3.1.3. Responsible AI, Security, and Governance
💡 First Principle: Generative systems fail probabilistically, so trust must be engineered — a layered defense where the platform moderates content, your configuration constrains behavior, governance policies bound what any maker can wire up, and humans stay reachable for what the agent shouldn't decide alone.
Plan the layers explicitly rather than assuming the platform handles it:
| Layer | Controls | Planner's question |
|---|---|---|
| Content behavior | Moderation levels on generative answers, custom instructions, scoped knowledge | How wrong can an answer be, and how do we bound it? |
| Transparency | Disclosure that users talk to AI, citations, confidence behavior | Does the user know what they're talking to and where answers come from? |
| Human oversight | Escalation/handoff paths, approvals in flows (3.2.2), computer-use supervision | What must a human see before it happens — or after it fails? |
| Platform governance | Environment strategy, DLP policies on connectors, admin center visibility and controls, sharing limits | What can makers build and share at all — and who can see that it happened? |
DLP deserves its own sentence because it binds Phase 2 to this section: data loss prevention policies classify connectors and can block combinations outright, so a tool that works in dev may be policy-impossible in production — a governance check that belongs in planning, not in deployment week. Similarly, environment strategy (which environments exist, who makes in them, what data classes they touch) is the container every later ALM decision (Phase 4) lives inside.
⚠️ Common Misconception: Responsible AI is handled automatically by the platform. The platform supplies the controls — moderation, DLP, audit surfaces, supervision hooks — but an accountable human must choose, configure, and monitor them. "We enabled generative AI, so compliance is covered" is the exam's favorite wrong assumption.
Reflection Question: For a financial-advice-adjacent agent, walk the four layers and name one concrete control per layer you'd insist on before launch — and one you'd monitor forever after.
