The detail
Email is one of the most common AI workloads — drafting, summarizing, classifying, and extracting from messages. A standalone 120-word email is about 160 tokens, but real emails rarely arrive standalone: signatures, disclaimers, and quoted history often triple the token count of what a human would consider "the email".
For high-volume processing the overhead matters more than the message. Stripping quoted threads and signatures before sending messages to a model routinely cuts token spend by 50-70% with no loss of information, because the model only needs the newest content plus a compact summary of the thread.
At $0.5-$5 per million input tokens, classifying or summarizing a single email costs thousandths of a cent — it is the millions-of-emails scale where tokenization discipline turns into real savings.