Knowledge management with AI: fundamentals, architecture and practice

01

What is knowledge management, and what does AI change about it?

Knowledge management shapes how an organisation captures, structures, finds, applies and keeps its knowledge current. AI changes mainly two phases: written knowledge can be searched by meaning and turned into answers backed by sources, and spoken accounts of experience can be transcribed and structured faster. AI does not change who owns sources or keeps them maintained.

Classic knowledge management rarely failed because documents were missing. It failed because nobody could find them, nobody kept them current, and the one thing that mattered most was never written down. Wikis, shared drives and document management solved the storage problem, not the retrieval problem: anyone who did not know the right search term came up empty.

Language models shift that bottleneck. Semantic search finds documents even when the question and the text use different words. Retrieval-augmented generation turns those results into an answer with a source. That makes the quality of the underlying material more visible than ever: a system that cites an outdated work instruction fluently and with a source reference does more damage than a search that simply fails to find it.

Ten knowledge pages cover each phase and building block in more depth, linked at the relevant point throughout.

02

What is the difference between explicit and tacit knowledge?

Explicit knowledge is spoken or written down: manuals, policies, minutes. Tacit knowledge sits in experience and skill and is hard to put into words, such as how an experienced technician recognises a fault by sound alone. AI works almost entirely with explicit knowledge and only helps with tacit knowledge once people articulate it first.

The chemist and philosopher Michael Polanyi coined the observation, in The Tacit Dimension (University of Chicago Press, 1966), that we know more than we can tell. Ikujiro Nonaka built on this in A Dynamic Theory of Organizational Knowledge Creation (Organization Science, 1994) with a model in which organisational knowledge emerges through a continuous exchange between tacit and explicit knowledge, across four modes of conversion. That model is a useful, level-headed way to place the role of AI.

The table shows why most AI projects in knowledge management start with combination: the material is already there. Where the knowledge that matters most still has to be drawn out of people's heads, for instance ahead of a generational handover, Capturing knowledge before it retires sets out methods such as expert interviews, lessons learned and knowledge maps.

03

What phases does knowledge management have, and which AI technique fits each one?

A workable model distinguishes five phases: capture, structure, find, apply and maintain. Capture suits OCR and transcription, structuring suits extraction and knowledge graphs, finding suits semantic search and RAG, applying suits agents with approval points, and maintaining suits measuring answer quality. Access rights and operations run through every phase.

• Capture: Understanding documents with AI explains how scans, tables and forms become machine-readable, and when visual search beats OCR.
• Structure: Knowledge graphs and GraphRAG shows when relationships matter more than similarity.
• Find: Semantic search in the enterprise covers the search techniques, and Retrieval-augmented generation covers the answer systems built on top.
• Apply: AI agents in knowledge work takes retrieval the rest of the way into the workflow.
• Maintain: Checking AI answers covers hallucinations, faithfulness to sources and evaluation.

04

How is an AI knowledge base structured?

An AI knowledge base consists of connections to the sources, an ingestion step that carries over text, structure, metadata and rights, a search index, a retrieval layer with an access filter, a language model for answers backed by sources, and an interface with feedback. Every layer is interchangeable and can run on an organisation's own network or in EU hosting.

The pattern traces back to retrieval-augmented generation (Lewis et al., 2020): the knowledge sits in the index, not in the language model. Two decisions shape the architecture. First, how read permissions carry from the source systems through to retrieval, covered in Access control in AI knowledge systems. Second, where the model runs; what running one with open weights requires is covered in Local language models. AI knowledge management describes how iiterate implements such systems.

The last layer is often underrated. An interface where users can flag an answer as wrong or outdated connects the system to its maintenance: the flag reaches the person responsible for the cited source. What gets logged is the question, the retrieved sources and the rating, without personal content wherever possible. Over time, these logs build the test set against which every later change is measured.

05

What framework does the ISO 30401 standard offer for knowledge management?

ISO 30401 sets requirements and guidance for establishing, implementing, maintaining, reviewing and improving a knowledge management system. The first edition appeared in November 2018 and was amended in 2022 and 2024; a revision has been out for vote since 2026. The standard covers organisation and process, not software.

According to its introduction, the standard is meant to guide organisations while also serving as a basis for audits and certification. Its requirements are written for organisations of any type and size. The draft of the second edition, ISO/DIS 30401, has been out for vote since 25 June 2026 according to ISO, and is meant to replace the 2018 edition.

06

How widespread is AI use among companies in Germany and the EU?

According to the Federal Statistical Office, 26% of German companies with ten or more employees used AI technologies in 2025, rising to 57% among those with 250 or more. Eurostat put the EU figure at 20.0%, up from 13.5% a year earlier. The most common technology was analysing written language, the field that knowledge management with AI builds on.

These figures measure use, not benefit: neither survey captures whether answers are correct or sources are kept current. The AI readiness check offers a structured way to see where an organisation stands.

07

How ready is your organisation's knowledge management for AI?

A self-check asks whether knowledge exists in writing and can be found, whether ownership and versions are clear, whether access rights are maintained, whether real test questions exist, and whether tacit knowledge is being captured. If only a few points hold true, it usually pays to sort out the material and ownership first and build an AI system only afterwards.

As a rule of thumb: ticking six or more points means a prototype can be meaningfully measured against real data. At three to five, a narrowly scoped knowledge area makes sense, with the missing points cleared up first. With fewer than that, the work still lies in organisation and material. The RAG readiness check covers the state of the documents specifically.

08

When does AI not help with knowledge management?

AI does not help where the problem is organisational: when nobody owns the sources, when people withhold knowledge out of concern for their own role, when documents contradict each other, or when the one thing that matters was never written down. An AI system makes these gaps visible; closing them is up to the organisation itself.

Culture cannot be installed. Where employees treat knowledge as personal job security, an assistant changes nothing about that; where leadership recognises documentation as part of the job, it quickly becomes useful.