Your company already knows the answer. It forgot where it put it.
This page has a calculator for what forgetting costs, a fictional company's Brain to question, and a section on where that memory should run.
A Company Brain is a wiki that a language model writes and keeps current from a company's raw documents. Every page links to the sources it was built from. Every answer links to the pages it used. Every change is versioned, dated and checked by a named owner, so anyone can trace a sentence back to the paragraph behind it.
Interaction workers spend about 19% of the workweek searching for and gathering information. McKinsey Global Institute, 2012, based on IDC data 1
The famous 9.3 hours never existed. The real cost is worse, and we show the math.
The often quoted "9.3 hours a week" is not in the 2012 McKinsey report it is credited to.2 The report gives a share: 19% of a 46.5-hour week, for interaction workers, which is about 8.8 hours (our calculation).1
Search time is the visible cost. Departures are the larger, hidden one. In a 2018 US survey, 42% of institutional knowledge was unique to the individual employee, and knowledge workers spent 5.3 hours a week waiting for colleagues or recreating what already existed.3
Every expert who leaves takes a part of your company with them.
| Figure | CHF per year |
|---|---|
| Search cost (S) | CHF 3'192'000 |
| Plausible saving range | CHF 399'000 to 798'000 |
| Knowledge at risk (K) | CHF 294'000 |
Show the formula
Search cost per year S = H × A × q Plausible saving range = S × r × 0.5 to S × r × 1.0 Knowledge at risk per year K = d × A × u × (t / 12) All results rounded to the nearest CHF 1'000.
Defaults: q = 19% (McKinsey Global Institute, 2012), or 25% with the Atlassian toggle (Atlassian, 2025); u = 42% (Panopto, 2018, US survey). H, A, d and t are placeholders. r is an assumption, shown as a band from half of it to all of it.
Survey-based defaults, not your data. Replace every default with your own number. How we grade evidence
The link in your address bar holds your numbers, so you can share it.
Search is not knowledge. Compiled is.
Search finds documents and leaves the reading to you, every time. It ranks the 2019 slide deck next to the 2026 procedure and does not say which one is current. A Company Brain does the reading once, writes down what it found and keeps that page current. It works in three layers.
Raw
The original documents: contracts, procedures, minutes, email threads, interview notes. Never edited, only added to.
Wiki
Pages the model writes and rewrites, one topic per page. Every sentence carries its source. Every page keeps its history.
Index
The map: which pages exist, what they cover, who owns them and when each was last checked.
A new document updates only the pages it touches, and each change is logged. When two documents disagree on a figure, the model does not average them. The page shows both figures, both sources and a conflict tag until the owner decides.
Each page has a named owner and a "last checked" date. A page nobody has checked for a year says so in plain view. The Brain is also where decision records live, so next year's board can see what was decided and on which evidence.
A wiki that nobody verifies is a rumour mill with a search bar.
- Ingest. A new document lands in raw, unchanged.
- Compile. The model rewrites only the pages it touches, each sentence with its source.
- Flag. Where two sources disagree, the page shows both and a conflict tag.
- Index. The map updates: page, owner, version, last checked.
- Verify. The owner checks the page and the new version is dated.
The LLM-wiki idea, a model that builds and maintains a linked wiki from raw sources, was described by Andrej Karpathy in a public note (2026). Credited as an idea; no affiliation.5
Every answer shows its sources. If it cannot, it says so.
Demowerk Präzision AG is a fictional parts maker: 14 documents, 24 wiki pages. Ask one of eight questions. Click any [n] to see the page, its version and the paragraph behind it. Then switch to retrieval only.
RAG re-reads. A Company Brain remembers.
Retrieval-augmented generation (RAG) cuts your documents into chunks. When a question arrives, it fetches the closest chunks and asks a model to answer from them. Each question starts from zero, so the model meets every contradiction fresh, and nobody can say which version of a figure it used.
Retrieval keeps improving. A 2024 study on contextual retrieval cut top-20 retrieval failures by 35%, by 49% with BM25 keyword matching added, and by 67% with reranking, but the model still receives fragments.6
A compiled wiki merges duplicates, tags conflicts and gives every page a date and an owner before anyone asks. Retrieval keeps a job inside a Company Brain: it finds the right page instead of the nearest fragment.
| Criterion | Fine-tuned model | Retrieval (RAG) | Compiled wiki |
|---|---|---|---|
| Freshness | As old as the last training run | As fresh as the last re-index | Updated page by page as documents arrive |
| Auditability | None: the answer comes from weights | Chunk ids, no owner, no version | Sentence to page to paragraph, with owner and version |
| Time to correct a wrong fact | Retrain | Fix the document, re-index, hope the right chunk wins | Edit one page |
| Running cost | High per retrain, low per question | Retrieval and long prompts on every question | Compile once per change, short prompts per question |
Your company's memory should not live in someone else's jurisdiction.
A Company Brain holds contracts, margins and the reasons behind decisions. Where it runs decides which law applies to it.
Three setups work: a Swiss data centre, your own hardware on-premise, or your own cloud tenant in a region you choose. In each, the encryption keys stay with you.
Open-weight models make on-premise practical, because the model runs where the documents are. One example: Apertus, an open multilingual language model by EPFL, ETH Zurich and CSCS, released on 2 September 2025.7
Every query is logged: who asked what, and which pages answered. The log shows which knowledge gets used and who saw what.
Switzerland's revised data protection act has applied since 1 September 2023; certain intentional breaches carry fines of up to CHF 250'000, imposed on the responsible individual rather than the company.8 On 9 November 2023 the FDPIC stated that the current act applies directly to AI-supported data processing.9
The same boundary logic runs our audits.
Your documents
- Where
- Your systems, unchanged
- Kept
- As today
Compile
- Where
- Swiss data centre, open-weight model
- Kept
- Nothing after the job
Wiki and index
- Where
- Swiss data centre
- Kept
- Every version, until you delete it
Query log
- Where
- Swiss data centre
- Kept
- Retention period you set
Keys
- Where
- Your key management
- Kept
- With you
Solid: infrastructure you controlDashed: a host under contract with you
Where this breaks.
- The Brain is only as good as the documents it is fed. That's why we recommend to automate ingestion.
- A model can still misread a source, which is why every sentence links to the paragraph it came from.
- Tacit knowledge that was never written down needs interviews or meeting transcripts.
Questions.
How is this different from our intranet search?
Search returns documents and leaves the reading to you. The Brain answers from pages that were already read, reconciled and dated, and every sentence links to the paragraph it came from. Inside the Brain, search finds the right page.
Who verifies pages?
A named owner per page, usually the person accountable for that topic in the business. The model drafts and updates; the owner checks. Every page shows its owner and its last-checked date, and the history shows what changed and when.
Can it run without data leaving Switzerland?
Yes. It can run in a Swiss data centre or on your own hardware with an open-weight model, so documents, pages and query logs stay in Switzerland. Which setup fits depends on your documents and your IT, and we settle it before the first document is ingested.
What happens when two documents contradict each other?
The page keeps both statements, links both sources and carries a conflict tag. The owner resolves it, and the resolution becomes a new, dated version. Until then, every answer that uses the page shows the conflict.
How long until a first domain is useful?
It depends on how many documents the first domain holds and how clean they are. If documents flow in at a steady pace, expect one to two weeks.
Sources.
- McKinsey Global Institute. The social economy (full report). 2012. WE
- Dacide check of the same McKinsey Global Institute report (2012): the "9.3 hours a week" figure attributed to it does not appear; the report states a percentage. Full report PDF. CC
- Panopto. Workplace knowledge sharing survey, press release. 2018. PO
- Atlassian. State of Teams 2025. 2025. PO
- Andrej Karpathy. LLM wiki, public note. 2026. PO
- Anthropic. Introducing Contextual Retrieval. 2024. WE
- ETH Zurich. Apertus: a fully open, transparent, multilingual language model, press release. 2025.WE
- Swiss Confederation, Fedlex. Federal Act on Data Protection (FADP), SR 235.1. In force since 1 September 2023.WE
- Federal Data Protection and Information Commissioner (FDPIC). Current data protection legislation is directly applicable to AI. 9 November 2023, updated May 2025.WE