Internal AI Assistant on a Company Knowledge Base in Berlin: Cost and Security

The verdict in three sentences

An internal AI assistant connected to your knowledge base (RAG) costs EUR 20,000 to 60,000 excl. VAT to build and EUR 100 to 800 per month to run for 300 users. The real challenge is not the language model but document-level access control and hosting inside the European Union, two non-negotiable requirements for engineering and industrial firms in Berlin. With 30 minutes saved per employee per week, the project pays for itself in under four months.

What a RAG assistant covers in 2026

The RAG (Retrieval-Augmented Generation) principle is simple: instead of letting an AI model answer from memory, each question is fed the relevant passages from your own documents (quality procedures, maintenance manuals, engineering notes, HR policies). The assistant answers and cites its sources, so the employee can check.

For the CTO of a mid-sized company, the project breaks down into clear building blocks. Document ingestion (PDF, Word, Confluence, SharePoint, PLM) often accounts for 30% of the effort, because technical documents contain tables, drawings and multiple versions. Vector search and passage ranking come next, then the interface (chat embedded in Teams or the intranet) and finally the security layer.

Security and hosting: the choices that move the budget

In a company bound by confidentiality requirements (aerospace or automotive supply chain, defence, export control), the question is not whether generative AI can be used, but where the data travels. Three architectures dominate in 2026.

The most underestimated point is rights filtering. If a production technician queries the assistant, they must never see a passage from an HR file or a classified customer contract. Each indexed chunk must carry the rights of its source document, and those rights must be checked on every query, not only at indexing time. Add query logging (kept 6 to 12 months), encryption at rest and a GDPR data protection impact assessment if personal data sits in the corpus. In Germany, involving the works council early is also wise, since usage logs can be seen as employee monitoring.

Timeline and success metrics

A serious project runs 8 to 12 weeks: 2 weeks of scoping, 3 to 4 weeks of ingestion and search, 2 weeks of security and interface, then a 2 to 3 week pilot with 30 to 50 users before company-wide rollout.

Mini case study

Jonas, CTO of a 300-employee engineering company in Berlin, launches a RAG assistant for EUR 42,000 excl. VAT, hosted in Germany, with estimated usage of EUR 350 per month. Each employee saves an average of 30 minutes per week looking for procedures. Over 45 working weeks, that is 300 × 0.5 × 45 = 6,750 hours per year. At a loaded cost of EUR 55 per hour, the gain reaches about EUR 371,000 per year, against a first-year cost of 42,000 + 4,200 = EUR 46,200. Even if only a third of the time saved is truly productive, payback comes in under five months.

FAQ

Are our technical documents used to train the model?

No. In a properly contracted RAG architecture, documents stay in your index and are only sent to the model at question time. Enterprise offers from the main providers guarantee no reuse for training, with query retention limited to 0 or 30 days.

How many documents can an assistant index?

A mid-sized company commonly indexes 20,000 to 200,000 documents, meaning a few million passages. Vector storage stays modest at around EUR 50 to 200 per month; the main cost is the initial ingestion.

What happens when a procedure is updated?

An automatic sync, daily or hourly, re-indexes modified documents and removes obsolete versions. Budget EUR 3,000 to 6,000 excl. VAT if you have many sources (SharePoint, PLM, DMS).

Can the assistant make up an answer?

The risk exists but drops sharply when source citation is mandatory and the assistant is allowed to say "I don't know". A set of 200 reference questions lets you measure a target error rate below 5% before rollout.

Do we need an internal team to maintain the tool?

Half a day per week from a designated owner is usually enough, backed by a maintenance contract of EUR 400 to 1,200 excl. VAT per month covering model updates and monitoring.

Let's scope your project. Send us your document volume, user count and hosting constraints: we price a RAG assistant between EUR 20,000 and 60,000 excl. VAT, delivered in 8 to 12 weeks. Detailed quote within 48 h. WhatsApp +221 77 596 93 33.