AI readiness depends on the data foundation

The constraint is the context, not the model

Frontier models are widely available infrastructure. What is specific to your organization is whether you can tell an agent, in machine-readable form, what "active study", "customer", or "revenue" means, who may see which attribute, and how fresh the answer is.

An agent pointed at an undocumented physical schema does not fail visibly. It returns confident, well-formatted, incorrect answers, which is harder to detect than an error.

What AI readiness means in practice

Machine-consumable metadata that can be queried at execution time. Column-level lineage that runs end to end. Quality expectations learned from observed distributions rather than assumptions encoded years earlier. Mastered entities so the agent resolves to one customer. A semantic layer where each metric is defined once. And access control inherited from the source, so retrieval cannot surface what the user could not otherwise open.

Permission leakage through vector stores is the most common security gap I see in enterprise RAG. Documents get embedded and their access controls do not travel with them.

Sequencing

Build the trust plane before the agent plane. Organizations adopt AI-assisted authoring roughly three times faster than AI-assisted verification, which means defects arrive faster too. A roadmap that places the semantic layer in a later phase will need revisiting.