When your AI talks to their AI, the 1:1 stops being a conversation

There is a new failure mode with a name now. An employee could not figure out her manager, kept getting messages she suspected were AI-drafted, and started asking her own AI how to read them and what to send back. She put it plainly: it is just his AI and my AI going back and forth ^source. Skillsoft's Leena Rinne calls this social offloading, when the interpersonal work that needs judgement or empathy quietly gets handed to a tool ^source. The part worth sitting with is not that people use AI to write. It is that once AI sits between two people as a drafting layer, the relationship stops being built by either of them.

The reps you outsource are the ones that build the relationship

A working relationship between a manager and a report is not a single event. It is reps. Reading each other, following through on what you said, remembering what mattered last time and bringing it up. Those are the small, repeated acts that turn two org-chart boxes into two people who actually work together.

Every one of those reps is exactly what a drafting layer absorbs. If AI reads the message for you and writes the reply for you, you are not learning to read that person, and they are not learning to read you. Rinne's point lands here: keep asking AI how to respond to your boss and you never actually learn to work with your boss ^source.

This is not a hypothetical for people already stretched thin. Orgs have been flattening for years, and the spans are getting extreme. One new AI unit at Meta is running roughly one manager per 50 engineers, double the 25-to-1 that used to be treated as the outer limit ^source. When a manager is that far from the people they lead, coaching is already the first thing that falls away, and younger workers get thrown into the deep end without much of it ^source. Put a drafting layer on top of that and the last bit of direct contact goes too.

The tempting next step hollows out the one thing you have left

When AI already drafts the messages, the obvious move is to let it do more: write the 1:1 agenda, then the summary, then the feedback. Each step sounds like a time save, and each one takes a little more of the pair out of their own conversation.

The trouble is that the recurring 1:1 is often the last thing the two of them still fully own. No approval, no rollout, just a standing half hour that belongs to them. Hand its agenda, its summary, and its feedback to AI and you have not made the relationship efficient. You have quietly removed the people from it and kept the calendar invite.

Reading the relationship is not the same as speaking for it

So the line matters, and it is a sharp one. There is a difference between an AI that reads a relationship and an AI that speaks for one.

This is where 1on1 draws its edge on purpose. The AI reads the recurring relationship so the manager walks in prepared. It pulls last session's context forward, so the meeting picks up instead of resetting. It carries action items with open and closed status across sessions, so a commitment that keeps slipping is visible as a pattern rather than a vague feeling. It keeps the 1-5 session-health rating as a trend, so a pair sliding over six weeks reads as a line. There is a manager-only coaching read on top, kept on the manager's side.

What it never does is talk in the room. It does not write your half of the conversation or the report's half. It is not a transcription bot pointed at two people. It prepares the manager and then gets out of the way, because the point is to support the conversation, not to hold it for you.

The test any operator can run

Here is the question to put to any 1:1 tool, and it cuts fast: what does its AI write, and what does its AI read.

A tool that authors your half of the conversation is replacing the relationship, one draft at a time. A tool that remembers the relationship, the prior sessions, the open commitments, the health trend, is supporting it and leaving the talking to you. Same category, opposite effect, and the difference is the whole thing.

Try it on one pair. Run the recurring 1:1 for four sessions with the AI reading the history and preparing you, and speaking for neither of you in the room. Then check whether the conversation still feels like yours. If it does, you have found the version of AI in a 1:1 that is worth keeping. If it does not, you spent four sessions learning something cheap and important. ^source