Human in the Loop (HITL)
In practice human in the loop is an approval step. The agent drafts the email and a person sends it. The model proposes a refund and a person approves anything over a set amount. The model flags contract clauses and a lawyer reviews them. Put the checkpoint where a mistake would be expensive or hard to undo, and let the rest run on its own.
Regulation expects it for higher-risk uses. Article 14 of the EU AI Act requires high-risk AI systems to be designed so that people can oversee them, understand their limits, override their output and stop them. A review step only works if the reviewer has the time and information to disagree, and the same article names the risk of automation bias, the habit of approving whatever the system suggests. Pair human review with guardrails for the checks a machine can do, and with LLM-as-judge scoring to decide which items most need a person's attention.
Example
You prepare refund decisions for the support team.
For each ticket, return JSON: {"ticket_id", "decision": "approve" | "deny" | "needs_review", "amount", "reason"}.
Rules:
- Use "needs_review" for any refund over $200, any customer with more than two refunds this year, and any case where the policy doesn't clearly apply.
- Never issue a refund yourself. A person approves every decision before it's sent.When to use it
- Actions that are hard to undo: sending, paying, deleting, publishing
- Decisions that affect a person's money, health, job or legal position
- Early rollouts, before you have data on how often the system is wrong
When not to use it
- Low-stakes, easily reversed actions where review only adds delay
- Review steps so frequent that people approve without reading; reduce the volume or raise the threshold instead
Articles that use this
- Magentic Just Hit 1.0, and Your Whole Agent Team Lives or Dies by the Manager's Ledger PromptAgentic
- Your Agent Should Write Python, Not Emit JSON: Prompting for CodeAct Now That It Ships in Microsoft Agent FrameworkAgentic
- Gemini Spark Skills Are Just Trigger-Condition Prompts: How to Write Always-On Agent Instructions That Don't Go RogueAgentic
- Prompt Patterns That Catch MCP Tool Poisoning Before Your Agent Executes ItSecurity

