Production agentic-AI costs at scale routinely run multiples of POC projections, and a layered optimisation programme covering model tiering, vendor prompt caching, batch APIs, context-window discipline, and observability budgeting closes most of the gap.

Correction log

1. 28 Apr 2026Rewritten 27-28 Apr 2026 from 27 Jul 2025 WordPress-migrated original. Original used a fictional CTO scene (Marcus Chen, $4.2B logistics company, 9:47 AM Tuesday Seattle), fabricated case figures ($2.1M to $187K monthly, named-company before/after teardowns), fabricated expert quotes (Patricia Williams VP of Engineering at Walmart; David Park Principal at Goldman Sachs), and banned phrases (plot twist, the dirty secret, revolutionary, emoji subheads). Rewrite extracts the verifiable cost-driver categories with primary-source citations from Anthropic's published multi-agent token-ratio research, vendor prompt caching and batch-API pricing pages, McKinsey State of AI, Andreessen Horowitz on LLM inference economics, and Gartner's April 2026 I&O finding. Approved + published 28 Apr 2026.