Email drafting

When to use this scenario

Automated email drafting spans two very different workloads: real-time single-message composition (sales reps, customer success) and batch triage-and-draft pipelines (support queues, newsletter personalization). The distinction matters because batch jobs tolerate 10–30s latency in exchange for deep batch pricing discounts, while real-time composition must respond in under 2 seconds.

For high-volume batch drafting — a support team processing 1M inbound emails per month — Gemini 2.5 Flash at roughly $0.075/million input tokens costs $45/month for input alone. The same workload on GPT-5 ($1.25/million) is $750. Quality differences for standard email triage are marginal at that cost gap.

The critical quality axis is tone fidelity, not factual accuracy. Evaluate models on whether they preserve brand voice, handle ambiguous sentiment correctly, and de-escalate angry customers without sounding scripted.

Common pitfalls

• Optimizing for cost per email without measuring re-draft rate — a cheaper model that requires 30% human rewrites often costs more in total
• Using the same model for outbound sales (requires creativity) and support triage (requires consistency) — different optimization targets
• Skipping a human-review layer for emails that trigger refunds, credits, or policy exceptions
• Underestimating prompt length: a full email thread context plus brand guidelines plus output format instructions often runs 400–800 tokens before the actual email body