\cleardoublepage \part{Operating AI Systems} \setbookpart{Operating AI Systems}

Part III: Operating AI Systems

Operations now runs AI systems as well as using them. Models are updated and retired by their providers on the providers' schedule, with notice periods that differ and that an emergency can shorten (147, Anthropic, n.d.); (145, Microsoft, 2026), prompts and retrieval indexes change like configuration, inference capacity is scarce and costly, and an AI incident is an operations incident with legal consequences. Part III treats these as operations questions.

Chapters in this part:

• Chapter 13, Operating AI Systems, treats drift, model updates and provider change (O1 Observe, O3 Change, O5 Sustain).
• Chapter 14, Incidents Involving AI, treats AI incidents: detection, triage and reporting (O2 Diagnose, O4 Restore).
• Chapter 15, The Economics of AI Operations, treats the economics of tokens, accelerators, latency and energy (O5 Sustain, O1 Observe).

Functions in this part. Part III concerns O1 Observe, O2 Diagnose, O3 Change, O4 Restore and O5 Sustain as they apply to AI services.

The thread from the series. The platform is that of AI Platform and Security, Chapters 6 to 23; containment of agents that exceed their mandate is that of AI Platform and Security, Chapter 19; the incidents of 2025 and 2026 are read in AI Threat.

What you will learn in this part:

• How model and prompt versions are managed as configuration
• How a model update is canaried and rolled back
• Which clocks start in an AI incident
• How the cost and latency of AI services are budgeted

Part IV turns to the operator's own AI.