Superhuman's AI drafts 60% unedited — and we still can't trust it

email client superhuman released a new version of its auto-draft feature, utilizing a mix of frontier models from openai and anthropic to generate replies that mimic user tone. beta testing data indicates that 40% of drafts were sent within 24 hours, and 60% of those were sent without any edits. the system learns from user behavior over time to correct mistakes, such as refusing unreasonable meeting times, but still requires oversight. this shift moves beyond older gpt-3.5 implementations to offer higher intelligence, signaling a trend where ai agents handle increasing amounts of routine communication.