3.3. Managing Cost and Responsible AI
💡 First Principle: Two kinds of "limits" protect a deployment, and they protect against different failures. Cost controls (budgets, alerts, TPM ceilings, tagging) protect against runaway spend; content safety / responsible-AI policy protects against harmful output. Confusing which control stops which failure is a common scenario trap.
Why care: production AI fails in two expensive ways — a runaway agent loop burning tokens, or a harmful generation causing reputational/compliance damage. The planning domain expects you to know which lever addresses which, and that both are configured before deployment, not bolted on after an incident.
⚠️ Common Misconception: "Content safety is something you add after the model generates a bad output." Content filtering runs on both input prompts and output completions, is configured as a reusable policy associated with the deployment, and ships on by default (though not tuned for your use case). It's a pre-deployment configuration, not a post-incident patch.
