A team is building an AI agent that drafts personalized marketing emails.

Bis correct — for tasks involvingtone, style, brand alignment, and personalization,model-graded evaluationis the best choice.

UiPath’sagent evaluation frameworksupports multiple types of evaluation:

Model-graded: LLMs score or classify outputs based on nuanced criteria (e.g., tone match, relevance)

Human-graded: For subjective tasks

Deterministic: For strict accuracy checks (e.g., regex, classification)

In creative tasks likeemail drafting, deterministic methods (D) or length-based metrics (A)fail to capture nuance.

A/B testing (C) is useful in live experiments, but not for structured evaluation during development.

Model-graded evaluations enablescalable quality checksfor outputs that mustfeel human, on-brand, and context-aware— essential for personalized communication.