AI News

The article argues that data architectures built for human analysts are not ready for autonomous agents, because agents lack the human ability to supply missing context, notice anomalies, and question unreliable values. It defines five requirements for AI-ready data: trusted, contextual, traceable, governed, and operational. The proposed architecture builds upward from data contracts and quality controls, through a context layer, to governed access, with observability running across every layer. Data contracts should enforce schemas, quality rules, and freshness SLAs in code; failed data should be quarantined before agents can access it. The same principle applies to documents and vector indexes, where stale re-embeddings, missing provenance, malformed chunks, and embedding problems can corrupt retrieval. A semantic context layer should separately define the domain model, metric calculations, and capabilities an agent may invoke, with definitions versioned, tested, and routed through instead of raw schemas. For access, the article distinguishes retrieval, real-time queries, and write-back actions. It recommends exposing read-only resources first, then carefully designed capabilities rather than one-to-one API wrappers, and using preconditions, ownership, permissions, reversibility classes, and deterministic checks for actions. Retrieved text may inform a proposal but must never authorize an action directly. Governance requires agentic lineage that records not only what was accessed but why, with traces and spans for each step, sources, reasoning, and decisions. The article recommends delegated user access, short-lived credentials, least privilege, and staged autonomy from shadow mode to supervised and guarded operation. Its practical starting points are to instrument from day one, contract critical data, build a focused context layer, and read before write.