Volume V asserts that ambient inference approaches conversational spend while producing a fraction of the recognised value. That is either your largest exposure or a non-issue, and only your own numbers decide which. Every figure below is editable, including the prices.
Illustrative defaults. Replace them with your contract rates — the shape of the answer matters more than the absolute number, but only your rates make the absolute number yours.
Every lever below the third is a quality decision wearing a finance label. Run the labelled set before and after, and record the quality delta next to the saving in the same document — a cost reduction with no measured quality cost has not been measured.
Storage is almost never where the money is, and ambient inference almost always is. Suggestions nobody asked for were a quarter of this bill — invisible until someone runs exactly this query, and fixable with one routing decision rather than an architecture change.