When Models Embody Expertise
What It Means for AI to “Capture” Expert Knowledge
Organizations want to “capture expert knowledge in AI” - to preserve departing expertise, scale scarce specialist capacity, or democratize elite skills.
This goal is more achievable for some kinds of expertise than others. Understanding the distinction helps set realistic expectations.
What Models Can Capture
AI models effectively capture:
Pattern recognition expertise. The experienced radiologist’s ability to spot abnormalities. The expert trader’s recognition of market patterns. When expertise is about recognizing patterns in data, models can learn those patterns.
Explicit decision rules. The underwriter’s documented criteria. The diagnostic flowchart. When expertise has been codified into rules, models can apply those rules (and potentially improve them).
Statistical regularities. The intuition about what works based on experience. When that intuition reflects patterns in historical data, models can learn the patterns.
What Models Struggle to Capture
AI models struggle with:
Tacit judgment. The expert’s sense of when standard rules don’t apply. The intuition that something is wrong before being able to say what. This knowledge resists articulation and therefore resists training.
Context sensitivity. The ability to recognize when the current situation differs from past patterns in ways that matter. Models extrapolate; experts know when not to.
Social knowledge. How to communicate with patients, clients, colleagues. What matters to this stakeholder. How to navigate organizational dynamics.
Adaptive expertise. The ability to handle novel situations by reasoning from principles. Models pattern-match; experts can go beyond patterns.
The Partial Capture Problem
Most expertise is a mix. The skilled diagnostician combines pattern recognition (capturable) with patient communication (not) with judgment about edge cases (partially).
Attempts to capture expertise in models often:
- Succeed at the pattern recognition component
- Fail at the tacit judgment component
- Miss the social and adaptive components entirely
The result: a model that handles typical cases but fails at the cases where expertise was most valuable.
Realistic Expectations
Setting appropriate expectations:
Triage and screening. Models can identify cases that need expert attention, extending expert reach without replacing judgment.
Decision support. Models can surface relevant information and suggest considerations, while humans provide judgment.
Training acceleration. Models can help novices learn pattern recognition faster, compressed experience without replacing mentorship.
Documentation and consistency. Models can ensure standard approaches are applied consistently, catching oversights.
Designing for the Limits
Given these constraints:
Keep experts in the loop for judgment. Automate pattern recognition; preserve human judgment where it matters.
Design for the edge cases. The model handles the typical; humans handle the atypical. Design workflows that route appropriately.
Capture expertise iteratively. Don’t try to capture everything at once. Start with explicit knowledge, add refinement as gaps become visible.
Maintain expertise. If models handle routine work, experts may lose practice. Design for ongoing skill maintenance.
The Succession Problem
The desire to “capture expert knowledge before they leave” faces a fundamental challenge: the most valuable expertise is often the hardest to capture.
What can be captured quickly - documented procedures, explicit rules - is valuable but not irreplaceable.
What makes an expert irreplaceable - tacit judgment, adaptive capability, relational skills - is exactly what models capture poorly.
The honest answer: AI can capture some of what experts know, extend their capacity, and assist their successors. It cannot replace what made them experts.
What kinds of expertise in your organization does AI capture well? What kinds resist capture? How does that affect your AI strategy?
Related: 06-molecule—tacit-vs-explicit-knowledge, 06-atom—tacit-knowledge