The verdict in three sentences
A RAG chatbot (retrieval-augmented generation) answers employee questions using only your internal documents, with the source cited, for EUR 15,000 to 40,000 excl. VAT of setup and EUR 100 to 800 a month in usage. The non-negotiable point is document-level access control: an employee must never get an answer drawn from a contract or HR file they are not allowed to open. With EU hosting and a rigorous test set, a correct answer rate above 85% is achievable in 2 months.
How an enterprise RAG works
Documents (SharePoint, Google Drive, Confluence, intranet, procedure PDFs, contracts) are split into passages, turned into vectors and stored in a search index. When an employee asks a question, the system retrieves the 5 to 10 most relevant passages among those they are allowed to see, then the language model writes an answer citing its sources. If no passage answers, the chatbot must say so instead of making something up.
| Component (2026 order of magnitude) | Standard version | Secured mid-cap version |
|---|---|---|
| Source connectors | 2 sources, EUR 3,000 to 6,000 | 4 to 6 sources, EUR 8,000 to 14,000 |
| Indexing and chunking | EUR 2,000 to 4,000 | EUR 4,000 to 7,000 |
| Access rights sync | By group, EUR 2,000 | Per document via Entra ID or Google, EUR 5,000 to 9,000 |
| Interface (Teams, Slack, web) | EUR 2,000 to 4,000 | EUR 3,000 to 6,000 |
| Test set and evaluation | EUR 1,500 to 3,000 | EUR 3,000 to 5,000 |
| Total setup | EUR 15,000 to 20,000 excl. VAT | EUR 25,000 to 40,000 excl. VAT |
| Monthly usage cost | EUR 100 to 300 | EUR 300 to 800 |
Security: five requirements to enforce
Usage cost depends mostly on question volume: for 500 employees asking 8 questions a month on average, i.e. 4,000 requests, the model bill is between EUR 80 and 250 depending on the model, plus EUR 50 to 300 for vector database hosting. The security budget, however, is not negotiable.
| Requirement | What to check | Risk if missing |
|---|---|---|
| EU hosting | EU region for the model and vector database | GDPR non-compliance |
| No data reuse | Contractual no-training clause | Know-how leakage |
| Per-document access rights | Filtering before retrieval, not after | An intern reads payroll |
| Logging | Who asked what, which sources | No audit possible |
| Purge and updates | Deleted document removed within 24 h | Answer based on an obsolete procedure |
| Mandatory citations | Link to the source passage | Undetected hallucinations |
Reaching 85% correct answers
Quality is not declared, it is measured. Before going live, build with your experts a set of 150 to 300 real questions with the expected answer. Every version of the system is scored on that set. First attempts often land around 65 to 75%; you gain 10 to 20 points by removing duplicate documents, adding metadata (date, approved or draft status) and tuning chunking. Typical schedule: weeks 1 to 3 for connectors and permissions, weeks 4 to 6 for evaluation and tuning, weeks 7 and 8 for a pilot with 50 users.
Need a professional website?
Kolonell builds websites that attract clients, optimized for the Sénégalese market. Free quote in 2 minutes.
Mini case study
Isabelle, COO of a 500-employee services mid-cap in Lille, estimates each employee loses 3 hours a month looking for a procedure or asking an expert. If the chatbot recovers 1.5 hours, that is 750 hours a month. Valued conservatively at EUR 35 fully loaded and counting only 30% of that time as genuinely reallocated, the gain is EUR 7,875 a month. For a EUR 32,000 excl. VAT setup and EUR 600 monthly usage, the project pays back in about 4.5 months. Legal and HR experts, asked 40% less often, regain time for complex cases.
FAQ
Can the chatbot make up an answer? The risk exists, hence mandatory source citations and an "I don't know" answer when no relevant passage is found. On a well-built test set, the hallucination rate drops below 3%.
Should we host an open-source model in-house? Only for very sensitive data (defence, health). An open-source model on a dedicated GPU server costs EUR 1,500 to 4,000 a month, versus EUR 100 to 800 for an EU-hosted API.
How many documents can be indexed? A base of 50,000 documents is common and cheap: initial indexing costs a few tens of euros. The difficulty is document quality, not quantity.
Does GDPR allow this project? Yes, with an impact assessment if personal data is indexed, EU hosting and employee notice. Budget 2 to 4 days of work with the DPO.
Can it be integrated into Microsoft Teams? Yes, it is the most adopted interface in mid-caps. The extra cost is EUR 2,000 to 4,000 and adoption often exceeds 60% of staff within 3 months.
Let's scope your project. Tell us your document sources, your directory (Entra ID or Google) and the number of users: we will price a secured RAG between EUR 15,000 and 40,000 excl. VAT, delivered in 2 months. 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.
