context. Depending on the field, these may include concepts, claims, procedures, cases, diagnostic questions, evidence, counterexamples, definitions and explicit boundaries.

Each unit should retain provenance: who created it, which source supports it, when it was reviewed and what scope it covers. Attribution allows the system to distinguish the expert’s position from a general model answer or an external source.

Structure does not require forcing every idea into a rigid template. Some knowledge is best expressed as a narrative explanation; some as a decision process; some as a worked example. The structure should follow the way the knowledge is used.

Retrieval makes knowledge available at the right moment

In many systems, a general model remains responsible for language and reasoning while a retrieval layer selects relevant expert material for the current question. The model can then answer with access to that material rather than relying only on what it learned during training.

Retrieval quality depends on more than text similarity. The system may need to consider the user’s role, the stage of a process, jurisdiction, version, concept relationships and source authority. A smaller body of carefully structured knowledge can outperform a larger but poorly organised archive.

Expert knowledge needs ownership and maintenance

A knowledge system becomes unreliable when no one knows which material is current. Experts should be able to review, correct, replace and retire content. Changes should preserve version history where the difference matters.

The AI should also be able to reveal uncertainty and source boundaries. If the knowledge base does not support a conclusion, the system should say so rather than fill the gap with a confident generalisation.

Related answers