Why AI drafts sound almost right
Sounding like you is the easy part. Knowing what you’d never say is the work.
- Based on
- Building a word-for-word approved claims list with a client, and reading the edits they make to drafts.
- Not claimed
- Legal advice on what you can claim.
Getting an AI to sound roughly like you takes an afternoon. Give it a few dozen of your emails and it picks up your greetings, your rhythm and the way you sign off. Most people stop there, because the drafts look right.
Look closer and they’re almost right. The tone is yours, but a sentence promises next-day dispatch you don’t offer. A thank-you note promises plastic-free shipping, when the claim you’ve actually verified is plastic-free packaging. A reply to a wholesale buyer uses the warm, chatty voice you keep for loyal customers, and lands as over-familiar. None of it is dramatic. All of it is the kind of thing a founder notices, and a customer or a regulator might too.
That gap is what aligning an AI should mean. It isn’t about sounding like you. It’s about knowing what you’d never say, and that’s harder than it looks, because most of it has never been written down. It lives in your head as instinct: the claims you can stand behind, the promises you don’t make, who gets which tone, and when to stop and ask you.
So the real work is unglamorous. We write down the claims you can make, word for word. We write down what you’d never say. We decide which kinds of message the assistant may draft and which come straight to you. And we treat anything that isn’t on the approved list as not approved. An unknown never quietly becomes a yes.
Then the record does the rest. Every edit you make to a draft is information. If you keep changing the same phrase, that’s a rule we haven’t written yet. I’d rather spend weeks finding those rules with you than ship something that sounds like you and occasionally speaks for you in ways you’d never sign off.
The voice is the easy part. The judgement is the work.