Knowledge extraction: how to capture what your best experts know
Summarize this article with AI
Tacit vs. explicit knowledge: why your internal procedures fall short
Your company has procedures. Manuals. Internal wikis. That’s explicit knowledge. The stuff that’s written down somewhere.
But the knowledge that actually makes the difference lives elsewhere. It’s tacit knowledge.
It’s your top salesperson’s reflex to detect in 30 seconds whether a prospect is serious. Your support manager’s intuition that a certain type of question masks a deeper problem. Your CTO’s ability to estimate project complexity by scanning the first 3 lines of the brief.
This knowledge has one key trait: the person who possesses it often struggles to explain it. They « feel » things. They « know ». But ask them to formalize their method, and you’ll get a simplified version that captures maybe 20% of what they actually do.
That’s exactly why internal procedures are insufficient. They capture the what. Knowledge extraction captures the how and the why.
The true cost of knowledge loss
When an expert leaves your company, you lose much more than a team member.
During those ramp-up months:
- Prospects get less qualified. Conversion rates drop.
- Customers receive lower quality service. Satisfaction declines.
- Technical decisions slow down. Projects fall behind schedule.
- New hires are trained less effectively. The cycle repeats.
Hypothetical example: a senior salesperson carries €2M in annual revenue. They leave. Their replacement takes 8 months to hit 80% of their performance. The revenue gap? Hundreds of thousands of euros.
Knowledge extraction takes a few weeks. It protects years of accumulated expertise.
Extraction techniques that work
Knowledge extraction isn’t a casual interview. It’s a structured process rooted in cognitive science. Proven techniques.
Ask the expert to recount critical situations they’ve handled. Complex cases. Edge cases. Moments when their expertise made the difference. This is where tacit knowledge emerges most naturally.
The expert works through a real case while verbalizing each step of their reasoning. « I look at X first, because if I see Y, it means Z. » This reveals the automatic patterns and mental shortcuts the expert uses unconsciously.
Map out the expert’s decision trees. At each choice point: which criteria? Which thresholds? Which exceptions? This produces decision diagrams directly usable by an AI agent.
Present varied scénarios to the expert — from straightforward to complex — and observe how their approach shifts. Edge cases are most revealing: that’s where expertise diverges from standard procedure.
How an extraction session works
A typical session runs 60 to 90 minutes. Shorter and you stay surface-level. Longer and concentration fades, diminishing transcript quality.
Before the session
The extractor prepares a domain-specific interview guide. They identify priority use cases, documented critical situations, and areas where knowledge is most concentrated.
During the session
The interview is recorded (audio) and transcribed in real-time. The extractor alternates between CIT and think-aloud techniques. They probe for details: « And there, how do you know it’s the right moment? » « What makes you choose that approach over this one? »
Key questions:
- « Tell me about a recent case where you made the difference. »
- « If you had to train your replacement in 1 hour, what would you tell them? »
- « What signals do you detect that junior staff miss? »
- « What’s the most common mistake you see from newcomers? »
After the session
The transcript is cleaned, structured, and enriched with the extractor’s notes. Raw verbatims are preserved for reference. Insights are categorized: processes, decisions, exceptions, metrics, illustrative anecdotes.
The deliverable: what a knowledge base contains
After extraction and structuring, the deliverable is a version-controlled Markdown knowledge base of 50 to 200 pages per expert. It contains:
- Documented processes — each step with its variations and conditions
- Decision trees — the selection criteria at each branch point
- Scripts and templates — the exact phrasings that work (emails, calls, presentations)
- Objection responses — each common objection with 2 to 3 tested replies
- Edge cases — exceptional situations and how to handle them
- Reference metrics — thresholds, internal benchmarks, success indicators
- Glossary — domain-specific vocabulary with its nuances
This Markdown format isn’t arbitrary. It feeds directly into an AI agent’s context. Each module becomes an operational knowledge base. No conversion. No reformatting. The language model ingests and executes.
Realistic timeline: how long it takes
The complete process for one expert follows this pace:
Identify priority domains, prepare interview guides, schedule sessions with the expert.
3 to 5 sessions of 90 minutes, spaced 2 to 3 days apart. This pace gives the expert time to reflect between sessions — often the best insights arrive at the next session: « I was thinking about what we discussed, actually… »
Verbatims transform into structured modules. The expert validates each module in a 60-minute review session. Corrections and additions get integrated.
Total: 3 to 4 weeks per expert. Their time investment? 6 to 8 hours. The rest is borne by the extraction team.
First deployment with 2 to 3 experts: 6 to 8 weeks. Extractions overlap.
Concrete example: extracting a VP of Sales
A real-world example, from the field.
VP of Sales, B2B SaaS, 12 years expérience. 180 prospects per month. 34% conversion rate. His peers? 18%.
What extraction revealed
Session 1 (CIT) — 5 complex qualification situations. A pattern emerges: he always asks the same second question (« What’s your current process? »), but interprets the answer using 4 simultaneous criteria.
Session 2 (Think-aloud) — 3 inbound requests handled live. He checks the prospect’s LinkedIn profile before reading the form. He adapts his approach based on seniority in under 10 seconds.
Session 3 (Decision Analysis) — His qualification decision tree has 23 nodes. Some counter-intuitive branches: a prospect who says « we’re in a rush » gets low priority (signal of a panicked buyer comparing 10 solutions).
The deliverable
87 pages of Markdown. 4 decision trees. 12 conditional email scripts. 31 objection responses with profile-based variants. 8 qualification signals his peers didn’t know about.
This deliverable fuels an AI qualification agent that pre-processes every inbound prospect using this VP’s method.
Ready to capture your team’s expertise?
Discover how knowledge extraction powers an AI competency center tailored to your business.
Explore the AI competency centerFrequently asked questions
Does the expert need to invest a lot of time?
Plan 6 to 8 hours total, spread over 3 to 4 weeks. Sessions last 60 to 90 minutes. The rest of the process (structuring, formatting) is handled by the extraction team.
What if the expert struggles to articulate their expertise?
That's normal — and exactly why extraction techniques exist. Think-aloud protocol and Critical Incident Technique are designed to surface tacit knowledge the expert uses unconsciously. The extractor guides the process.
Is the deliverable directly usable by an AI?
Yes. The base is structured Markdown, designed to be injected into a language model's context. Each module (decision tree, script, objections) is a standalone block an AI agent can use directly.
Can we extract knowledge from multiple experts on the same domain?
Absolutely. It's even recommended. Extracting 2 to 3 experts on the same domain lets you cross-reference approaches, identify shared best practices, and build a more robust base. Disagreements between experts get documented too.
How do we protect confidentiality of extracted information?
The knowledge base remains your company's property. It's stored in your environment. Access is controlled. A confidentiality agreement covers the entire extraction process.
