Open Data Infrastructure
Foundation for AI Control Plane Architecture
A control-plane map for AI programs across catalogs, policy, metadata, lineage, evaluations, tool registries, and operational review.
AI programs do not fail only because models hallucinate. They fail because nobody controls the data path the model depends on.
AI needs a control plane
A foundation for AI is not a pile of vector databases, prompts, and model endpoints. It is a control plane that can answer basic operational questions: who can access which data, which tools are exposed, which metadata is trusted, which lineage applies, which evaluations ran, and which owner can stop the system.
Without that plane, every AI application builds its own thin governance layer. That feels fast until the company has ten different policy models and no shared answer for why an agent saw a field.
The components are not optional
The control plane needs catalogs for data identity, policy engines for decisions, metadata systems for ownership and meaning, lineage for data movement, observability for runtime behavior, evaluation systems for quality, and tool registries for action boundaries. MCP-style tool interfaces make the tool boundary more explicit, but the data layer behind those tools still has to be governed.
NIST AI RMF gives a risk-management frame. OpenLineage and DataHub-style metadata help with data movement and asset context. OPA-style policy keeps decisions outside one application code path. OpenTelemetry helps operational signals travel in a standard way.
Core idea: the AI control plane should coordinate data access, context, tools, policy, lineage, and evaluation as one operating surface.
ODI ties the plane to data
For related ODI context, read Open Data Infrastructure as the foundation for AI, data product ownership for AI, and ODI reference architecture for agentic analytics.
The control plane should not live only in the AI application tier. It has to reach into the catalog, table, policy, lineage, and serving layers. That is where the durable evidence lives.
What breaks first
- Tool registries expose capabilities without data product ownership.
- Policy decisions are implemented separately in each application.
- Evaluations test model output but ignore data freshness and permissions.
- Lineage exists for pipelines, but not for context retrieval or tool calls.
Architecture questions
Ask where policy runs, where tools are registered, where metadata is authoritative, where evaluation results live, and where a reviewer can see the whole path. If no single plane can answer, the architecture is still a set of parts.
Sources to start with
These primary and authoritative sources anchor the claims in this guide.
- Model Context Protocol specification
- Open Policy Agent documentation
- OpenLineage object model documentation
- OpenTelemetry metrics documentation
AI needs a model choice. Production AI needs a control plane.