The verdict in three sentences
In an industrial, banking or contact centre group, whether in Tunis or Singapore, procedures, work instructions and internal notes add up to 2,000 to 50,000 pages nobody finds quickly. A RAG AI assistant (retrieval-augmented generation) answers in natural language while citing the source document, respects each employee's access rights, and costs 40,000 to 120,000 TND (about EUR 12,000 to 36,000, or USD 13,000 to 39,000) to set up, then 0.5 to 3 TND (EUR 0.15 to 0.90) per user per day; a Singapore build typically costs 2 to 3 times more because of local engineering rates. The measured 2026 gain: procedure search time divided by 5, and fewer findings in quality audits.
How a RAG assistant works and what it costs
The principle is simple. Documents (PDF, Word, intranet pages, tickets, ISO procedures) are split, indexed in a vector database, then for each question the assistant retrieves the relevant passages and asks the language model to write an answer from those passages only, with references. It does not train on your data: it consults it, so a procedure can be updated without retraining anything.
| Component | Function | Setup cost (TND, 2026 order of magnitude) |
|---|---|---|
| Document connectors (SharePoint, DMS, file server, intranet) | Ingest and resync documents | 6,000 to 20,000 |
| Extraction and chunking (OCR of scans, tables, Arabic, French, English) | Make documents searchable | 5,000 to 18,000 |
| Vector database and hybrid search | Find the right passages | 6,000 to 15,000 |
| Rights management (per-document permission inheritance) | Show only what the user may read | 8,000 to 25,000 |
| Interface (web, Teams, mobile) and source citations | Daily use and verification | 6,000 to 20,000 |
| Evaluation, 200 to 500 test questions, acceptance | Measure accuracy before launch | 5,000 to 15,000 |
| Total project | 40,000 to 120,000 |
Usage cost depends on the model and volume: an employee asking 10 to 20 questions a day costs 0.5 to 1.5 TND per day with a standard API model, up to 3 TND with a premium model or dedicated hosting. For 300 active users over 22 days, plan 3,300 to 20,000 TND a month (EUR 1,000 to 6,000).
Security: the three hosting choices
The CIO of a bank or an exporting manufacturer asks about confidentiality first. In Tunisia, organic law no. 2004-63 on personal data and INPDP requirements apply as soon as documents contain employee or customer data, and banks must also follow Central Bank of Tunisia outsourcing circulars. In Singapore, the PDPA and MAS technology risk guidelines play the same role.
| Hosting option | Confidentiality | Indicative monthly cost (300 users) | Answer quality | Best for |
|---|---|---|---|---|
| Major model API, no data retention by provider | Good with the right contract | 3,300 to 8,000 TND | Very high | Industry, contact centres |
| Dedicated cloud in a chosen region, hosted model | Very good | 8,000 to 15,000 TND | High | Multi-country groups |
| Open model on local servers (on-premise or local data centre) | Maximum | 10,000 to 20,000 TND, plus 60,000 to 150,000 TND of GPU hardware | Good to high | Banks, insurers, public sector |
| Classic full-text search without AI | Maximum | 500 to 2,000 TND | Low on free-form questions | Small document bases |
Whatever the choice, four rules apply: filter documents by rights before calling the model, log every question, mask personal data, and decline to answer when no source is found. An assistant that invents a procedure is more dangerous than a slow search engine.
Rollout in three steps
Start with 2,000 to 5,000 pages and one pilot department (quality, customer support or compliance), 6 to 8 weeks. Measure on 200 real questions the rate of correct answers with a source, targeting above 85 %. Then expand by document batches, naming an owner per batch in charge of removing obsolete versions. A full 50,000-page project takes 4 to 6 months.
Mini case study
Need a professional website?
Kolonell builds websites that attract clients, optimized for the Sénégalese market. Free quote in 2 minutes.
Mrs. Ben Salah, quality director of a 600-agent contact centre in Tunis, manages 9,000 pages of customer procedures. An agent looks up a procedure 6 times a day at 4 minutes each, 24 minutes a day. The RAG assistant, quoted at 75,000 TND (about EUR 22,500) with usage at 0.8 TND per agent per day, brings this to 5 minutes. Gain: 19 minutes per agent per day, 190 hours a day for the site. At a loaded cost of 12 TND per hour, that is about 2,280 TND a day recovered for 480 TND of usage. The net gain of 1,800 TND per working day pays back the investment in about 42 working days, two months, and average call time drops by 25 seconds.
FAQ
Does the assistant understand Arabic, French and English?
Yes. 2026 models handle Modern Standard Arabic, French and English and answer in the language of the question. Arabic scans need suitable OCR, 10 to 15 % of the extraction budget.
Can an employee see a confidential document through the assistant?
Not if rights are inherited from the DMS or SharePoint: the assistant filters passages before sending them to the model. This workstream is 15 to 20 % of the budget and must never be cut.
Do we retrain the AI when a procedure changes?
No. RAG reindexes the updated document within minutes. Only poor chunking or a forgotten obsolete version degrades answers.
What accuracy should we target?
At least 85 % correct answers with a source on a 200-question test set before launch, then above 90 % after 3 months of corrections.
What is the minimum budget to test?
A proof of value on 2,000 pages and 30 users costs 15,000 to 25,000 TND (EUR 4,500 to 7,500) and ships in 5 to 6 weeks, with a measurement report before any decision.
Let's scope your project. Tell us your document volume, tools (SharePoint, DMS, intranet) and hosting constraints: we price a RAG assistant from 40,000 to 120,000 TND (EUR 12,000 to 36,000) with a pilot in 6 to 8 weeks. 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.
