Building an AI Email Assistant That Does Not Hallucinate Your Calendar Away

We have built seven AI-powered email assistants in the past eighteen months. Five are still in production. Two were shut down within weeks. The difference came down to one principle: never let the AI act without a human-reviewable checkpoint.

The two failures shared the same flaw. The client wanted full autonomy: email arrives, AI reads it, AI drafts a response, AI sends it. No human in the loop. One system confidently scheduled a meeting for Saturday at 3am because it misinterpreted a timezone abbreviation. The other responded to a frustrated customer with a cheerful tone that read as dismissive. Both required extensive damage control.

Our current architecture has four stages. Stage one is classification using Gemini Flash Lite -- categorizing emails as needs-response, informational, spam, or urgent. Accuracy is around 95 percent after tuning. Stage two is draft generation using Claude Sonnet, with the full email thread, CRM context, and the client's communication style guide. Stage three is the review queue where every draft gets human approval. This adds thirty seconds to two minutes per email but eliminates catastrophic failures entirely. Stage four is learning -- every human edit gets logged and used to improve the prompts weekly.

The economics work. Our clients are executives spending two to three hours daily on email. The system reduces that to thirty to forty-five minutes of review time. At an executive's hourly rate, it pays for itself in the first month.

Key technical lessons: thread context is critical because single-email responses miss context and sound robotic. CRM integration is non-negotiable because the AI needs relationship history. Timezone handling is a minefield, so we now explicitly convert all times to the sender's timezone with a confirmation phrase.

Our recommendation: start with the review queue. Always. The temptation to go fully autonomous is strong, but testing with synthetic emails does not capture the long tail of real-world ambiguity. The review queue is your safety net. Remove it only after months of production data prove the AI handles your specific email landscape without correction.