The biggest reason enterprise AI portfolios grow out of control is the absence of a shared architecture. Each team picks its own platform, its own LLM interface, its own vector store and its own governance model. Two years later the company has 200 solutions that cannot be compared or composed.
L1 contains operational systems (ERP, CRM, core systems), IoT streams and third-party enrichment data. This is the enterprise "truth" in the form in which it is born. L1 owners are typically the business or IT system owners. A good AI architecture does not try to change L1. It recognises source-system boundaries and builds controlled contracts on top.
L2 is the direct copy of source systems into a data lake or lakehouse. No transformations, no quality control — just stored and versioned raw data for traceability. This layer enables "honest" use cases and root-cause analysis.
In L3 raw data is transformed into core data products: each is an owned, documented, quality-assured and SLA-bound unit representing one business entity (customer, contract, product, transaction). Core data products are the foundation of everything above. In practice a single business unit gets by with 15–40 core data products when each is disciplined to one business entity. At this scale, governance and ownership stay tractable.
L4 is the business-logic layer. Here multiple core data products are combined, and purpose-specific metrics and entities are computed: "customer lifetime value", "order forecast for week X", "credit risk for segment A". The semantic layer is the intelligent boundary of AI. BI reports, copilots and autonomous agents all live on top of it — using the same data product.
L5 holds all models: foundation models (GPT, Gemini, Claude), domain-fine-tuned models, classical ML (regression, gradient boosting), time series and image/audio. What matters is that the model layer is standardized — one prompt registry, one model registry, one versioning mechanism. Without this every team picks its own.