The verdict in three sentences
A group that refuses to expose procedures, contracts and HR data to a consumer AI has three realistic options in 2026: Microsoft 365 Copilot, a Mistral model hosted in France, or a custom RAG assistant plugged into its own sources. Copilot is quick to deploy but becomes expensive beyond 200 licences and remains tied to US cloud providers. For 1,000 employees, a custom assistant with sovereign hosting costs three to five times less per year, provided access rights are handled seriously.
Copilot, Mistral, custom: 2026 cost comparison
Assumption: a group of 1,000 employees, 400 active users, 120,000 indexed documents (procedures, contracts, internal memos, HR records).
| Option | Upfront cost | Estimated annual cost | Data hosting | Document permissions respected | Customisation |
|---|---|---|---|---|---|
| Microsoft 365 Copilot (1,000 licences) | EUR 5,000 to 20,000 scoping | about EUR 330,000 (USD 30 x 12 x 1,000) | Microsoft cloud, EU region | yes, inherits SharePoint | low |
| Copilot limited to 400 users | EUR 5,000 to 20,000 | about EUR 132,000 | Microsoft cloud, EU region | yes | low |
| Mistral Le Chat Enterprise | EUR 3,000 to 15,000 | EUR 15 to 25 per user per month, i.e. EUR 72,000 to 120,000 for 400 | France or on-premise | partial, depends on connectors | medium |
| Custom RAG assistant + Mistral API hosted in France | EUR 25,000 to 70,000 | EUR 40,000 to 90,000 (inference, hosting, maintenance) | France, SecNumCloud possible | yes, filtering by AD group | full |
| Custom RAG assistant + open-source model on dedicated GPUs | EUR 40,000 to 90,000 | EUR 60,000 to 140,000 (2 to 4 rented GPUs) | your servers or trusted cloud | yes | full |
Inference costs often surprise on the upside: 400 users asking 15 questions a day with 6,000 tokens of context come to about 800 million tokens a month. With a mid-sized model billed at EUR 0.5 to 2 per million tokens, the monthly bill runs between EUR 400 and 1,600.
What a serious internal assistant must include
| Component | Role | Indicative budget | Timeline |
|---|---|---|---|
| Source connectors | SharePoint, file servers, Confluence, DMS, ERP | EUR 5,000 to 15,000 | 2 to 4 weeks |
| Indexing and chunking | OCR, table extraction, incremental updates | EUR 4,000 to 10,000 | 2 weeks |
| Permission management | Results filtered by Active Directory or Entra ID groups | EUR 4,000 to 12,000 | 2 to 3 weeks |
| Interface and citations | Web and Teams chat, cited sources linking to the document | EUR 4,000 to 10,000 | 2 weeks |
| Security and logging | SSO, encryption, query logs, masking of sensitive data | EUR 3,000 to 10,000 | 1 to 2 weeks |
| Evaluation and quality | 200-question test set, accuracy rate, user feedback | EUR 3,000 to 8,000 | 2 weeks |
What sinks most projects is not the model but document permissions. An assistant that answers a salesperson with an excerpt from an employment contract or a board memo destroys trust in a single day. Filtering must happen before generation, at retrieval level, never afterwards.
Sovereignty and compliance: the real questions to ask
The US Cloud Act applies to US providers even when data is stored in Europe. For sensitive data (health, defence, HR data, trade secrets), favour a host qualified SecNumCloud (OVHcloud, Outscale, Cloud Temple), whose premium is around 15 to 30% over a standard cloud. GDPR also requires a data protection impact assessment, and the EU AI Act mainly asks for transparency and user training for internal use.
Mini case study
Nathalie, CIO of a 1,000-employee industrial group in Lyon, compares two scenarios over three years. Copilot for 400 users: EUR 132,000 a year, i.e. EUR 396,000 over three years. Custom RAG assistant: EUR 55,000 build plus EUR 60,000 a year (inference EUR 12,000, SecNumCloud hosting EUR 18,000, maintenance and improvements EUR 30,000), i.e. EUR 235,000 over three years. Difference: EUR 161,000, with access open to all 1,000 employees at no per-licence cost. If each active user saves 20 minutes a week searching for procedures, 400 x 20 min x 45 weeks = 6,000 hours a year, about EUR 270,000 of time valued at EUR 45 an hour.
Need a professional website?
Kolonell builds websites that attract clients, optimized for the Sénégalese market. Free quote in 2 minutes.
FAQ
Is a custom assistant as good as Copilot?
On internal documents, often better, because indexing is tailored to your formats and sources. In 2026, mid-sized Mistral models reach 85 to 92% correct answers on a well-built business test set.
Is our data used to train the model?
No, neither with Mistral's enterprise API nor with an open-source model hosted on your infrastructure. The contract must state it explicitly, with log retention capped at 30 days.
How long does a first pilot take?
Six to eight weeks for a pilot on 2 to 3 sources and 50 users. Company-wide rollout follows within 2 to 3 months.
Can it be integrated into Microsoft Teams?
Yes, the assistant can run as a Teams bot with Entra ID authentication, for EUR 3,000 to 6,000 of extra development.
Who maintains the assistant after delivery?
A maintenance contract of EUR 1,500 to 3,000 a month covers monitoring, model updates, new sources and answer quality tracking.
Let's scope your project. Describe your document sources, user count and security constraints: we will price a sovereign pilot between EUR 25,000 and 70,000, delivered 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.
