A week later, someone pastes a paragraph into a thread and asks whether that is what you agreed. The paragraph came from an assistant. You remember approving something. You do not remember if it was this sentence, a shorter one, or a version you edited and then lost.
That is not a memory problem. It is a missing signature. Most AI-written mail has no moment where accountability attached, and no label that would tell a later reader who wrote which part.
Approval is the act you can point at
The confirm step is a signature, not a speed bump. Before send, you see the package as the other person will see it. You leave the draft alone, or you change the note. That choice is the whole record.
Leave it alone and you approved the assistant's wording. Change it and you claimed different words. An emptied box in the portal is refused, because deleting the sentence and sending the draft anyway would record an approval of something you just removed.
Without that step, "I sent it" and "I meant it" collapse into one event. A later argument has nothing to stand on except your recollection of a chat.
Labels are how a stranger reads the signature
A later reader — a colleague, an auditor, you in six months — should not have to guess which sentences were yours. Provenance labels sit next to the note. Verbatim is sacred: only the server-gated path may produce it, and only when the final wording differs from the draft. Echoing the draft is approval. A reply typed on a magic link is typed by the recipient, never verbatim.
The brief, the options, the assumptions each carry their own mark. A guess labelled as a guess can be corrected. The same sentence unmarked becomes a fact somebody later "agreed to."
You cannot prove a feeling. You can prove a labelled package that went out after a confirm.
What you can prove, and what you cannot
You can show what left: the topic, the ask, the note, the time, the labels. The sender's trail shows the send, whether notification was accepted, whether their assistant pulled it, whether they acknowledged or replied.
You cannot show that a person read a page because a scanner fetched the link. You cannot show that a Slack reaction meant yes. You cannot unsend a message that already arrived and rewrite history; you send a correction.
That limit is the honesty. A product that pretended a prefetch was a read would give you a stronger story and a worse one.
How to leave a record you can use
Approve on purpose. If the note is yours, change it so the label can say so — or accept that you are signing the draft. State the rejected option so it cannot reopen as "we never considered that." Keep the session off the package. The chat is not the evidence. The briefing is.
When someone quotes a paragraph later, you open the package, not the transcript. The question "did I approve this?" has an answer. That is a different kind of confidence from "I think I said something like that."