Knowledge: Knowledge Retention
Capturing knowledge before it retires: knowledge transfer with AI
When an experienced specialist leaves, more than a folder of documents leaves with them. Organisations can capture experience-based knowledge methodically, hand it over in a structured way and make it usable for onboarding, and a clear line separates what proven methods deliver from where AI helps or does not.
Short answer
Knowledge transfer before retirement means capturing a person's experience-based knowledge in time, checking it, and passing it to successors. Methods such as expert interviews, shadowing, after-action reviews and knowledge maps do the capturing. AI speeds up transcription, structuring and querying the results, but replaces neither the conversation nor the expert's own check.
Definition
Knowledge transfer: Knowledge transfer is the planned handover of knowledge between people or organisational units, so that recipients can apply it independently. In a generational handover, it mainly concerns tacit experience-based knowledge that is not written down and can only be captured and passed on through conversation, observation and guided practice.
In the glossary: Knowledge management, Retrieval-augmented generation, Retrieval, Large language model, Hallucination, Human in the loop, Gold-standard test set, On-premise
01
How big is the wave of retirements in mid-sized companies, really?
According to the Federal Statistical Office (Destatis, 2026), around 13.3 million people in the workforce reach statutory retirement age by 2040, 30.0 percent of the 2025 workforce. The IfM Bonn (2025) estimates around 186,000 business handovers between 2026 and 2030. No official statistic captures how much experience-based knowledge is lost; that can only be assessed within an organisation itself.
The figures describe two separate movements. Destatis (2026) counts people in the workforce who will reach retirement age within the next 15 years, and notes that younger age groups do not replace the large birth cohorts in absolute numbers. The IfM Bonn (2025) estimates handovers because owners step down for personal reasons. In mid-sized companies, the two often coincide.
The most important number is missing from both sources: how much knowledge depends on a single person. How knowledge management is structured overall is explained in Knowledge management with AI. The focus here is the moment of handover itself.
02
What is the difference between explicit and tacit knowledge?
Explicit knowledge is written down, or can be: work instructions, drawings, parameter lists. Tacit knowledge sits in a person's skill and judgement, for instance in an ear for a machine running roughly, or a feel for which customer wants to be treated differently. It is hard to put into words, and is therefore captured differently from simply collecting documents.
Michael Polanyi described the core idea in 1966, in The Tacit Dimension: people know more than they can say. Someone who has done a job for decades describes the normal case when asked. The exceptions that make up their experience often only come to mind when they occur.
Ikujiro Nonaka described in 1994, in Organization Science, how organisations create knowledge through the interplay of both forms. In practice there are four transitions: working together (socialisation), telling and writing down (externalisation), combining documents (combination), and applying and practising (internalisation). Anyone who only collects documents is serving just one of these four transitions.
| Type of knowledge | Example | Where it sits | How to surface it |
|---|---|---|---|
| Explicit, documented | An inspection plan, a maintenance manual | Drives, wikis, ERP systems | Collect, check validity, make queryable |
| Explicit, but scattered | Reasoning in emails and tickets | Mailboxes, ticket systems | Analyse, condense, get it confirmed |
| Tacit, tellable | Why a supplier is approached differently for rush orders | In someone's head, in stories | Expert interview, the story method |
| Tacit, visible only in doing | Setting up a machine by ear | In movements and perception | Shadowing, pairing, guided practice |
| Knowledge about knowledge | Who helps with which question | In a person's network | A knowledge map, handing over responsibilities |
03
Which knowledge should a company capture first?
The knowledge that is critical to operations, depends on a few people, and will soon stop being available should come first. These three criteria, criticality, exclusivity and timing of departure, produce a ranking. Knowledge that is well documented or spread widely across the team can wait, even if a person who also holds it is leaving soon.
Take a machine-tool manufacturer with 150 employees (a hypothetical example): it has a foreman who is the only person who can set up a special machine, and a bookkeeper whose tasks a colleague already shares. The foreman ranks higher, even if the bookkeeper leaves sooner.
Decision path
Which capture method fits this area of knowledge?
Work through one area of knowledge at a time.
All questions and results as a list
- Can this knowledge basically be written down as a rule, a process, or a decision criterion?
- Yes, continue with: Do documents already exist that are scattered or probably outdated?
- Not really, continue with: Does the skill show mainly in doing, for example in movements, in sound, or in handling material?
- Do documents already exist that are scattered or probably outdated?
- Yes, Result: Collect the documents and have them checked
- No, Result: A structured expert interview
- Does the skill show mainly in doing, for example in movements, in sound, or in handling material?
- Yes, Result: Shadowing and guided practice
- No, continue with: Is it about judgement calls in rare situations, such as faults, complaints or exceptions?
- Is it about judgement calls in rare situations, such as faults, complaints or exceptions?
- Yes, Result: The story method and critical incidents
- No, the area is still unclear, Result: A knowledge map first
- Result: Collect the documents and have them checkedWorth discussing: which documents are valid, who will be responsible for them going forward, and which gaps the expert identifies while reviewing them.
- Result: A structured expert interviewWorth discussing: specific cases, decision criteria and warning signs. The expert checks the condensed result themselves.
- Result: Shadowing and guided practiceWorth discussing: which tasks the successor carries out themselves under guidance. Videos of the movements supplement the practice, they do not replace it.
- Result: The story method and critical incidentsWorth discussing: remembered situations where something went wrong or narrowly went right, and how the person spotted it early.
- Result: A knowledge map firstWorth discussing: which topics the person covers, who can stand in, and what is documented. Then work through each sub-area again.
04
Which methods can capture experience-based knowledge?
Five methods have proved themselves: the structured expert interview, shadowing at the workplace, after-action reviews of completed projects (lessons learned), the story method for rare critical incidents, and the knowledge map as an overview. They surface different kinds of knowledge and complement each other. None of them depends on AI; AI mainly changes the effort needed for note-taking and analysis.
Five capture methods compared
| Criterion | Expert interview | Shadowing | After-action review (lessons learned) | Story method | Knowledge map |
|---|---|---|---|---|---|
| Surfaces mainly | Tellable experience and criteria | Action knowledge visible only in doing | Insights from a completed project | Judgement in rare situations | Who knows what, and where it is missing |
| Output | A transcript, a condensed note, open questions | Observation notes, photos, short videos | A record with causes and recommendations | Cases with warning signs and responses | An overview of topics, people, documentation status |
| Where AI helps | Transcription, structuring, follow-up questions | Little, at most tagging | Summarising, comparing with earlier records | Transcription, pulling out signals | A draft from the org chart and document store |
| Limit | Delivers the normal case unless asked about specific cases | Rare cases go missing without the right occasion | Only works if recommendations have an owner | Memory is selective | Shows gaps, does not close them |
An after-action review asks about a completed project: what was planned, what happened, why did it deviate, what will be done differently in future? The story method asks for stories about specific situations instead of processes, because they contain the warning signs and trade-offs that no work instruction records. The knowledge map usually comes first, because it shows where the more time-consuming methods pay off.
05
How do you run an expert interview that delivers more than generalities?
A productive expert interview asks about specific cases rather than processes, about warning signs rather than rules, and about exceptions rather than the normal case. It is prepared, recorded or noted down, then condensed and approved by the expert. The most effective question is often: how do you notice something is wrong before others do?
Scope the area
One conversation covers one area of knowledge, not an entire career. Existing documents are reviewed beforehand, so the conversation starts at the gaps.
Clarify the framework
Purpose, use and storage location are discussed openly. The person sees and corrects the result before it is passed on.
Start with cases
"Tell me about the last time the machine gave you trouble" draws out more than "How is the machine maintained?"
Ask about signals and trade-offs
How do you spot that early? What did you rule out? What would a newcomer get wrong here?
Ask to see artefacts
Sticky notes, personal spreadsheets and marked-up drawings reveal which information actually gets used.
Condense, reflect back, approve
The transcript becomes a structured note with open questions. Only the version the expert has approved gets filed, with its area, owner and status.
- Which decision do you make today without thinking, that a newcomer would agonise over?
- Which mistake became costly here once, and what changed afterwards?
- If you could pass on only three things to your successor, what would they be?
06
What is a knowledge map, and how is it created?
A knowledge map is an overview of which areas of knowledge exist in a department, who has mastered them, who can stand in, and how well they are documented. It is created through conversation with managers and experts, often starting as a simple table. Its value lies in making gaps and single points of dependency visible before someone leaves.
The scope is what matters: an area of knowledge should be narrow enough that "Who else can do this besides person X?" has a clear answer. "Production" is too broad; "setting up the special machine for short runs" fits.
| Area of knowledge | Expert | Backup | Documentation | Action needed |
|---|---|---|---|---|
| Setting up the special machine | Foreman A | none | outdated manual | high: pairing and interview |
| A major customer's complaints | Sales rep B | Sales rep C, partly | Ticket history | medium: story method |
| Preparing the annual accounts | Bookkeeper D | Bookkeeper E | Checklist exists | low: review the checklist |
| Special customs clearances | Shipping clerk F | none | none | high: interview, hand over responsibilities |
The "backup" column is the most revealing: where it says "none", the area is a single point of dependency, regardless of when the person leaves. AI can propose a first draft from the org chart and the document store. That draft does not show who truly masters an area; only conversation clarifies that.
07
Where does AI concretely help with knowledge transfer?
AI helps in three places: it puts conversations into writing through speech recognition, it structures and condenses transcripts into drafts, and it makes checked content queryable through retrieval-augmented generation, with a reference to its source. The actual knowledge gain still happens in the conversation and in the expert's check, not in the model.
- 01ConversationInterview, story, after-action review
- 02TranscriptionSpeech recognition, on-premise too
- 03DraftStructure, key points, open questions
- 04ApprovalThe expert checks and corrects
- 05FilingArea, ownership, status
- 06QueryingAnswers with a source reference
Transcription. The Whisper paper (Radford et al., OpenAI, 2022) showed that a speech-recognition model trained on large, multilingual audio data works robustly across many recording conditions without further training. Such open models can also run on an organisation's own hardware. Proper names, model designations and numbers are still occasionally misheard, and belong in the check.
Structuring. A language model turns a long transcript into a structured note with key points, criteria, warning signs and open questions. The open questions prepare the next conversation. The draft stays a draft until the expert has read it.
Querying. Retrieval-augmented generation, described by Lewis and others in 2020, searches for matching passages and has the language model answer from them. The pipeline is explained in Retrieval-augmented generation explained; for scanned legacy records, see Understanding documents with AI.
08
Where does AI hit its limits in capturing experience-based knowledge?
AI can only process what has been captured. It cannot recognise which exception was missing from a conversation, it can produce content that sounds plausible but is wrong, and it cites outdated documents just as convincingly as valid ones. Judgement that only shows up in doing cannot be extracted from text, even with AI.
Four common assumptions
Only partly
Documents record what someone was able to write down. The feel for exceptions, and knowing who to ask, are usually missing, and that is exactly what gets lost when someone leaves.
Traces, at best
Emails show decisions, rarely reasons, and contain third-party data. A language model fills gaps plausibly. That is useful as a list of questions for a conversation, not as captured knowledge.
No
It answers questions about what is written down. New starters do not yet know what to ask, and action knowledge only comes from guided practice.
Started too late
Many critical situations only arise on specific occasions, such as the annual accounts. The successor needs the opportunity to experience them under guidance.
There is also an organisational limit: a retrieval system does not, on its own, distinguish between a valid and an outdated version of a manual. Every piece of filed knowledge therefore needs an owner and a visible status.
09
What does a structured handover before departure look like?
A structured handover combines a knowledge map, capture and guided practice into a plan for each area of knowledge. It does not end with filed documents, but with the successor having carried out typical and rare tasks themselves, under guidance. A checklist helps keep track of points that easily get lost in the final working phase.
Checklist
Handover checklist per departing employee
Only tick what is actually done. The list stores nothing.
10
How do new employees use captured knowledge during onboarding?
New employees make the best use of captured knowledge in three layers: guided onboarding with a named contact, a collection of checked content to read, and a search tool or assistant that answers questions with a source reference. The AI layer lowers the barrier to asking, but does not replace the person who explains how things connect.
Onboarding is the test of whether the transfer worked. Say a new service technician (a hypothetical example) asks an internal assistant why a product line is tested differently, and gets no sourced answer. That is not a failure, but a signal of a gap that should go to the owner of that area of knowledge.
- Guided onboarding: a named contact, a plan organised by area of knowledge, guided tasks of increasing difficulty.
- Checked content: approved notes, instructions and case descriptions with an owner and a status.
- A queryable knowledge base: answers with a source reference, so new starters learn where knowledge sits.
- A feedback channel: unanswered questions go to the responsible person.
Trust is built in two places: the assistant only shows what the person asking is allowed to see (see Access control in AI knowledge systems), and it makes clear which source an answer rests on, or that none was found.
11
Which roles does a knowledge transfer need, and who maintains it afterwards?
A knowledge transfer needs at least five roles: the expert, the successor, someone facilitating capture and condensing, a person with domain responsibility for each area of knowledge, and a manager who schedules time for it. With AI, operations and data protection are added. Without named owners for upkeep, captured knowledge goes stale before it is needed.
| Role | Task | Typical mistake |
|---|---|---|
| Expert | Tells, shows, approves drafts | Never sees the result |
| Successor | Learns, practises under guidance, asks questions | Named only at the very end |
| Facilitator | Prepares conversations, runs them, condenses | Only asks about the process |
| Domain owner per area | Maintains content, closes reported gaps | Missing, content goes stale unnoticed |
| Manager | Prioritises areas, creates time | Expects transfer alongside a full workload |
| IT and operations | Provides transcription, filing and search | Picks tools before the process is clear |
| Data protection and works council | Clarifies recording, storage, deletion | Consulted only after recordings already exist |
ISO 30401:2018 offers a reference framework for ongoing operation. The standard sets requirements for a knowledge management system and, by its own account, applies to organisations of any type and size. A single handover does not need it. The connection to the whole knowledge lifecycle is described in Knowledge management with AI; implementing a knowledge assistant is shown in AI knowledge management.
Read more on iiterate.de
Foundations in the Knowledge section
- Knowledge Knowledge management with AI The lifecycle from capture to upkeep, with the right AI technique for each phase.
- Knowledge Retrieval-augmented generation explained How checked content becomes a queryable knowledge base with source citations.
- Knowledge Checking AI answers How test questions and spot checks make answer quality measurable.
- Knowledge Access control in AI knowledge systems How access rights carry through into an assistant's answer.
Implementation and self-test
- Consulting AI knowledge management A knowledge assistant that answers from an organisation's own documents and names its source.
- Consulting RAG implementation Implementing a RAG application on company data, including on an organisation's own infrastructure.
- Tool RAG readiness check A self-test of whether the data and the organisation are ready for a queryable knowledge base.
- Glossary Knowledge management in the glossary The short definition of the term, with related entries.
Further reading
- Signal Introducing AI in mid-sized companies A five-phase road map that starts with a measurable problem.
- Signal What a RAG system costs to run Index upkeep and evaluation as ongoing costs after the build.
- Signal Personal knowledge management with AI Where the line falls between local notes and rented inference.
- R&D Lab iiterate Handbook A topic turned into a bilingual learning handbook with sources, comprehension checks and a completion record.
Frequently asked questions
When should knowledge transfer start before a retirement?
Early enough that the successor can carry out the important tasks themselves, under guidance, before the experienced person leaves. What matters is the rarest critical occasion in that area of knowledge, such as the annual accounts, an audit or seasonal maintenance. If that occasion falls after the last working day, there is not enough time left, no matter how well things were documented.
May conversations with experts be recorded and transcribed by machine?
That is a matter of data protection and, where applicable, co-determination, to be settled with the relevant departments in an organisation. It makes sense from a methods perspective to settle purpose, storage location, access and deletion with the relevant departments before the first conversation. Transcription can technically run on-premise. An overview is given in GDPR and AI.
What can be done if the person has already left?
Then reconstruction from traces is what remains: change histories, tickets, records, personal spreadsheets and colleagues' memories. AI can summarise such material and surface contradictions. The result is hypotheses, not established facts. Every statement gets marked as reconstructed and is confirmed or corrected the next time it comes up in real use.
What conditions make it easier for experienced specialists to share their knowledge?
Three things help methodically: scheduled working time instead of transfer as an afterthought, a conversation format that asks about experience rather than testing knowledge, and a commitment that the person sees every result before it is passed on. People who see their experience carefully prepared tend to share more. Personnel-policy questions do not belong in this method.
How can you check whether the knowledge transfer succeeded?
Most reliably with real tasks: the successor solves typical and rare cases themselves, while the experienced person only observes. For a knowledge base, a collection of test questions with known correct answers also helps, run before approval and after every change. How to build one is described in Checking AI answers.
Can a company outsource knowledge transfer entirely to a service provider?
Parts of it, yes: facilitation, transcription, building a queryable knowledge base and its technical operation can be handed to an external provider. What cannot be outsourced is prioritising areas of knowledge, the experts' time, approving content, and domain responsibility for upkeep. These roles stay inside the company, or the captured knowledge ends up belonging to nobody.
Read on
Related topics
Sources
- 01 13,3 Millionen Erwerbspersonen erreichen in den nächsten 15 Jahren das gesetzliche Rentenalter Statistisches Bundesamt (Destatis), 2026-06-23 · destatis.de
- 02 Unternehmensnachfolgen in Deutschland 2026 bis 2030, Daten und Fakten Nr. 37 Institut für Mittelstandsforschung Bonn (IfM Bonn), 2025-11 · ifm-bonn.org
- 03 The Tacit Dimension (Michael Polanyi) University of Chicago Press, 1966 · press.uchicago.edu
- 04 A Dynamic Theory of Organizational Knowledge Creation (Ikujiro Nonaka), Organization Science 5(1), 14-37 INFORMS, 1994 · doi.org
- 05 ISO 30401:2018 Knowledge management systems, Requirements International Organization for Standardization (ISO), 2018 · iso.org
- 06 Robust Speech Recognition via Large-Scale Weak Supervision (Radford et al.) arXiv, 2022-12-06 · arxiv.org
- 07 Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks (Lewis et al.) arXiv, NeurIPS 2020, 2020-05-22 · arxiv.org
- 08 Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile (NIST AI 600-1) National Institute of Standards and Technology (NIST), 2024-07-26 · doi.org
Let us talk about your project
Whether it is a prototype, an internal tool or an AI application: describe briefly what you are building or want to take into production.