Data layer
Sources, data preparation, historisation, access rights — the basis of every analysis and every model.
Why architecture comes before the first model
Many AI initiatives start with a tool and end as an isolated solution: separate logins, separate data, separate costs. With the second and third use case, this repeats. A target architecture defines once how data, models and applications interact — and makes every further use case cheaper and safer.
For us, architecture planning does not mean building everything at once. It defines the framework within which implementation proceeds step by step.
The layers of an AI architecture
We structure the target architecture into five layers that are decided separately yet considered together.
Model layer
Selection of language, vision and forecasting models; frontier providers, European providers or open models on premise.
Integration layer
Interfaces to ERP, CRM, MES and specialist applications; orchestration of workflows and agents.
Application layer
Assistants, automations and dashboards that employees and customers work with.
Operation & governance
Monitoring, cost control, model maintenance, roles and responsibilities, AI policy.
On premise, cloud or hybrid
We do not decide the question of where to operate ideologically but with the Sellium TCO model: token volume and deployment routing, electricity price, PUE and hardware lifecycle, lease or buy — set against requirements for data sovereignty, latency and scaling.
The result is often hybrid: sensitive processes with your own models in-house, non-critical tasks via external providers with clear contract terms.
- Existing hardware and available data centre space are factored in
- Routing rules define which use case may use which model
- Exit scenarios: switching providers remains technically possible
Security and compliance by design
Certification constraints such as ISO 27001, TISAX or BSI baseline protection flow into the architecture instead of being retrofitted later. That includes network segmentation, identity and access management, logging and a clear approach to personal data.
The EU AI Act requires different evidence per risk class. The architecture provides for documentation, logging and human oversight so that high-risk applications can be met later without rebuilding.
What stands at the end of architecture planning
A documented target picture with decision rationales, a staged roadmap and the cost view over the lifecycle — as the basis for investment decisions, tenders and implementation with us or your existing IT partners.
- Target architecture with layers, components and interfaces
- Model and provider selection with routing rules
- TCO comparison of operating variants
- Security and compliance concept
- Implementation roadmap (1–3, 4–6, 6–12 months)