The verdict in three sentences
A 350-person firm in Amsterdam can give its staff an in-house AI assistant that answers from company documents while keeping data in Dutch data centres. Expect roughly EUR 30,000 to 70,000 to build it in 2026, plus EUR 1,000 to 3,500 a month to operate depending on the model and active users. Success hinges on permission-aware retrieval and on having your GDPR and AI Act paperwork completed before launch.
What an in-house AI assistant is, and what it must never do
The standard architecture is retrieval-augmented generation: your documents (project files, procedures, templates, support knowledge base) are split, indexed and searched, and a language model writes an answer that cites its sources. The model does not learn your data; it reads relevant extracts at question time.
The main risk sits in access control. If the index mixes HR files, client contracts and engineering notes, a junior consultant could receive an answer built from a document they cannot open. The assistant must inherit permissions from your directory (Entra ID, Google Workspace groups) and filter sources before generating anything.
| Hosting option | 2026 market ballpark (EUR per month, excl. VAT) | Strengths | Trade-offs |
|---|---|---|---|
| Open-weight model on a dedicated GPU server in the Netherlands | 1,800 to 3,500 | Full control, data stays on your servers | More hands-on operations |
| European provider API with NL or EU processing | 1,000 to 2,500 | Fast start, pay per use | Provider dependency |
| Managed private cloud with ISO 27001 hosting | 2,000 to 3,500 | Strong assurance, managed patching | Fewer model choices |
| On-premises hardware | Upfront hardware investment | Data never leaves the building | Needs in-house skills |
The build budget for 300 users
| Item | 2026 market ballpark (EUR, excl. VAT) | Comment |
|---|---|---|
| Discovery, source and permission mapping | 4,000 to 8,000 | Workshops with each department |
| Connectors (SharePoint, file shares, DMS) | 6,000 to 16,000 | Depends on number of sources |
| Indexing and permission filtering | 8,000 to 18,000 | The most sensitive work |
| Chat interface with citations | 5,000 to 10,000 | In Teams or the intranet |
| Logging, monitoring, usage dashboard | 3,000 to 8,000 | Traceability |
| Compliance: DPIA, AI Act documentation | 4,000 to 10,000 | With your DPO |
| Total | 30,000 to 70,000 | 10 to 14 weeks |
Kolonell's prices for custom applications are EUR 6,100 to 10,670 (4,000,000 to 7,000,000 FCFA) for medium complexity and EUR 10,670 to 22,870 and above (7,000,000 to 15,000,000 FCFA and above) for complex builds, with Premium maintenance at EUR 183 a month (120,000 FCFA). A multi-source assistant with fine-grained permissions usually sits in the complex band, with hosting billed directly by the Dutch provider you choose.
GDPR and the AI Act: what to document
Regulation (EU) 2024/1689, the AI Act, applies in stages. The AI literacy obligation in Article 4 already requires organisations to ensure that people using AI systems understand how they work and where they fail. A document assistant generally falls outside the high-risk uses listed in Annex III, unless it is used to assess candidates or employees, in which case the regime changes and legal review is needed.
On GDPR, your documents contain personal data. The Autoriteit Persoonsgegevens is the Dutch supervisory authority, and a DPIA is required where processing is likely to result in a high risk: let your DPO make that call early.
| Requirement | Deliverable | Owner |
|---|---|---|
| Record of processing | Entry for the assistant | DPO |
| DPIA | Signed off before go-live | DPO and CIO |
| Processor agreement | Article 28 GDPR terms with host and model provider | Legal |
| AI literacy | Usage guide and staff awareness sessions | HR and IT |
| Excluded sources | List of folders never indexed | Department heads |
| Logging | Prompt logs kept for a limited period | IT |
Mini case study
Need a professional website?
Kolonell builds websites that attract clients, optimized for the Sénégalese market. Free quote in 2 minutes.
Illustrative example: Pieter is CIO of a 350-person engineering consultancy on the Zuidas and rolls the assistant out to 300 users. Conservative assumption: half use it and save 20 minutes a week finding specs, templates or past project answers. That is 50 hours a week, about 217 hours a month, at a loaded cost of EUR 65 an hour: EUR 14,100 a month. The build costs EUR 55,000 and hosting EUR 2,500 a month. Net monthly gain is EUR 11,600, so payback arrives in under 5 months, before counting fewer tickets to the internal help desk.
FAQ
Is our data used to train the model?
Not in a properly configured retrieval setup: the model reads extracts at question time only. Get that commitment in writing, along with Dutch or EU hosting and prompt retention limited to, say, 90 days.
How many users can one GPU server handle?
A mid-range dedicated server typically handles a few dozen concurrent requests, enough for 300 users with 10% to 20% active at once. Beyond that, add a second node.
Do we need a DPIA?
Likely, once the assistant indexes HR files or client data at scale. Allow 2 to 4 weeks with your DPO and start during discovery.
Can we start with one department?
Yes, and you should: a 6-week pilot with 30 to 50 users measures real time savings before opening to 300 people.
What if we want to switch models later?
With a clean architecture, swapping the model leaves the index and permissions untouched. Budget 3 to 5 days of answer-quality testing.
Let's scope your project. List your document sources and target user count, and we will scope the build, budget and Dutch hosting costs. 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.
