What to Hand to the Agent, and What to Keep: The Messages an AI Should Never Write for You
We build an agent that reads your inbox and drafts your replies. Here is the research on the messages it should not touch: apologies, bad news, first contact, anything where the effort is the point. And the caveat on summaries that even careful people miss.
What the recipient is actually reading
In 2023, researchers at Ohio State showed people a warm, well-written reply from a close friend to a message about burnout. The text was identical for everyone. For some readers it was labelled as written with AI's help; for others, with a human's help; for the rest, by the friend alone. The AI-assisted version was ju…
Keep these
Working from that evidence, here is the list we use ourselves. An agent can draft any of these if you want a starting point, but the words that go out should be yours, and the person should be able to tell.
Hand over these
The flip side is just as real, and it is most of the inbox.
The caveat on summaries
Summaries are where people over-trust the tool, and the research is specific about how. A 2025 study in Royal Society Open Science ran 4,900 summaries of scientific papers through ten leading models and found most of them overgeneralised: findings about a sample became claims about everyone. The models' summaries were…
The part nobody measures: what it does to you
There is one more cost, and it lands on the sender. A 2025 survey of 319 knowledge workers by Carnegie Mellon and Microsoft Research found that the more people trusted a generative tool, the less critical thinking they reported doing with it, and most said they put in less effort at every stage from understanding a ta…
How we built for this line
Kanvo's agent drafts and waits by default. Nothing it writes leaves your account until you send it, and when you let it send on its own, that is a decision you make per kind of conversation, with a person approving the rest. Every message an agent sends carries a record of who authorised it and whether it disclosed it…
Read next
Say it's AI: what the research says about disclosing AI-written messagesWorkslop and the one-ask messageAI that replies to your messages, safely
Kanvo Team
Writes here about what Kanvo is building and why — the send path, the evidence it leaves, and the decisions behind both.
Related Articles
Thirteen experiments say telling people a message was written with AI makes them trust you less. The same research says being found out costs far more, that suspicion now falls on everyone, and that the EU, Utah and California no longer leave you the choice. Here is how to disclose without losing the room.
Say It's AI: What the Research Actually Says About Disclosing AI-Written Messages
Thirteen experiments say telling people a message was written with AI makes them trust you less. The same research says being found out costs far more, that suspicion now falls on everyone, and that the EU, Utah and California no longer leave you the choice. Here is how to disclose without losing the room.
Workslop and the One-Ask Message: How to Write for a Reader Who Has an AI Too
AI made polished text free, so the cost of a message moved to the person reading it. 40% of desk workers received "workslop" last month and spent nearly two hours on each piece. The message that still works in 2026 carries one ask, in the first line, with the context a human or their assistant needs.
Enjoyed this article?
Subscribe to our newsletter for more insights and updates.