Picture the scene. My assistant drafts an email to your team. Your assistant reads it, summarizes it for you, drafts a reply on your behalf, and sends it. My assistant reads the reply, summarizes it for me, drafts the next one. Three rounds in, neither of us has read a sentence the other person actually wrote. We have, however, both spent five minutes approving things, and the calendar invite has been sent.
I do not yet know whether this future is good. I am increasingly sure it is the future we are building.
The asymmetric case is fine
Start with what's working. Asymmetric drafting — where one party uses an assistant to compose and the other reads the message attentively — is roughly a pure productivity gain. It is not different in kind from spell-check, autocomplete, or a junior writing the first draft for a senior to edit. The signal in the message — the request, the offer, the question — still moves human-to-human. The assistant is a faster typewriter. Most of the public discussion about AI in email assumes this case, and most of the public discussion is correct that this case is fine.
The case I am worried about is the symmetric one, where both ends of the channel are automated.
What changes when both sides automate
Three things shift, and the change is not a matter of degree. It is a change in what the medium is.
Attention stops being spent. Email's value as a coordination medium has always rested on a costly signal: the sender spent some attention to compose, the receiver spent some attention to read. The cost is what makes email a useful signal of I considered this. When both ends automate, the cost collapses to almost zero on both sides. The artifact still exists — the message in the inbox, the reply, the timestamp — but the underlying claim I thought about this and chose to send it is now false on both sides. The protocol remembers the form. The substance has left.
Signal-to-noise inverts. Most of the cues we use to read an email's importance are side effects of human composition: the length, the formality, the time of day, the typos, the structure. Those cues were generated involuntarily by the writer's state of mind. When an assistant generates the message, those cues become style choices made by the model, and the receiver's instincts — trained for a decade on the involuntary cues — start systematically misfiring. A message that reads urgent isn't urgent. A message that reads casual was generated by the same model that generated the urgent one yesterday. The reader's calibration breaks.
The protocol assumes a human at each end. Email's accountability model — you sent this, you signed it, you are responsible for what it says — assumes a sender who composed and a receiver who read. When both are automated, both ends can plausibly disclaim the content. My assistant misunderstood. I would never have agreed to that. The summary was wrong. The accountability fragments and the question of whose word counts becomes genuinely unclear. This is not an edge case. This is what happens by default once both sides delegate.
What the email is actually for
If neither side is reading and neither side is writing, what is the email loop doing? A few candidate answers, each unsatisfying.
It is producing a legal record — a timestamp, a written exchange, a paper trail that can be subpoenaed or audited. The artifact has value even if no human ever read it.
It is coordinating calendars and tasks — the email is the wrapper around an action that the assistants negotiate (a meeting time, a file transfer, a contract signature). The substance is the action, not the message.
It is performing the form of considered communication — both sides demonstrating that they are responsive, professional, and engaged, even when the engagement is delegated. This is the most uncomfortable answer because it admits that much of what we treat as substance was already form.
I think all three are happening at once, and the proportions vary by context. None of them justifies the volume of the loop.
What might replace it
The honest replacement for AI-to-AI email is not a better email client. It is a different protocol — one that admits both sides are automated and stops pretending otherwise.
A few directions I find plausible:
- Structured proposals. Instead of prose, the assistants exchange machine-readable proposals: here is what my principal is asking for, here are the constraints, here is the deadline. The receiving assistant evaluates against its principal's preferences and either accepts, counters, or escalates. The human reads only the result.
- Escalation gates. The loop runs at machine speed until a defined boundary is crossed — a dollar amount, a calendar conflict, a topic flagged as sensitive — at which point the human is brought in with full context. Most loops resolve without escalation. The human's attention is reserved for the ones that need it.
- Explicit attestation. Messages carry metadata about whether they were drafted by a human, by an assistant, or jointly. The receiving end's calibration adapts. This is technically easy and politically hard.
None of these are email. They are something else dressed up in email's familiar surface, the way email itself was once a letter dressed up in computer.
What I still don't know
- Whether the loop will collapse on its own. It is possible the volume becomes self-evidently absurd and the protocol moves to one of the directions above. It is also possible the volume is the point — that demonstrating engagement is what the channel was for all along, and AI just made it cheaper.
- How to handle the legal-record case. A signed contract email exchange has weight precisely because each side composed it. If both are automated, the law has not yet decided what that means. I suspect the answer will be ugly and project-specific for a while.
- Whether attestation will be honored. A protocol that asks parties to declare this was written by an assistant depends on the parties wanting to declare it. Most won't, because the declaration carries cost.
The unifying observation is that AI does not just speed up email. It changes what email is for, and the change is not gradual. The medium was a costly signal of attention. When the cost goes to zero on both sides, the signal is gone, but the artifact remains. We will spend the next few years figuring out what we actually wanted from the artifact in the first place.