Someone challenges a paragraph in a shared doc and the author says, out loud, "oh, that part was the model — ignore it." Everyone moves on. Over the next week, three people quietly re-read a month of that person's messages, trying to work out which parts they were meant to have taken seriously.
Nobody reports that. It just happens, and a colleague's writing carries a little less weight than it did.
The absence of a norm is not neutral
A team that never discusses this doesn't end up with no policy. It ends up with one per person, none written down, none visible to anyone else. One appends "drafted with AI assistance" to everything, including two-line replies they typed themselves. Two never mention it and draft half of what they send. One sends beautifully structured messages at 6am that nobody believes were read first.
Everyone else has quietly recalibrated. That is the real cost and the one nobody raises: readers skim what would once have got full attention and re-check what they would once have accepted. Discounting is silent — nobody tells you they've started reading your writing as machine output, so nobody gives you the chance to fix it.
It lands hardest on the careful: whoever agonised over one paragraph gets the same discount as whoever forwarded unread output, because from outside the two look identical. That is ambiguous attribution taxing the diligent at everyone else's rate.
Five norms worth the argument
Own your asks and commitments in your own words. These are the sentences that create obligations — the request, the deadline, the price, the agreement. If a message asks or promises, those sentences should be typed by the person whose name is on it. Everything else can be drafted without harm. The narrowness is the point: broad rules like "write your own emails" get abandoned within a month, while cheap and obviously right ones survive.
Verify anything factual you'd be held to. A model's confident wrong number, sent under your name, is your wrong number. Fluency and accuracy are unrelated, so nobody downstream can grade a claim by how well it reads. The rule isn't "check everything" — it's "check what would be expensive to be wrong about," a much shorter list than it sounds.
Disclose at the sentence level, not in a footer. Readers don't want to know whether a model was involved; assume it was. They want to know which sentences you stand behind. One line at the top — "the summary is my assistant's, the recommendation in section three is mine" — carries more than a hundred footers. The longer argument is in what you actually owe the recipient.
Never send unreviewed. Not because unreviewed output is always bad, but because review is the only moment a person becomes answerable. Skip it and your team's position on any claim is "an unsupervised process wearing our name said that."
Agree what's off-limits for drafting entirely. Apologies, condolences, thanks, praise, feedback on someone's work. A generated apology is worse than none: if the sentiment wasn't worth the sender's attention, the recipient will correctly conclude it wasn't a sentiment. Expect this to be the item your team actually argues about; the argument is the useful part. Some messages shouldn't be drafted at all covers where people draw the line.
A norm nobody can check becomes theater
Three of those five run purely on individual discipline. Nobody can tell from outside whether you verified the figure, typed the commitment, or read the draft before sending, and a norm whose violations are undetectable degrades into a statement of values — sincerely held, quietly dropped the first week everyone is behind.
Checkable norms leave a trace. "Commitments go in a numbered list at the top" is checkable — the list is there or it isn't. So are "the ask is the first sentence" and "figures that matter carry a source."
Structural norms are held by the tool, so nobody has to remember. A channel that requires a human to view the rendered message before releasing it enforces "never send unreviewed" without anyone's willpower.
The third pile — real, important, entirely unverifiable — should be short. Two will hold. Nine will not, and a list people have visibly stopped following is worse than none: it teaches everyone the page is decorative.
Writing them down without producing a policy document
Half an hour, one page, somewhere your team actually reads.
- Draft the list first, then argue it. Arguing from a blank page produces the vaguest possible wording.
- Decide the off-limits categories together. That's where real disagreement lives, and a line only one person believes in isn't a norm.
- Say what happens when someone breaks one. "We mention it" is enough — a norm with no response to violation is a preference.
- Date it for revisiting. Tools change underneath these agreements, and a stale page trains people to stop opening it.
Where the channel can hold a norm for you
Prefer structure wherever it exists; structure doesn't get tired.
RelayLink has no team settings — nothing to configure centrally, no policy to push to anyone. It makes two of these norms properties of the channel for whoever uses it. Sending is two steps: an assistant creates a draft, the human reviews the package exactly as it will arrive, and a separate confirmation releases it. No single call composes and delivers, which turns "never send unreviewed" from a promise into a property. Provenance is recorded per field rather than claimed — a sender's note is labelled verbatim, human-authored only when the wording genuinely differs from the assistant's draft, so approving a suggestion is recorded as approval, not authorship.
That covers two of five. Verifying facts, owning the asks, and keeping the human messages human stay yours — write those down anyway.
If you'd rather the review step be enforced than remembered, connect your assistant and send one briefing to see the approval gate.