Most of what's been written about AI agents and email so far focuses on drafting quality — whether the model writes convincing copy. That's a small part of the actual shift. A chatbot's whole interaction model is synchronous: a person types, the model replies, the conversation ends when the tab closes. Email is the opposite shape — asynchronous, often one message every few hours or days, frequently with other people or systems in the thread who never see a "chat window" at all. Teaching an agent to use email well isn't a cosmetic change from chat to a different UI; it's a change in what the agent has to be good at.
A chatbot never has to decide when to follow up, because the human is right there to ask again. It never has to draft something that reads correctly out of context, because the whole conversation is the context. And it never has to handle silence — no reply at all — as a normal, frequent outcome rather than an error. An agent that corresponds by email has to do all three: decide when enough time has passed to follow up, write a message that makes sense to someone who wasn't watching the conversation unfold, and treat no-reply as the default rather than the exception.
This is where drafting as its own step, separate from sending, earns its place in the design. A chat reply can be half-formed and still work, because the human will ask a follow-up if something's unclear. An email can't — it has to carry its own purpose, tone, and key points in one shot, because there's no back-and-forth inside the message itself — a genuinely different failure mode than a bad chat reply, since there's no second turn in the same window to correct it. Giving the drafting step explicit tone and key-points parameters, rather than one freeform prompt, is what makes that possible without a human reviewing every draft before it goes out.
Some of what an agent sends by email needs to be exact: a confirmation number, a precise disclaimer, a quote it already computed and shouldn't paraphrase. Raw mode exists for exactly that case — the agent sends the exact wording it already has, rather than handing a purpose brief to a drafting step and hoping the output matches. A correspondent that only ever drafts, and never sends exact text, is missing half of what makes email a reliable channel rather than just a more patient chat window.
The harder shift is temporal. A chatbot has no concept of "it's been three days and no one replied" — that's an email-native problem, and it's the center of a pattern for agents that follow up on their own rather than waiting to be asked. It requires the agent to track what it's sent, notice what hasn't gotten a reply, and decide on its own whether a nudge is warranted — decisions a chat-only agent never has to make, because in a chat, waiting silently isn't really an option.
Receiving mail has the same asymmetry in reverse. A chatbot's "receive" is just the next message in the same window. An agent with an inbox has to actually watch for new mail arriving on its own schedule, decide whether what arrived needs a response at all, and in many cases triage a backlog rather than process one tidy turn at a time — the subject of building an agent that handles its own inbox end to end. Whether it finds out about new mail by asking (polling the message list on some interval) or by being told (a webhook firing the moment something arrives) is its own design decision, and the two approaches suit different kinds of correspondence — a support agent that needs to react within seconds wants a push; a research agent checking back once a day doesn't need one at all.
The skills email forces onto an agent — deciding when to follow up, writing for a reader with no shared context, handling silence as a normal outcome — aren't really about email. They're the skills any agent needs once it operates asynchronously with parties that aren't sitting in a chat window waiting for it. Email is simply the channel where those skills are most immediately testable, because the entire medium is built around exactly that asynchrony.
None of this is a claim that agents have become good correspondents in any human sense — tone can still misfire, follow-up timing can still be wrong, and a triage decision can still be a bad one. What's changed is simply that the infrastructure for trying now exists: an address, a way to draft and send from it, a way to receive into it, and a way to pay for all three without a human approving each step.