Is the data ready?

Data readiness is the question "can this data, as it exists today, support the answers we are promising?" asked per source and per question, and answered with numbers. For an agent it breaks into seven dimensions: is the data there (availability), can the agent legally and technically reach it (access), is it recent enough (freshness), is it complete and valid (quality), does it join to the other sources it needs (joinability), does everyone agree what it means (semantics), and is that meaning written down where a model can read it (documentation). Warehouses and lakehouses are where you usually fix what fails: you build curated views at a declared grain, encode the definitions in a semantic layer, and give the agent its own governed, metered compute. You never point the agent at raw or operational tables and hope the prompt makes up for it.

• Readiness is relative to a question set. Data can be ready for "how many deposits mature this month?" and not ready for "which of my clients have one?".
• Score each dimension pass or fail. Do not average. A source with perfect quality and no access approval is not 86% ready; it is blocked.
• Fix data in data, not in prompts. Decoding, deduplication, conflict rules and freshness stamps belong in the curated view or upstream, where they are tested once.
• The semantic layer is where business meaning becomes executable. A metric defined once and signed by its owner beats a paragraph in a system prompt.
• Agent compute is a workload like any other. It needs its own warehouse, a statement timeout, a quota and a cost line.