Copilot vs. your own AI platform — the cost math that decides your AI strategy

The 30-euro seat is a subscription to uncertainty

Microsoft Copilot works well for standard office tasks. If your primary need is summarising emails, drafting slides inside PowerPoint, and asking questions about SharePoint documents, it delivers value from day one with zero infrastructure. But that is a narrow slice of what enterprises actually need from AI — and the slice that is most expensive to scale.

The published list price sits at 30 USD, approximately 26 to 28 EUR per user per month for standard seats. Copilot Studio — the agent builder that lets you connect custom data sources — adds further cost per seat and scales with complexity. This is a per-user, per-month license that does not decrease when usage drops and has no published volume discount. Your first 500 users cost €156,000 per year. Your next 1,500 cost another €468,000. The curve never bends.

An open AI platform — LiteLLM as the model gateway, Open WebUI for the interface, pgvector for the knowledge store, vLLM for inference — runs on infrastructure you control. The model is a commodity: GPT-4-class inference costs have fallen by a factor of approximately 50 over the last three years. The value sits in the integration layer, the data pipeline, and the governance framework. Those are fixed costs that do not scale with headcount.

The math below compares three scenarios at two company sizes over three years. Every assumption is stated. The numbers are not estimates; they are published prices and calculated arithmetic.

Below 200 users, Copilot is cheaper in absolute terms. Above 200, the open platform compounds the advantage because infrastructure stays fixed while licenses scale linearly with every new employee.

What the 1.3-cent day proves

Constantin, a senior architect at Pexon, ran an afternoon of work across document drafting, code generation, data extraction, and report summarisation. The total: 1.3 cents in token spend measured at the gateway. That is the cost of a complete workday for one person, using models that match or exceed GPT-4-class quality on the tasks that matter in an enterprise.

The math is straightforward. A typical knowledge worker sends 80 emails, reads 12 documents, and drafts 3 reports per day. At an average of 2,000 tokens per task, that is roughly 256,000 tokens per day per user. At a blended token price of 0.000008 EUR (which is conservative for open models routed through LiteLLM with fallback), the daily cost comes to approximately 0.002 EUR. Even at 0.02 EUR per user per day — ten times the actual spend — the monthly cost for 1,000 employees is €60,000 in tokens against €156,000 in Copilot licenses. Include the gateway, the UI, the vector store, monitoring, and managed operations, and the total lands at roughly €5 per user per month for a 1,000-person company. For 2,000 users, the per-head cost halves because the infrastructure does not grow proportionally with headcount.

Four numbers that define the cost structure

The open AI platform is not a product you license. It is an architecture you own, operated either by your own engineers or by a managed service partner. The cost structure is fundamentally different from any seat-based model, and the differences compound over time.

• A complete workday — drafting, coding, summarising, extracting — costs 1.3 cents in measured token spend at the gateway. Models that match GPT-4-class quality run on open weights through LiteLLM with zero data retention.
• The daily cost per user sits between 0.02 and 0.05 EUR for normal office work. That includes the model call, the retrieval layer, and the governance framework. It does not include the initial platform setup, which is a one-time fixed price.
• At 1,000 employees and €3,000 per month for managed operations, the total cost per user works out to approximately €5 per month including everything. At 2,000 users, the per-head cost drops to €2.50 because infrastructure does not scale linearly.
• The platform foundation — gateway, SSO, knowledge base, guardrails, audit logging, infrastructure-as-code in your repository — is delivered as a fixed-price project of approximately €24,500. Every piece of code, every configuration, every deployment script belongs to you from day one.

The real decision: Windows or Linux?

Every AI platform purchase for the enterprise falls into one of three categories, and the third category exists only to capture margin from companies that want the appearance of Linux without the substance.

The first category is the closed ecosystem: Microsoft Copilot, Anthropic's paid tiers, OpenAI's enterprise offers. These deliver deep integration, zero setup, and an ecosystem that works immediately. The cost is a per-user seat price, no model choice, and a roadmap controlled by someone else. You are buying convenience and accepting vendor lock-in as the price.

The second category is the open stack: LiteLLM, Open WebUI, Hermes, n8n, LangGraph, pgvector, vLLM, NeMo Guardrails. Everything is built from established open-source components, connected through open interfaces, and operated by you or by a managed service partner. The cost is infrastructure and engineering — both fixed, both scale sub-linearly. You own the code, the data, and the exit.

The third category is the clones: Langdock, MyGPT, and the other German ChatUI vendors. They charge a seat price similar to Windows but deliver neither the integration depth of Microsoft nor the cost advantage of Linux. They are, in the words of the position document, Windows without Windows.

The sentence that sorts this for the data protection officer is the simplest possible. You can take Windows, or you can take Linux. But why should you take a clone of Windows?

Why this matters for indexation and search: the open platform is the only option where every piece of knowledge you build — skills, agents, prompts, workflows — lives in your own repository as code. At the end of every engagement, you receive complete code transfer and a documented exit clause. No proprietary SDK, no session-based licensing, no black-box API.

The 200-user threshold belongs on the first slide

State it yourself before the CFO finds it. At fewer than 200 employees, a seat-based model like Copilot is cheaper in absolute terms. The open platform becomes economically superior at roughly 200 users for managed service and at roughly 100 for self-operation. Above that line, the advantage compounds because infrastructure is fixed while licenses scale linearly. Sovereignty is not a cost item. It is the reason the math is worth looking at.

The Liebherr reference: 0.38 EUR per user per month for managed operations on a live 600-user platform — the real number behind the TCO model, not the theoretical projection.

Where Copilot is the right answer

Copilot is a tool, not a platform. For standard cases inside the Microsoft 365 context — drafting emails, summarising Teams meetings, generating slides — it works exceptionally well and requires zero setup. If your organisation has fewer than 100 employees and its AI needs stop at the M365 boundary, Copilot is the correct choice and should be bought without hesitation.

The moment you need AI on your own data — contract analysis, customer support automation, regulatory compliance checking, supplier document processing — Copilot Studio becomes a partial solution. It works, but it adds cost per seat, ties you to the Microsoft model ecosystem, and gives you no model choice. For orchestration you need Azure AI Foundry. For the data layer, Fabric. The seat price is a partial purchase, not a platform purchase.

And you still have no model choice. When a better or cheaper model arrives tomorrow — and they arrive quarterly — you are locked into Microsoft's roadmap. With the open platform, a model swap is a configuration change, not a project.

How to start without a platform project

The most common objection to the open platform is the perception that it is an 18-month project. It is not. The recommended entry point is a single use-case pilot over 6 to 8 weeks at a fixed price of €15,000. You pick one concrete workflow — document processing, internal research, knowledge retrieval — and prove cost savings, quality, and user acceptance against measurable baselines. The platform foundation happens underneath it, not before it.

The 90-day engagement breaks into three phases. Weeks 1 to 4: architecture and gateway foundation plus security concept with your IT, use-case selection, and baseline measurement. Weeks 4 to 10: the first productive use-case with real users, measured on quality, cost, and acceptance. Weeks 10 to 13: a decision package based on actual measurement data rather than assumptions.

At the end of the pilot, you have real numbers. Your team has used the tool. Your CFO has seen the cost delta. The platform decision becomes an extension, not a starting point.

# Minimal gateway configuration — LiteLLM proxy, open-source.
# Every model call passes through this point, where cost, budget,
# routing, fallback, and zero-data-retention are enforced.

# models:
#   - model_name: anthropic-claude-sonnet-4
#     litellm_params:
#       model: anthropic/claude-sonnet-4-latest
#       api_key: $ANTHROPIC_API_KEY
#
#   - model_name: deepseek-chat
#     litellm_params:
#       model: deepseek/deepseek-chat
#       api_key: $DEEPSEEK_API_KEY
#
#   - model_name: self-hosted-qwen3-32b
#     litellm_params:
#       model: vllm/qwen3-32b
#       api_base: http://inference:8000/v1
#       max_tokens: 8192
#       temperature: 0.7
#
# Cost control:
#   - Daily budget per user: 0.05 EUR
#   - Hard monthly budget: 2.0 EUR per user
#   - Fallback: self-hosted Qwen3 when external quota exceeds 80%
#   - Zero data retention: enforced at gateway level, never optional

The four things you own

The open platform gives you four assets that no seat license ever can: the code (every script, every configuration, every deployment pipeline in your repository), the data (your documents stay in your vector store, never leave your tenant), the model choice (switch between Anthropic, DeepSeek, Qwen, Mistral, or any OpenAI-compatible endpoint without changing the integration layer), and the exit (documented, tested, and ready to execute on day one).

Where to go next

• LiteLLM: the open model gateway that replaces Copilot Studio
• Sovereign private AI: open weights deployed in your own data centre
• Enterprise RAG platform with permission-aware retrieval
• EU hosting and GDPR: zero data retention architecture