A friend's father died on Tuesday. You have had the message box open since Wednesday. Your assistant would produce three warm, well-organized paragraphs in four seconds, and every one of them would be worse than the two clumsy sentences you would write yourself.
Most writing about AI drafting is about doing it well. This is the other list — the messages where the correct move is to close the assistant and type.
When the effort is the message
Condolences, apologies, thanks, praise. These four carry almost no information. An apology's content is not the words "I'm sorry" — it is evidence that you understood what you did, and a model that was not there cannot have understood it. Praise that names nothing specific tells the recipient you wanted credit for noticing rather than that you noticed.
For these the effort is the payload — generating them does not make the message faster, it removes it. And the clumsy version is not merely more virtuous, it is better: a sentence that starts twice, or says the awkward true thing plainly, is evidence a person was at the keyboard not knowing how to put it. Fluency is free and proves nothing. Hesitation still costs something.
A useful test: if the recipient found out you had generated it, would the message become worthless? For a status update, no — the facts are still the facts. For a condolence, yes, completely. Anything in the second category, write yourself. Badly is fine.
When a wrong fact would bind you
Prices, dates, scope, availability, promises — anything the reader will act on. The problem is not that models invent things; it is that generated text has no confidence gradient. A date read off your calendar and a date inferred from context arrive with the same certainty. When you write, uncertainty leaks: you hedge, you slow down on the number, you go check. A draft does none of that, and confident phrasing is not verification.
Once it goes out under your name it is your claim, and walking a wrong number back costs more than checking would have. Draft the surrounding message; put your hands on the sentence that binds you.
When reviewing costs more than writing
To own a draft you have to read it about as carefully as you would have written it, and anything cheaper is skimming — which is how the unverified detail above gets through. So the threshold is short: if the instructions you would have to give run about as long as the message, write the message. "Yes, Thursday works, I'll bring the deck" does not need a pipeline.
There is a second cost people do not count: a fluent draft anchors you. Your own phrasing starts to feel clumsy beside it, and you send the draft's framing rather than yours — not because it says more, but because it arrived first. Write your version first.
When it is actually a conversation
If the exchange runs at one line every thirty seconds, composition overhead is the entire cost. Drafting pays when the unit is a dense message travelling one direction; it is dead weight when the unit is a sentence and the value is turn speed.
The louder signal is the clarifying loop — you are on the sixth round of narrowing what the question even was. That is a medium problem, not a message-quality one, and it started three rounds ago. Call them. Fifteen minutes on the phone resolves what six asynchronous turns will not; asynchronous is a tradeoff you choose, not a default.
When the relationship needs to see you did it
This one is not about the message type but about what the reader concludes about you. Someone who reports to you reads polish as distance. A person whose time you are asking for as a favor is deciding whether you are worth the favor, and your message is the only evidence they have.
None of that is about detection — whether a message "sounds like AI" is a dead end, and people guess wrong both ways. It is the gap between your written register and your actual one, which surfaces because you end up in a room together, and it is paid in exactly the relationships you were trying to maintain efficiently.
Where drafting genuinely earns it
Remove those five and the residue is specific: a lot of context, one reader who lacks it, and a question that has to be impossible to miss. That is an organizing problem, and organizing is what models are good at.
The work is compressing three weeks of detail into the couple of hundred words a cold reader needs — the part humans do worst under time pressure, because we narrate the history instead of stating the current position. Nothing in that job description is a claim about your inner life, so nothing is lost by having help.
The parts only you can say stay yours
The rule that survives all five cases: an assistant can carry the context, but it cannot carry the stake.
On RelayLink that is structural rather than a discipline you must remember. The sender's note is the field built for your words, and the verbatim label is earned rather than asserted — echo the draft back unchanged and it ships as AI-drafted, approved unchanged, because approving is not authoring.
That is also the honest limit of the review step. Reviewing the exact package before it sends catches wrong facts and wrong tone, which is the whole reason there is no unsend afterwards. It does not make a generated condolence real: no approval workflow converts a message whose whole content was supposed to be your attention into one that has it.
So the test is what the message is made of. Facts somebody else needs: draft it, review it, send it. Evidence that you cared: type it yourself, unevenly. When it is the first kind, connect your assistant.