The correct option is Human in the Loop (HITL) because a licensed attorney must review, edit, and approve each AI draft before any client sees it.

Adding human oversight ensures accuracy, legal compliance, and accountability for outputs. HITL is a recommended practice for high‑stakes domains where expert judgment must validate model results before release, which matches a legal drafting workflow that requires professional approval.

Vertex AI Guardrails provides automated safety controls and content filters that help reduce harmful or out‑of‑policy responses, yet it does not require a human reviewer to approve outputs before delivery, so it is not the practice described.

Fine-tuning adapts a model using task‑specific training data to improve style or performance, but it does not add a human approval step to the generation workflow, so it does not match the scenario.

Grounding connects generation to authoritative data to improve factuality, often through retrieval of enterprise content, but it still produces automated outputs without mandatory human approval, so it is not the correct choice.

Full AWS Practitioner Certification Question