Your business knowledge, machine-readable
In most companies the most valuable asset is scattered: in the heads of long-serving staff, in mailboxes, on network drives, in systems that do not talk to each other. A holistic knowledge approach brings it together: capturable, findable and in a form a company-owned language model can actually work with.
We show the derivation, the principle behind it, the possible technology routes and what you get out of it.
Why this is the decisive topic right now
Three developments hit Swiss mid-sized companies at the same time. Each on its own would be manageable. Together they create a pressure that can no longer be sat out.
Knowledge is retiring
The large post-war cohorts are leaving the labour market. By 2030 Switzerland is estimated to be short of around 500,000 workers. Every retirement takes with it experience that was never written down: why a plant was built that way, why a customer gets an exception, how a special case was solved seven years ago.
Searching eats the working day
According to the widely cited McKinsey figure, knowledge workers spend around 1.8 hours a day searching for and gathering information, close to a quarter of their working time. In a company with 80 office staff that equals the output of roughly 18 full-time positions doing nothing but searching.
AI is waiting to be fed
Language models have been good enough for production use for two years. What is missing is not the model. It is the accessible body of knowledge underneath. Gartner expects around 80% of companies to run knowledge-grounded AI systems by 2027. Without an ordered knowledge base the technology simply cannot be applied.
From filing cabinet to knowledge network
The difference between «we have documented everything» and «we can use our knowledge» is not a question of volume but of linkage. On the left the normal state, on the right the target picture.
Four layers · and only one of them follows the fashion
Every holistic knowledge approach has the same structure, whatever products implement it. Keep the four layers cleanly apart and you can swap individual pieces later without rebuilding the body of knowledge.
What makes a body of knowledge machine-readable
A language model reads differently from a person. It has no memory of your company and no instinct for which version is the valid one. It therefore needs exactly what good documentation has always needed, only carried through consistently:
- Small, self-contained units. One matter per note instead of 80-page compendia. What is looked up as one piece must be findable as one piece.
- Properties on every object. Type, date, affiliation, validity, confidentiality: machine-readable, not buried in prose.
- Explicit links. A set of minutes points to customer, project and participants. These edges are the difference between search and understanding.
- One place of truth. Exactly one valid version per question. Three versions of the same price list make every answer useless.
- Provenance and age. Where does the statement come from, when was it last confirmed? Without provenance no answer can be checked.
- Permissions on the object. Rights must travel with the content, not hang off a folder. Otherwise the assistant answers questions it should not.
Failure patterns in practice
- «We do have an intranet»
- Storage without structure or upkeep: search returns 400 hits, none of them current.
- «The assistant makes things up»
- Almost always not a model fault but a content fault: contradictory or outdated sources.
- «AI will sort that out»
- A model can help sort a mess, but it cannot create information that is not there.
- «Platform first, content later»
- The most common expensive ordering. Without content every platform stays an empty shelf.
A knowledge store that has been running for years
We run this approach inside our own network, deliberately in the leanest conceivable form: open text files in a folder structure, no database server, no vendor lock-in. The principle is the same as in a corporate platform, only without its implementation project.
| Building block in our own operation | What it delivers | Equivalent at corporate scale |
|---|---|---|
| Open text files in folders | The store is readable, versionable and portable without special software. | Document store with enforced format and mandatory metadata |
| Properties in every file | Type, affiliation, date, confidentiality, filterable by machine. | Content types and managed metadata in the platform |
| Links inside the text | Relationships between objects are part of the content, not of the software. | Knowledge graph with entities and edges |
| Atomic activity records | Every mail, conversation and meeting as its own entry with a fixed structure. | Event data in the estate, kept current automatically |
| Daily state file | A short, always-valid situation report: the entry point for every enquiry. | KPI cockpit as context for assistants |
| Hybrid search index | Keyword and meaning combined, finds things even without the right words. | Vector search alongside classic full-text search |
| Attachments with their parent | No central image graveyard: the plan sits with the project that needs it. | Document management by object rather than by drive folder |
| Automatic versioning | Every change traceable, every state restorable. | Audit readiness and retention rules |
The decisive point in this example is not the tool. It is discipline. The store works because every new object follows the same rules: fixed location, fixed properties, fixed naming, checked for duplicates. Exactly those rules transfer to any corporate platform and without them, even the most expensive one will not help.
One principle, several routes
The principle is always the same: the implementation depends on where your knowledge already sits. We are technology-open: we build the approach where you have already invested, rather than adding another island beside it.
| Question | Route A · business system | Route B · data platform | Route C · file-based |
|---|---|---|---|
| Typical components | Knowledge graph and assistance layer of modern cloud ERP platforms, for example SAP Knowledge Graph, Business Data Cloud and the associated agent tooling | Data platforms such as Microsoft Fabric with OneLake, plus the document world of SharePoint and Teams as well as knowledge agents and Copilot surfaces | Open text formats, a self-operated search index, a language model of your choice, including one in your own data centre |
| Strength | Process knowledge and master data are already structured and current | Reach across the whole office world, permissions cleanly attached to content | Quick to stand up, inexpensive, entirely under your own control |
| Weakness | Knowledge outside the system stays outside at first | Licence and consumption costs need managing | Connections to business systems have to be built |
| Time to value | Months, depending on platform state | Months, data build-up and permissions dominate | Weeks, first productive value very early |
| Data sovereignty | Contractual assurances from the platform provider | Contractually governed, region selectable | Entirely in house, operable offline |
Product names here are deliberately examples, not recommendations. Which route carries depends on three questions: where your knowledge sits today, how strict your requirements for data handling and auditability are, and how quickly the first value has to become visible.
Within a few years this is the normal state
The integrated knowledge approach will become what an ERP system became in the nineties: no longer a competitive advantage, but a precondition. The lead is won in the window that is open now, by those who start early.
Scattered
Knowledge lives in heads, mailboxes and drives. Every answer needs a person who happens to know. Absences and departures hit immediately.
Filed and findable
There are defined locations, naming rules and a working search. Those who know what they are looking for will find it. This is the state many companies consider «done».
Linked and maintained
Objects carry properties and relationships; knowledge is created automatically where work happens, from meetings, mail and systems. The store no longer ages in silence.
Usable by machines
A company-owned assistant answers questions from your own knowledge, with sources and within permissions. Onboarding times typically fall from months to weeks.
Anchored in the process
Agents work alongside day-to-day business: they prepare quotations from comparable cases, answer service enquiries with the history behind them, onboard new staff. The company's knowledge takes effect without anyone searching for it.
The value is measurable
Knowledge management has a reputation for being an end in itself. That was true as long as it meant filing. Once the body of knowledge is usable by machines, it takes effect where it counts.
| Effect | How you notice it |
|---|---|
| Serving customers faster | Enquiries are answered at first contact instead of being passed around internally. A customer's history is available in seconds. Not once the responsible person is back from holiday. |
| Quotations in hours, not days | Comparable cases, the pricing used then and any special arrangements are findable. Costing no longer starts from a blank page. |
| Shorter onboarding | New colleagues read the history instead of asking for it. Market experience points to a reduction from around four months to roughly six weeks. |
| Departures become survivable | The knowledge of a retirement is documented instead of walking out with the person. Succession becomes plannable rather than risky. |
| Decisions stay on record | Who committed to what, when and on what grounds. In complaints and project disputes, what is documented is what counts. |
| Mistakes repeat less often | Whatever went wrong once is findable, including cause and fix. The second service case of the same kind costs a fraction of the first. |
| Contradictions surface earlier | When someone files something that contradicts an existing statement, the system flags it on the way in. Ambiguities and outdated versions come to light at capture. Not weeks later in a customer conversation. |
| AI becomes applicable | Every future assistant and agent tool draws on the same ordered body of knowledge. The investment pays into everything that comes after. |
| Audits get calmer | Evidence, versions and validity are on file. Compiling material for an audit or certification becomes a query. |
Not a large programme · one first area
We do not start with a platform decision but with the knowledge you miss most today. The first value should be there before the second invoice arrives.
Knowledge inventory
Half a day on site: which questions cost the most time today, where does the answer sit, who is the bottleneck? The result is a map of your knowledge, with the three places where starting pays off fastest.
Structure and pilot
We define the object model: which types, which properties, which rules and build it for one area. Capture runs automatically from day one, so the store does not depend on manual work.
Putting it to work
Search and assistant go on top of the store. Only once answers are reliable do we extend, to further areas, further sources and, where it pays, to agents in daily business.
Operation then passes to your own people. A body of knowledge that depends on external upkeep is not one, which is why handover is part of the approach from the start.
Where the figures come from
We name our sources so you can check the argument and because reliable figures in this field are rare while widely quoted ones are often old.
- Search effort of roughly 1.8 hours per working day: a figure originating in a McKinsey survey and widely cited; overview of the evidence at Cottrill Research and Valamis. The original survey dates from the early 2010s. We use it as an order of magnitude, not a point value.
- Adoption of knowledge-grounded AI systems by 2027 as well as cost and growth figures for this approach: AutomationFlow, market overview 2026 and Atlan, comparison of enterprise knowledge platforms 2026.
- Systematic literature review on knowledge-grounded language models in enterprise knowledge management: Applied Sciences (MDPI), 2026.
- Knowledge graph and assistance layer in the cloud ERP world: SAP Community, knowledge-graph-driven runtime and SAP News Center, May 2026.
- Knowledge agents and the permission model on the data platform: Microsoft Learn, Fabric overview and Microsoft Azure Blog, Build 2026.
- Skills situation and the standing of AI among Swiss mid-sized companies: KMU ZH Monitor 2026 and NZZ SME Barometer 2026.
- Knowledge loss on staff turnover and shortened onboarding: Knowledge management guide for Swiss SMEs.
Where is your most valuable knowledge?
In a first conversation we work out which area pays off first and which route fits your system landscape. Without obligation and as equals.
Get in touch → All solutions