Digital Africa11 min read

Internal AI Assistant on Company Documents (RAG): Cost and Security (2026)

Mohamed Bah·Fondateur, Kolonell
October 10, 2026
Share:
Internal AI Assistant on Company Documents (RAG): Cost and Security (2026)

Internal AI Assistant on Company Documents (RAG): Cost and Security (2026)

Digital Africa

The verdict in three sentences

An internal RAG AI assistant (retrieval-augmented generation) answers employee questions by citing company documents, without retraining any model. For a 600-person group, budget EUR 20,000 to 60,000 excl. VAT to build it and EUR 2 to 6 per user per month to run it, EU hosting included. The project hinges on one point: enforcing access rights document by document, otherwise the assistant becomes an internal data leak.

What a RAG assistant really costs in 2026

The budget depends less on the language model than on the work on sources: cleaning procedures, SharePoint or Google Drive connectors, document chunking, version handling. The ranges below are 2026 orders of magnitude observed on SME and mid-size company projects.

ItemSmall scope (1 department)600-employee groupNote
Scoping and source auditEUR 3,000EUR 6,000 to 9,000SharePoint inventory, duplicates, outdated versions
Connectors (SharePoint, Drive, intranet)EUR 4,000EUR 8,000 to 15,000Incremental sync every hour
Indexing and vector searchEUR 3,000EUR 6,000 to 12,000Vector database hosted in the EU
Access rights (ACL)EUR 2,000EUR 5,000 to 12,000Per-query filtering based on Entra ID identity
Interface (Teams, web, widget)EUR 2,500EUR 4,000 to 8,000Answers with clickable sources
Testing, evaluation, trainingEUR 2,500EUR 4,000 to 6,000Set of 150 to 300 test questions
Total buildabout EUR 17,000EUR 33,000 to 62,0006 to 12 weeks

Running costs stay modest: model calls, vector storage and hosting amount to EUR 2 to 6 per active user per month. With 600 employees including 350 regular users, that is EUR 700 to 2,100 per month, plus maintenance of EUR 500 to 1,500 excl. VAT per month.

Security: five non-negotiable requirements

A CIO should put these requirements in the specification before any demo. An assistant that gives a technician access to payslip content because payslips were indexed is the most common failure scenario.

RequirementWhat to ask forRisk level if missing
Inherited access rightsFiltering by Entra ID group on every queryCritical
EU hostingModel and vector database in an EU region, DPA contractHigh (GDPR)
No training on your dataContract clause from the model providerHigh
LoggingQuestion history kept 90 to 365 daysMedium
Systematic citationsEvery answer links to its source documentMedium (hallucinations)
Sensitive folder exclusionBlocklist for HR, legal, executiveCritical

Off-the-shelf offers such as Copilot for Microsoft 365 natively respect SharePoint rights but cost about EUR 28 to 30 per user per month, nearly EUR 200,000 per year for 600 licenses. A custom RAG focused on procedures is often 3 to 5 times cheaper over three years, with finer control over sources.

Custom build or market solution

CriterionCopilot Microsoft 365SaaS RAG platformCustom RAG
Annual cost, 600 usersEUR 200,000 to 215,000EUR 25,000 to 60,000EUR 15,000 to 40,000 + build
Deployment time2 to 4 weeks4 to 6 weeks6 to 12 weeks
SharePoint rightsNativeDepends on connectorBuilt on your groups
Model choiceImposedLimitedFree (Mistral, Claude, GPT)
Guaranteed EU hostingYes (EU Data Boundary)Depends on vendorYes
Business customizationLowMediumHigh

Mini case study

Nathalie, CIO of a 600-employee retail group, manages 4,200 procedures scattered across 38 SharePoint sites. Her 40-person support team spends about 45 minutes a day looking for the right version of a procedure. After deploying a RAG assistant for EUR 42,000 excl. VAT, search time drops to 15 minutes: 30 minutes saved per day per support employee.

Need a professional website?

Kolonell builds websites that attract clients, optimized for the Sénégalese market. Free quote in 2 minutes.

Prefer a call back?

Leave your WhatsApp number and a Kolonell expert will get back to you within 1 business day. Free, no strings attached.

You are :

Calculation: 40 people x 0.5 h x 210 days = 4,200 hours per year. At a loaded hourly cost of EUR 38, the gain reaches EUR 159,600 per year. With EUR 18,000 of annual costs (usage and maintenance), the project pays for itself in just over 3 months.

FAQ

Can a RAG assistant make up answers?

The risk exists but it shrinks when every answer cites its sources and the assistant is instructed to say "I don't know" outside the corpus. On a 200-question test set, a good project targets over 90% correct, sourced answers before going live.

Should SharePoint be cleaned before starting?

Yes, at least partly: on most projects, 20 to 40% of documents are duplicates or outdated versions. Allow 2 to 4 weeks of sorting with business teams, otherwise the assistant will cite expired procedures.

Which language model should we pick?

For French or English documents, Mistral, Claude and GPT perform comparably on document search. The choice is mostly driven by EU hosting and cost, which varies by a factor of 1 to 4 across models.

How long for a first pilot?

A pilot covering one department and 300 to 500 documents ships in 4 to 6 weeks for about EUR 15,000 to 20,000 excl. VAT. It measures the correct-answer rate before rolling out group-wide.

Does data leave the company?

Relevant excerpts are sent to the model with each question, hence the importance of a contract with EU-region hosting and no reuse for training. Full documents stay in your vector database, encrypted at rest.

Let's scope your project. We scope your internal AI assistant (sources, access rights, EU hosting), with an indicative budget of EUR 20,000 to 60,000 excl. VAT and a pilot delivered in 6 weeks. Detailed quote within 48 h. WhatsApp +221 77 596 93 33.

Tags:#AI assistant#RAG#knowledge base#enterprise AI#data security#2026 cost
Share:

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.