The most valuable knowledge in most organisations isn't in documents. It's in the heads of senior staff, in undocumented processes, in "the way we've always done it." When those people leave, the knowledge goes with them. AI can help, but knowledge extraction is harder than it looks, and most approaches get it wrong.
What You Need to Know
- Tacit knowledge (know-how that hasn't been documented) is the biggest knowledge risk in most enterprises. It's also the hardest to extract.
- AI is excellent at processing documented knowledge but can't extract knowledge that doesn't exist in any recorded form. The extraction step is human work. AI helps structure, validate, and make that knowledge accessible.
- The biggest mistake is treating knowledge extraction as a one-time project. It's an ongoing process. Knowledge changes. People leave and join. New edge cases emerge.
- Interview-based extraction, enhanced by AI structuring, is the most effective pattern we've found. Record experts explaining their decision-making, then use AI to structure, index, and make the knowledge searchable.
42%
of enterprise knowledge is tacit (undocumented), according to knowledge management research
Source: Nonaka and Takeuchi, updated estimates, 2024
The Scale of the Problem
Every organisation we work with has a version of this story: a senior claims handler who's been there 20 years and can spot a fraudulent claim by gut feel. A compliance officer who knows which regulatory interpretations apply in which situations. An engineer who knows why the system was built this way, because they built it.
When these people leave, the knowledge gap is immediate and expensive. Training replacements takes months or years. Some knowledge is simply lost.
The conventional solution is documentation. Write it down. Build a knowledge base. Create standard operating procedures. This helps, but it captures perhaps 60% of what an expert knows. The remaining 40%, the judgement calls, the pattern recognition, the "I wouldn't do that because last time it caused X", is the hardest to document and the most valuable.
Where AI Helps (And Where It Doesn't)
Where AI Helps
Structuring extracted knowledge. Once knowledge is captured in some form (interview transcripts, recordings, rough notes), AI can structure it into searchable, usable formats. Converting a 90-minute interview with a claims expert into a structured knowledge article with categorised scenarios, decision trees, and key rules.
Identifying gaps. AI can analyse existing documentation and flag areas where the coverage is thin. "You have detailed procedures for standard claims, but nothing for contested claims above $50K." This gap analysis guides further extraction.
Making knowledge accessible. A well-built RAG system on structured knowledge makes it available in context, when a less experienced staff member is handling a case similar to one the expert would handle differently.
Keeping knowledge current. AI monitoring can flag when documented knowledge might be outdated: regulatory changes, process updates, or new patterns in incoming data that don't match the documented procedures.
Where AI Doesn't Help
The actual extraction. AI can't interview experts. It can't observe undocumented processes. It can't identify what it doesn't know. The human facilitation of knowledge extraction is irreplaceable.
Judgement and context. An expert's "gut feel" is usually pattern recognition built over years of experience. AI can be trained to replicate some of these patterns, but the expert needs to articulate them first. "I look at the claim amount, the policy type, the claimant's history, and whether the timing seems right" is extractable. "It just feels wrong" needs more work.
Organisational politics. Knowledge is power. Some experts resist extraction because their knowledge is their job security. This is a change management problem, not a technology problem.
A Practical Extraction Process
Here's the process we've refined across several enterprise knowledge extraction projects:
Phase 1: Identify Critical Knowledge
Map the knowledge that matters most. Where are the single points of failure? Which roles have the longest ramp-up time for new hires? Which processes depend most heavily on individual expertise?
Prioritise by risk (what happens if this person leaves?) and value (how often is this knowledge needed?).
Phase 2: Structured Interviews
Conduct recorded interviews with subject matter experts. Not free-form conversations. Structured interviews with scenario-based questions:
- "Walk me through how you'd handle [specific scenario]."
- "What would you check first, and why?"
- "What's the most common mistake someone new to this role makes?"
- "When do the standard procedures not apply?"
The scenarios come from the process mapping in Phase 1. The interviews are typically 60-90 minutes, recorded with consent.
Phase 3: AI-Assisted Structuring
Process the interview transcripts using AI to extract:
- Decision rules and criteria
- Scenario-specific procedures
- Common pitfalls and exceptions
- References to other processes or data sources
The AI produces structured drafts. Human editors (ideally the original expert) review and refine them.
Phase 4: Integration into Workflows
The structured knowledge feeds into:
- A searchable knowledge base (RAG-enabled)
- Training materials for new staff
- AI-assisted tools that surface relevant knowledge in context
- Decision support systems that apply the extracted rules
Phase 5: Continuous Maintenance
Knowledge extraction isn't a project. It's a capability. New knowledge emerges. Existing knowledge becomes outdated. Regular review cycles (quarterly for high-priority knowledge) keep the knowledge base current.
Start Here
Identify the three people in your organisation whose departure would cause the most operational disruption. Schedule structured interviews with each of them this quarter. Even without AI processing, having their expertise recorded is a form of risk mitigation.
The technology for making knowledge accessible is mature - the bottleneck is getting the knowledge out of people's heads and into a form the technology can work with. That's a human process, and it requires deliberate investment, not just an AI tool.
Mak Khan
Chief AI Officer