“Implement a human in the loop. Especially in high-stakes environments such as healthcare or legal advice, requiring human review before responses are returned to the user can prevent errors from spreading.”

In Mata v. Avianca: the attorney used LLM output directly in a filing without independent verification. In Air Canada: the chatbot’s policy answers went live without any human reviewing them against the actual policy. In both cases, a single human check at the right point would have caught the error before it caused harm. Neither system had a review gate.

→ Map your workflows. Identify every path where LLM output can reach a regulated or high-stakes context: legal filings, medical records, financial reports, safety instructions
→ Add a mandatory checkpoint before any of those endpoints: output must be reviewed by a qualified human before it is acted on
→ “Human in the loop” means review before action — not review after the fact when harm has already occurred

Walk every workflow path from user input to output action. If any path reaches a legal filing, a medical record, a financial instrument, or a safety system with zero human review steps: that path is the gap. Flag it, require review, and confirm the process works in practice.