capinorium · Article · Strategy & SMEs
Securing experience-based knowledge before the baby boomers retire: what an AI knowledge system can do

A mid-sized company does not secure experience-based knowledge with AI alone, but in three steps: first, people capture the knowledge methodically; then the experts approve the results; and only after that does an AI-supported knowledge system make the verified content findable, with access rights. The technology makes finding and passing on knowledge easier. It does not replace the work of drawing tacit knowledge out of people's heads.
The time pressure is real. According to IW short report 39/2026 from the German Economic Institute, Germany's potential labour force will shrink by around 4.3 million people by 2036, because around 1.3 million reach retirement age each year and only around 800,000 enter (IW, 15 June 2026). Owners are ageing too: 57 per cent are 55 or older, and by the end of 2029 around 545,000 mid-sized companies are planning a succession (KfW Research, Fokus No. 526, 9 January 2026). The DIHK skilled labour report 2025/2026 explicitly cites the loss of company-specific knowledge as a risk. How much knowledge is actually lost, however, has not yet been reliably quantified. All the more important is the question of what role an AI system can realistically play.
Capture first: tacit knowledge is not in any document
The most valuable experience-based knowledge is rarely written down. The Fraunhofer IPK guide describes routines that run "unconsciously – out of routine" and only become visible through targeted questions. Expertise, it says, cannot be transferred technocratically.
The institute therefore proposes a moderated approach. In transfer conversations organised by topic, the expert talks, a moderator from outside the field asks questions, and the successor reflects. Knowledge and task maps, a network of contacts and a prioritised transfer plan record the results. The IPK product sheet names a generous handover period with overlapping staffing as the most reliable route. These methods date from 2016 but are not outdated: language models have not changed the problem of making tacit knowledge visible.
In its 2026 white paper, the Plattform Lernende Systeme confirms this limit: generative AI has difficulty with practically acquired, tacit knowledge. AI can nonetheless lighten the load here, for example by transcribing recorded conversations and pre-structuring drafts of rules of thumb or FAQs. Asking the follow-up questions remains a human task.
Then approve: the experts check what the system will later say
Before content goes into a knowledge system, it should be approved by the person whose knowledge it reflects. That sounds like bureaucracy, but it is the only point at which subject-matter errors can still be fixed cheaply. Approval also clarifies which statements only apply under certain conditions, for example to a particular machine or customer.
For project planning, this means the scarce resource is the time of the departing specialists. Those who only start in the final year of employment will hardly be able to capture and review all topics. The transfer plan should therefore define which knowledge is secured first and who is responsible for approval.
Only then into a RAG system with access rights
Retrieval-augmented generation (RAG) is suitable for accessing the verified content. The system searches its own documents for passages relevant to a question and has a language model formulate an answer from them. The model is not trained on the company data for this. Content can therefore be corrected or removed individually. Such applications are spreading: according to Bitkom, 22 per cent of German companies with 20 or more employees use AI in internal knowledge management, twice as many as in the previous year. Users cite data protection as the biggest barrier.
The Plattform Lernende Systeme recommends enforcing permissions via roles and policies directly at the retrieval stage. Researchers from Microsoft argue similarly: before loading content into the model context, it should be checked whether all participants are authorised (Bhatt et al., arXiv 2509.14608).
RAG does not eliminate errors. A Stanford study found hallucination rates of 17 to 33 per cent in commercial legal research tools using RAG (Magesh et al., 2024). The figure relates to legal research and cannot be transferred directly to technical knowledge bases. But it shows that even systems with source references give wrong answers. According to the Plattform Lernende Systeme, hallucinations are not easy to detect automatically. It advises validation loops and a combination of human and machine review. This includes test questions from your own operations, source references in every answer and a person with subject-matter responsibility for corrections.
Works council and employees belong at the start
A company's own knowledge system typically logs who asked what. Under the view widely held in German legal literature, the mere objective suitability for monitoring is sufficient to trigger co-determination under Section 87(1) No. 6 of the Works Constitution Act (BetrVG), even without any such intention (CMS). A decision by the Hamburg Labour Court, which saw no co-determination for privately used ChatGPT, cannot readily be applied to in-house systems with access to usage data (Gleiss Lutz, 27 March 2024). This is an assessment, not legal advice; individual cases should be reviewed by specialists.
Regardless of the legal position, success depends on the experts' willingness. The IPK explicitly provides for preliminary conversations for this. Anyone wanting to secure experience-based knowledge should therefore plan capture, approval and technology as one joint project and build the knowledge system as the last step. If you would like to assess whether and in what form such a system makes sense for your company, an assessment with capinori can be a suitable first step.
Frequently asked questions about experience-based knowledge and AI knowledge systems
How does a mid-sized company secure experience-based knowledge before retirement?
A mid-sized company secures experience-based knowledge most reliably with moderated transfer conversations, knowledge maps, a transfer plan and, where possible, overlapping staffing. These methods are described by Fraunhofer IPK. An AI knowledge system can then make the captured and expert-approved knowledge findable, but does not replace the capture.
Can AI capture tacit experience-based knowledge on its own?
No, AI cannot capture tacit experience-based knowledge on its own, because it lies in routines and judgements and is not in any document. The Plattform Lernende Systeme notes that generative AI has difficulty with this. AI can, however, transcribe conversation recordings and pre-structure drafts that experts then review.
What is retrieval-augmented generation (RAG) in a knowledge system?
Retrieval-augmented generation (RAG) is a method in which a system searches its own documents for passages relevant to a question and a language model formulates an answer from them. The language model is not trained on the company data. Content in a RAG system can therefore be corrected, removed or given access rights individually.
Does retrieval-augmented generation (RAG) prevent wrong answers?
No, retrieval-augmented generation (RAG) reduces the risk of wrong answers but does not prevent it. A Stanford study found error rates of 17 to 33 per cent in legal RAG tools; the figure applies to legal research, not directly to technical knowledge. Source references, test questions from your own operations and human review are needed.
Must the works council be involved in an AI knowledge system?
For an in-house AI knowledge system that logs usage data, under the widely held view in Germany the works council must generally be involved under Section 87(1) No. 6 of the Works Constitution Act, because objective suitability for monitoring is already sufficient (CMS). This is not legal advice; the specific case should be reviewed by an expert.
How widespread is AI in knowledge management at German companies?
According to Bitkom, 22 per cent of companies with 20 or more employees use AI in internal knowledge management, up from 11 per cent the previous year. The survey of 603 companies took place in summer 2026. AI users cite data protection as the biggest barrier. The figure says nothing about micro-enterprises.
Sources (11)
- German Economic Institute (IW): Deschermeier, P.; Schäfer, H.: Die Babyboomer erreichen das Rentenalter, IW-Kurzbericht 39/2026 (German). Study page; press release of 15 June 2026. Website
- KfW Research: Nachfolge-Monitoring Mittelstand 2025, Fokus Volkswirtschaft No. 526, 9 January 2026 (Dr Michael Schwartz) (German). PDF
- DIHK: Fachkräftereport 2025/2026, 19 December 2025 (German). Website
- Fraunhofer IPK: Wissen erfolgreich bewahren. Ein Leitfaden zum systematisch begleiteten Wechsel von Fach- und Führungskräften, 2016 (German). PDF
- Fraunhofer IPK: product sheet on knowledge transfer and securing experience, undated (German). PDF
- Plattform Lernende Systeme: Generative KI verantwortungsvoll einsetzen. Impulse für Unternehmen und Industrie, white paper, 2026, DOI 10.48669/pls_2026-2 (German). PDF
- Bitkom: press release on "KI in der Wirtschaft 2026", 14 September 2026 (German). Website
- Magesh, V. et al.: Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools, arXiv, 30 May 2024. arXiv
- Bhatt, S. et al.: Position: Enterprise AI Must Enforce Participant-Aware Access Control, arXiv 2509.14608. arXiv
- CMS: KI-Transformation und Betriebsrat – Mitbestimmung richtig gestalten, legal update (German). Website
- Gleiss Lutz: Arbeitsgericht Hamburg zu ChatGPT: Kein Mitbestimmungsrecht des Betriebsrats, 27 March 2024 (German). Website




