\cleardoublepage \part{Delivery and Operations} \setbookpart{Delivery and Operations}

Part V: Delivery and Operations

Most AI initiatives fail not at the proof of concept but between a convincing demonstration and a dependable, affordable production service. Part V addresses that stretch.

Testing probabilistic systems requires a different pyramid, with golden datasets, behavioural scenarios, red teaming and human judgement layered on conventional tests (Chapter 20). Observability must detect drift and quality decay that no exception handler raises, and it supplies the logs the AI Act's record-keeping and post-market monitoring depend on (Chapter 21). Cost and energy become design constraints when inference is billed per token (Chapter 22). Scaling from pilot to production then draws these together and demands golden paths and readiness reviews rather than heroics (Chapter 23).

Procurement, which fixes much of an AI system's cost and risk early, and the organisational side of scaling are treated in AI Governance (Chapters 7, 12 and 13).

What you will learn in this part:

• How to design an evaluation strategy with acceptance criteria as code
• How to instrument AI services with OpenTelemetry, detect drift and manage incidents
• How to model, control and attribute the cost and carbon footprint of AI workloads
• Why pilots stall, and how golden paths and readiness reviews let AI scale

Part V extends the testing, cost and delivery chapters of Architecture as Code to model-based systems.