The verdict in three sentences
A RAG AI assistant (retrieval-augmented generation) connected to 15 years of reports, design calculations and standards costs EUR 15,000-40,000 to build and EUR 200-1,500 per month to run for an 80-engineer firm. In Amsterdam, the deciding question is security: EU hosting, per-project access rights inherited from the directory and strict handling of client-confidential material under the AI Act and GDPR. With 30 to 60 minutes saved per engineer per week, the project pays back in under 6 months, provided answer quality is measured on a set of real engineering questions before rollout.
Copilot, ChatGPT or custom RAG: the 2026 options
Engineers lose time finding a 2014 test report, the rationale behind a material choice or the right revision of a quality procedure. Generic assistants cannot see these archives, often stored on file servers, in an old document management system or as scanned PDFs. Prices below are 2026 estimates.
| Option | Indicative cost | Data hosting | Per-project rights | Main limit |
|---|---|---|---|---|
| Microsoft 365 Copilot | EUR 30 per user per month, EUR 28,800 per year for 80 | EU (EU Data Boundary) | inherited from SharePoint | sees only M365, weak on scanned PDFs |
| ChatGPT Enterprise or similar | EUR 40-60 per user per month (estimate) | EU residency possible | limited | manual uploads, no archive indexing |
| Custom RAG, model via EU-hosted API | EUR 15,000-40,000 plus EUR 200-1,500 per month | EU (Netherlands or France) | yes, synced with the directory | project to manage |
| On-premises RAG with open-weight model | EUR 40,000-80,000 including GPU servers | in your building | yes | heavier maintenance |
| Classic enterprise search engine | EUR 5,000-15,000 | your choice | yes | no written answers |
Custom RAG does not retrain a model: it splits your documents, indexes them, retrieves relevant passages for each question and asks the model to write an answer citing its sources (document, page, revision). The engineer checks in one click.
Where the budget goes in a EUR 30,000 project
| Work package | Share | Amount | Content |
|---|---|---|---|
| Scoping and source mapping | 10% | EUR 3,000 | inventory, classification, exclusions |
| Ingestion and OCR | 25% | EUR 7,500 | reports, scanned PDFs, tables, drawings |
| Indexing and hybrid search | 20% | EUR 6,000 | chunking, vectors, keywords, metadata |
| Per-project rights and SSO | 15% | EUR 4,500 | Entra ID or LDAP, query-time filtering |
| Interface and citations | 15% | EUR 4,500 | chat, sources, history, user feedback |
| Evaluation and acceptance | 15% | EUR 4,500 | 200 engineering questions, accuracy rate |
Running costs depend on question volume and the chosen model.
| Usage profile | Questions per month | Model cost | Index hosting | Monthly total |
|---|---|---|---|---|
| Pilot, 15 engineers | 1,500 | EUR 60 | EUR 150 | EUR 210 |
| 80 engineers, moderate use | 8,000 | EUR 320 | EUR 250 | EUR 570 |
| 80 engineers, heavy use | 20,000 | EUR 800 | EUR 300 | EUR 1,100 |
| Heavy use, premium model | 20,000 | EUR 1,200 | EUR 300 | EUR 1,500 |
Security, client confidentiality and the AI Act
For a Dutch engineering consultancy working on infrastructure, water and energy projects, three topics must be settled during scoping. First, classified or defence-related material and export-controlled data (EU dual-use regulation 2021/821) stays out of the index or on an on-premises instance. Second, client NDAs with public bodies and industrial clients require per-project filtering: an engineer must never receive an excerpt from a project they cannot access. Third, purchased standards (NEN, ISO, EN) are licensed: check that your subscription allows internal indexing.
Under the AI Act, an internal document assistant is minimal risk. AI literacy obligations (user training) have applied since February 2025 and most of the regulation since 2 August 2026. Require contractually that your data is never used to train the model, and keep question logs for 12 months for audits, in line with guidance from the Dutch data protection authority (AP).
Mini case study
Need a professional website?
Kolonell builds websites that attract clients, optimized for the Sénégalese market. Free quote in 2 minutes.
Pieter is technical director of an 80-engineer consultancy in Amsterdam. Central assumption: 45 minutes saved per engineer per week over 44 working weeks, so 2,640 hours per year. At a loaded hourly cost of EUR 85, that is EUR 224,400 per year. Cautious assumption at 30 minutes: 1,760 hours, or EUR 149,600.
Year-one cost: a EUR 30,000 RAG project and moderate use at EUR 600 per month, so EUR 37,200. Even with the cautious assumption, payback comes in 3 to 4 months. Copilot for 80 people would cost EUR 28,800 per year without reaching the file server archives.
FAQ
Can the assistant make up answers?
The risk exists, hence three safeguards: mandatory source citations, refusal to answer without a relevant passage and evaluation on 200 engineering questions. A well-tuned project reaches 85 to 95% correct, sourced answers on that test set.
Is our data used to train the model?
No, with a professional EU-hosted API the contract excludes training on your data and retention rarely exceeds 30 days. For the most sensitive projects, an on-premises open-weight model keeps all data inside your network.
How long does the project take?
Allow 8 to 14 weeks: 2 weeks of scoping, 4 to 8 weeks of ingestion and development, 2 to 4 weeks of pilot with 15 engineers. Rollout to all 80 users follows in 1 to 2 weeks.
How are per-project rights handled?
Each indexed document carries the access groups from your directory (Entra ID or LDAP). Filtering happens before retrieval, so the model never receives a passage the user is not entitled to see.
What does the AI Act require for internal use?
An internal document assistant is minimal risk: no CE marking or conformity assessment. You must train users and document the tool, about 1 to 2 days of work.
Let's scope your project. Describe your document sources, the number of engineers and your confidentiality constraints: we will price a RAG pilot from EUR 15,000, delivered in 8 to 14 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.
