Email remains the backbone of professional communication. Despite the rise of Slack, Teams, and a dozen other messaging platforms, the average professional still receives over a hundred emails per day. Many of those emails require a response, and the quality of that response matters -- it reflects your professionalism, your attention to detail, and your respect for the relationship with the sender. Building an AI system that can handle this responsibility required us to solve several hard problems simultaneously.
Understanding intent, not just content
The first challenge was classification. Not all emails are created equal. A message from your CEO asking for a project update requires immediate, thoughtful attention. A vendor confirmation email requires a quick acknowledgment. A newsletter requires nothing at all. Traditional email filters use rules-based systems -- keywords, sender addresses, subject line patterns -- to sort messages. These systems are brittle and miss context. An email from an unknown sender might be a cold sales pitch or a potential customer reaching out for the first time. The difference matters enormously, and it depends on understanding the content, not just the metadata.
Ivo's classification system uses a multi-layered approach. The first layer analyzes the sender's relationship to you based on your communication history. The second layer examines the content for intent signals -- is this a request, an update, a question, a social message? The third layer considers temporal context: is this email part of an ongoing thread, does it reference a meeting happening today, does it have a deadline? Together, these layers produce a nuanced priority score that determines how Ivo handles the message.
Learning your voice
Drafting responses is where the real complexity lives. Anyone who has used a generic AI to write an email knows the result: technically correct but tonally wrong. It sounds like a robot wrote it, or worse, it sounds like a different person wrote it. For Ivo to be trusted with email responses, the output had to be indistinguishable from what the user would write themselves. We built a personal language model for each user that learns from their sent messages over time. This model captures not just vocabulary and sentence structure but subtler patterns: how formally you address different contacts, whether you use exclamation points, how you sign off, whether you prefer bullet points or prose. The model also adapts based on relationship context. The way you write to your manager is different from the way you write to a client or a close colleague, and Ivo learns these distinctions.
The confidence threshold
Not every email should be auto-drafted. Some messages are sensitive, ambiguous, or politically complex in ways that require human judgment. We built a confidence scoring system that determines whether Ivo should draft a response, flag the email for your attention, or simply stay silent. The threshold is deliberately conservative. In our testing, we found that users strongly preferred false negatives (Ivo staying quiet when it could have helped) over false positives (Ivo drafting a response that missed the mark). Trust is fragile, and one bad email draft can undo weeks of reliable performance.
We also built an escalation framework for edge cases. When Ivo encounters an email it cannot confidently classify or respond to, it creates a brief summary of the message with its best guess at the appropriate action, then asks the user to decide. These escalations serve double duty: they ensure nothing falls through the cracks, and they generate training signal that helps Ivo handle similar situations better in the future.
Speed matters
Email intelligence is useless if it is slow. A draft that appears two hours after the email arrived has already missed the window of relevance. Ivo processes incoming emails in near real-time, typically generating a classified priority and draft response within thirty seconds of the message hitting your inbox. Achieving this required significant engineering investment in our inference pipeline. We use a tiered processing architecture where lightweight models handle initial classification and routing, while more capable models generate the actual draft text. This lets us maintain low latency without sacrificing quality.
Handling nuance at scale
Perhaps the most challenging aspect of email intelligence is handling the sheer variety of communication styles, subjects, and contexts that flow through a typical inbox. A single user might receive emails about project timelines, budget approvals, personal lunch plans, customer complaints, partnership proposals, and team birthday celebrations -- all in the same morning. Each of these requires a different tone, level of formality, and depth of response. We addressed this by building what we call contextual frames -- reusable templates of communication patterns that Ivo recognizes and applies. These are not rigid templates in the traditional sense. They are flexible behavioral patterns that adapt to the specific content while maintaining the appropriate register.
The result is an email intelligence system that handles roughly seventy percent of incoming email autonomously, surfaces another twenty percent with helpful context and suggested actions, and quietly files the remaining ten percent that requires no action. For our users, this translates to an average of fifty-two minutes saved per day -- nearly an hour of reclaimed time that used to be spent reading, sorting, and typing responses to messages that a well-trained AI can handle just as well.
Building email intelligence is not a solved problem. Every week we encounter new edge cases, new communication patterns, and new ways that human language surprises us. But the foundation is solid, and it gets better with every message it processes. We are excited to keep pushing the boundaries of what AI can do with the most fundamental communication tool in professional life.