Why it matters
Raw models will discuss anything, promise anything and occasionally invent anything: fine in a sandbox, unacceptable wearing your logo. Every public AI embarrassment you've read about was a guardrail nobody built. They're where responsibility becomes engineering.
How it works
Layers, not a switch: scoped system instructions, input filtering, grounding in approved content, output checks (topics, claims, tone), hard limits on actions (nothing irreversible without a human), logging, and escalation paths. Sized to stakes: internal drafts need less than public answers.
What to do about it
For any AI touching customers or systems, ask to see the guardrail spec before the demo. If the answer is a blank look, the project isn't ready to wear your name.
