What is actually being valued
RAG and vector store value is the marginal model quality, latency, and unit economics they create against a documented baseline - not the storage cost of the index. Embeddings, chunking strategy, and evaluation sets all count.
Attribution to outcomes
Value is attributed to the data layer only when a controlled experiment, ablation, or production A/B shows a measurable shift in answer quality, deflection, or revenue per session.
Decay and refresh
Vector assets decay as upstream content and model behavior drift. The valuation carries an explicit refresh cadence and an obsolescence layer.
Questions buyers actually ask.
Only against a defined retrieval task and a controlled baseline. Without those, an embedding has no defensible standalone value.
Yes - fine-tuning corpora, preference data, and evaluation sets all sit inside the AI-native scope.
The data layer is valued; the model is treated as a substitutable execution surface with its own rights and cost profile.
Yes. The band is decision-grade and pairs with the AI investment posture in the capital-readiness brief.