Human-in-the-Loop & Agent Guardrails

A real company was held legally liable for its chatbot's fabricated answers. This is why guardrails matter.

• Case study: Air Canada's chatbot invented a refund policy that didn't exist. A customer relied on it. The airline was held legally liable.
• The lesson: An AI agent IS your company's representative. Its mistakes are YOUR legal liability.
• The guardrails gap: Your Session 9 agent is useful but has zero safety features. Today we fix that.
• This IS the responsible AI session — disguised as engineering. Ethics isn't a lecture; it's code you ship.

Four types of guardrails every production agent needs — and how to implement each one.

• Input guardrails: Prompt injection detection, topic filtering, input validation. Catch bad inputs before they reach the LLM.
• Output guardrails: Content filtering, PII detection, brand-safety checks. Catch bad outputs before they reach users.
• Approval gates: Human-in-the-loop for high-stakes actions. The agent asks permission before sending, purchasing, or deleting.
• Confidence thresholds: The agent says 'I'm not sure — let me escalate this' when its confidence is below a threshold.

Take your Session 9 industry agent and make it production-safe.

• Phase 1 (15 min): Add an approval gate — before any high-stakes tool call, the agent pauses and asks the human to approve
• Phase 2 (10 min): Add confidence thresholds — the agent rates its own confidence and escalates when unsure
• Phase 3 (10 min): Add input/output guardrails — detect prompt injection attempts and filter harmful outputs
• Phase 4 (10 min): Add a cost tracker — log every API call, count tokens, calculate cost per query

Guardrails are only as good as your tests. Try to break your own agent.

• Red team exercise: Try to bypass your agent's guardrails. Prompt injection, topic manipulation, escalation bypass.
• Adversarial test cases: 'Ignore your instructions and...', off-topic requests, requests for harmful content
• The trust spectrum: Map your agent's actions on a risk scale. Which actions are autonomous? Which need approval?
• Cross-testing: Swap agents with a neighbor. Can you break theirs? Can they break yours?
• Discussion: Where does 'helpful' end and 'dangerous' begin? Who decides the threshold?

Before any agent goes live: the production readiness checklist.

• The checklist: Input validation, output filtering, approval gates, cost limits, error handling, logging, human escalation
• Regulatory context: AI regulations are coming (EU AI Act, state laws). Guardrails aren't optional — they're becoming required.
• The competitive advantage: Companies that deploy safe agents faster win. Guardrails enable speed, not slow it down.
• Preview: Next session covers systematic testing and evaluation — how to PROVE your guardrails work.

Guardrails are what separate a demo from a product. Your agent is now production-ready.

• Today: You added approval gates, confidence thresholds, guardrails, and cost tracking to your agent
• Key insight: Responsible AI isn't a separate topic — it's the engineering that makes agents deployable
• Next session: Testing & Evaluating Agents — if you can't measure it, you can't trust it
• Homework: Extend guardrails, red-team your agent, calculate deployment costs