The verdict in three sentences
An AI assistant connected to internal documentation using RAG (retrieval-augmented generation) costs between EUR 20,000 and 60,000 for a 900-employee company with 15,000 documents, plus EUR 200 to 1,500 a month in model usage. The project succeeds or fails on two points that rarely show in demos: the quality of document chunking and enforcement of per-document access rights. Done well, it saves about 30 minutes per day per technician, which pays back the investment within months.
What a serious RAG project includes
A prototype answering over 50 PDFs takes a week. A reliable tool over 15,000 procedures, technical sheets and internal notes, with versions changing every month, is a real project.
| Work package | Content | Indicative 2026 cost (EUR excl. VAT) | Duration |
|---|---|---|---|
| Scoping and document audit | Source inventory (DMS, SharePoint, PDFs, wiki), quality, duplicates, outdated versions | 3,000 to 6,000 | 2 weeks |
| Ingestion and chunking | Connectors, table and diagram extraction, section chunking, metadata | 5,000 to 15,000 | 3 to 5 weeks |
| Search and answer engine | Vector store, hybrid search, source citations, refusal outside the corpus | 5,000 to 15,000 | 3 to 4 weeks |
| Per-document access rights | Sync with the directory (Entra ID, LDAP), filtering before generation | 3,000 to 10,000 | 2 to 3 weeks |
| Interface and integration | Web, Teams or shop-floor mobile app, per-answer user feedback | 2,000 to 8,000 | 2 to 3 weeks |
| Evaluation and acceptance | Set of 200 to 400 reference questions, measured accuracy | 2,000 to 6,000 | 2 weeks |
The low end covers a single document source and a web interface. The high end includes several sources, fine-grained rights and a mobile interface for technicians on the shop floor or on site.
Running costs and hosting choices
Usage cost depends on question volume and the chosen model. For 900 employees with 150 regular users, expect 3,000 to 8,000 questions a month.
| Monthly item | Budget option | Mid-range option | Sovereign option |
|---|---|---|---|
| Language model | Compact model API, EU region: EUR 150 to 400 | Advanced model API, EU region: EUR 400 to 1,000 | Open-source model hosted in Germany: EUR 800 to 1,500 (GPU) |
| Vector store and storage | EUR 50 to 150 | EUR 100 to 250 | EUR 150 to 300 |
| Re-indexing updates | Included | Included | Included |
| Maintenance and improvement | EUR 300 to 600 | EUR 500 to 900 | EUR 700 to 1,200 |
| Monthly total | EUR 500 to 1,150 | EUR 1,000 to 2,150 | EUR 1,650 to 3,000 |
| Data location | EU | EU | Germany, no transfer outside the EU |
For most mid-sized companies, the mid-range option in an EU region, with a contractual commitment not to use data for training, is enough. The sovereign option is justified for defence, healthcare or sensitive trade secrets, and is often requested by German works councils.
The traps that sink the project
The first trap is documentary: if three versions of the same procedure coexist, the assistant will sometimes cite the wrong one. The initial audit must flag outdated documents. The second is security: an assistant that shows a technician a confidential HR note destroys trust in a day. Filtering must happen before generation, not after. The third is adoption: without business champions and a "useful or not" feedback button, usage drops after six weeks.
Mini case study
Need a professional website?
Kolonell builds websites that attract clients, optimized for the Sénégalese market. Free quote in 2 minutes.
Julian, CIO of a 900-employee industrial company, targets 120 maintenance technicians who search 15,000 documents for procedures every day. Assuming a realistic gain of 15 minutes a day on average, half the 30-minute target, over 210 working days: 120 x 0.25 h x 210 = 6,300 hours a year. At a loaded hourly cost of EUR 48, that is EUR 302,400 a year. The project costs EUR 45,000, plus EUR 1,200 a month to run, or EUR 59,400 in year one. Payback comes in under three months of real use, even with partial adoption.
FAQ
Can the assistant make up answers?
The risk exists; it shrinks by enforcing source citations and refusal outside the corpus. On a set of 300 reference questions, aim for 85 to 92% correct answers before go-live.
Can our documents be used to train the provider's model?
Not if the contract says so: enterprise offers in EU regions exclude training on your data. Check the clause and the query retention period, often 0 to 30 days.
How long before a first usable version?
Allow 6 to 8 weeks for a pilot on one department and 2,000 to 3,000 documents, then 2 to 3 months to extend to all 15,000 documents.
Do we have to clean all documentation first?
No, start with the 20% most consulted documents. The audit usually identifies 10 to 25% of outdated documents to archive.
Can we measure usage and quality?
Yes, a dashboard tracks question volume, the share of answers rated useful and unanswered questions, which reveal gaps in the documentation.
Let's scope your project. Tell us about your document sources, document volume and target user groups: we will price the pilot, the indicative budget and the rollout plan. Detailed quote within 48 h. WhatsApp +221 77 596 93 33.
Mohamed Bah
Fondateur, Kolonell
Passionate about digital and entrepreneurship in Africa, Mohamed has been helping Sénégalese businesses with their digital transformation since 2020. Founder of Kolonell, he believes every SME deserves a professional and accessible online présence.
