A healthcare-AI-agent glossary for engineers (what these terms mean in the code)

Half of "agentic AI" vocabulary is marketing. Here's what the terms map to in an actual implementation.

• Agent — a workflow with tool-use, retrieval, and a control plane, not a single prompt. It reads and acts; a chatbot just returns text.
• Human-in-the-loop — a durable interrupt in the workflow: consequential actions halt, route to an approver, and resume after sign-off.
• RAG / grounding — retrieval feeds the prompt and a citation is attached per claim; a decision without a source is rejected at the output layer.
• PHI redaction — a detect/redact step on the input path, before prompt assembly — order matters, not the checkbox.
• Audit trail — append-only event log: request → retrieval → decision → citation → approver; exportable.
• Governed AI — the sum of the above enforced as invariants, not a wrapper.
• On-prem / air-gapped — same governance model, deployed inside the customer network.
• Escalation / confidence threshold — below-bar results short-circuit to needs_human with a structured reason.

Names aside, the architecture is the point. We ship it as governed healthcare AI agents at IntelliBooks Studio — more at intellibooks.ai/overview.