There's a small line of text in every RelayLink package that we consider the most important thing we ship: the provenance label on the sender's note. Sometimes it reads (verbatim, human-authored). Sometimes it doesn't. The difference is the whole product.
The question every AI-mediated message raises
When a message arrives via someone's assistant, the recipient silently asks: did the person say this, or did the machine? Get that answer wrong in either direction and something breaks.
Treat the model's words as the human's, and you'll hold someone to a position they never took. Treat the human's words as the model's, and you'll discount the one sentence they actually meant. Most AI communication tools leave this ambiguous — the output is fluent, the attribution is vibes.
RelayLink's position: attribution is data, and it must be earned, not asserted.
How a label gets earned
Every field in a briefing carries provenance. Assumptions are marked stated by sender or inferred by sender's AI. Excerpts say who spoke. The package states whether the sender reviewed it before it went out. And the sender's note — the one place a human speaks directly — follows one strict rule:
The verbatim label appears only when the human's final wording differs from anything the AI drafted.
If your assistant suggests a note and you send it unchanged, the package says the note was approved — true, and useful. It does not say you wrote it, because you didn't. Echoing a draft back is approval, not authorship. The system checks, server-side; the label is not the model's to grant.
Words a recipient types through the reply link or by email get their own attribution: typed by that person, via that path. Honest, but distinct — the verbatim label belongs to one path only, so it can't be diluted.
Why so strict? Because labels only work if they can't lie
A provenance label is a promise about process. The moment "human-authored" can be produced by a model with a cooperative human clicking OK, the label means nothing and every recipient goes back to guessing.
This is the same reason the label logic lives in one place in our codebase, and every surface — email, web, the assistant's own view — renders from that single source. A label that could drift between surfaces would be worse than no label: it would be confident misinformation about who said what.
What recipients get to do with this
Provenance turns a briefing from prose into evidence. A recipient (or their assistant) can weigh each part for what it is:
- The verbatim note — this is the person, respond to them.
- AI-inferred assumptions — check these first; they're where good briefings go wrong.
- Approved framing — the human vouched for it, at the altitude of a signature, not a draft.
That's how disagreements stay about the substance instead of about what was "really said." The transcript-forwarding world can't offer that; it can only offer more transcript.
The bet
Our bet is that AI correspondence becomes trustworthy the same way contracts did: not by making the parties more honest, but by making the record unambiguous. Provenance labels are that record. Guard the words that are yours — and let the machines be plainly the machines.