The verdict in three sentences
An AI assistant on your internal knowledge base (a RAG architecture, retrieval-augmented generation) costs EUR 20,000 to 60,000 (about GBP 17,000 to 52,000) for a 300-employee London company, plus EUR 200 to 1,500 a month for inference and hosting. Its value rests on two non-negotiable requirements: every answer cites its sources and respects per-document access rights, otherwise the tool is neither reliable nor compliant. With 30 minutes saved per employee per week, the project, delivered in 2 to 4 months, pays back in well under a year of use.
What an enterprise RAG project covers
The operations director sees teams spending a lot of time hunting for the right version of a procedure in SharePoint, Google Drive or the intranet: 15,000 documents, duplicates, scanned PDFs, contradictory memos. A RAG assistant does not replace the knowledge base: it indexes it, retrieves the relevant passages and writes a short answer with links to the original documents.
The building blocks of a serious project:
- Connectors to sources (SharePoint, Google Drive, Confluence, intranet, DMS).
- Ingestion and chunking, OCR for scans, version handling.
- Vector index and hybrid search (semantic plus keyword) to find the right passages.
- Access control: an employee never sees an excerpt from a document they cannot open.
- Interface in Teams, Slack or a web portal, with clickable citations.
- Evaluation: a set of 200 reference questions to measure quality before and after every change.
| Item | 2026 range (EUR, excl. VAT) | Note |
|---|---|---|
| Scoping, document audit, question set | 3,000 to 7,000 | Find the 20% of documents behind 80% of questions |
| Connectors and ingestion (2 to 4 sources) | 4,000 to 12,000 | OCR, versions, nightly sync |
| Index, hybrid search, tuning | 4,000 to 12,000 | The core of answer quality |
| Per-document rights management | 3,000 to 10,000 | Sync with directory (Entra ID, Google) |
| Teams or web interface with sources | 3,000 to 10,000 | Thumbs up or down feedback |
| Evaluation, acceptance, change support | 3,000 to 9,000 | 30-user pilot before rollout |
Running costs and hosting choices
Inference cost depends on question volume and model. For 300 employees asking 5 questions a day on average, expect EUR 300 to 900 a month with a strong model. UK or EU hosting is standard for HR, legal or client documents.
| Option | Indicative monthly cost | Advantages | Watch points |
|---|---|---|---|
| Hosted model API in UK/EU | EUR 200 to 900 | High quality, no infrastructure | No-training data agreement |
| Open-source model on UK/EU cloud | EUR 600 to 1,500 | Full control, data in your tenant | Slightly lower quality, maintenance |
| Generic office-suite copilot | EUR 25 to 35 per user | Instant rollout | EUR 7,500 to 10,500/month for 300 staff, little control over sources |
| Custom RAG assistant | EUR 200 to 1,500 plus EUR 500 to 1,200 maintenance | Focused on your procedures, reliable citations | 2 to 4 month project |
The comparison with a per-user generic copilot is often decisive: for 300 people, the licence alone exceeds the full cost of a custom project within a year.
Pitfalls to avoid
- Indexing everything unsorted: outdated versions produce wrong answers. An audit archiving 30 to 40% of documents improves quality more than a pricier model.
- Forgetting access rights: without per-document filtering, an intern could retrieve an excerpt of a pay agreement. This is the first thing the ICO and your DPO will check.
- Not measuring: without a reference question set, you cannot tell whether an update helps or hurts.
- Launching without a pilot: 30 users for 3 weeks are enough to spot documentation gaps.
Mini case study
Need a professional website?
Kolonell builds websites that attract clients, optimized for the Sénégalese market. Free quote in 2 minutes.
Oliver, operations director of a 300-employee services firm in London, estimates each employee loses at least 30 minutes a week looking for procedures. Over 45 working weeks, that is 300 × 0.5 × 45 = 6,750 hours a year. At an average loaded cost of EUR 45 an hour, the loss reaches about EUR 304,000 a year.
The project is priced at EUR 45,000, plus EUR 800 a month for inference and EUR 900 a month for maintenance, about EUR 20,400 a year. Even if the assistant recovers only half the lost time (EUR 152,000), the first-year cost (EUR 65,400) is covered in just over 5 months.
FAQ
How much does a RAG project cost for a 300-employee company?
EUR 20,000 to 60,000 depending on the number of sources and rights complexity, plus EUR 200 to 1,500 a month for inference. A single-source pilot starts around EUR 12,000.
Does the data stay in the UK or EU?
Yes, the index and models are hosted in the UK or EU, with a contract forbidding reuse of data for training. This fits UK GDPR and typical DPO requirements.
Can the assistant make up answers?
The risk exists but is reduced: the assistant answers only from retrieved passages and cites its sources. On a 200-question set, we target over 90% correct, sourced answers before rollout.
How long does it take?
2 months for one or two sources with simple rights, 4 months for several sources, a complex directory and a full pilot.
Should we clean up the knowledge base first?
Strongly recommended. Archiving duplicates and outdated versions (often 30 to 40% of volume) clearly improves answer accuracy.
Let's scope your project. Tell us your document sources, user count and access constraints, and we will price a RAG assistant between EUR 20,000 and 60,000 with a 2 to 4 month timeline. 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.