Twenty seconds between a draft and a mistake.
An AI wrote something for you. Before it goes anywhere — before you paste it into a reply, a post, an invoice, an email — read it against three questions. It takes about twenty seconds once you know them, and it's the entire difference between AI that helps and AI that embarrasses you in front of a customer.
The three questions
1 — Facts
Is everything in here actually true?
Not "does it sound right." True. Check every number, every date, every name, every claim about what you offer. The AI didn't verify any of them. It produced text that fits the shape of your business, which is not the same as text that matches it.
Specifically: did it resolve everything it flagged as uncertain? If it wrote [CHECK: confirm the delivery date] and you send it with that still in there, you've just sent your customer a note that says you weren't paying attention.
2 — Tone
Does this sound like me, or like a machine doing an impression of me?
Read it out loud if you're not sure. Out loud is brutal and fast. Any sentence you'd never say — "we would be delighted to assist you with your enquiry" — is a sentence that tells your customer they're talking to a bot, or to someone who's stopped caring. Both cost you the same.
3 — Promise
Did it commit me to something I can't deliver?
This is the expensive one, and it's the one people skip. The AI is agreeable by construction. It will happily confirm a date, imply a discount, or say "no problem" to something that is, in fact, a problem. Every promise in that draft is one you'll be held to. Read the draft as if it's already been sent, and ask what you now owe.
The three verdicts
FIRE — it's right. Send it. Most drafts, once your memory document is good.
FIX — one thing is off. Edit it and send. Normal.
BIN — it's wrong in a way that's faster to rewrite than to repair. Also normal, and it's not a failure.
The bit that makes it compound
If you're binning a lot, the AI isn't the problem. Your memory document is thin.
Every bin is information: the draft went wrong because the AI didn't know something. So don't just fix the draft — find the line that would have prevented it, and add it to the document the AI reads.
- Bad draft quoted a price you don't offer any more → add the current price.
- Bad draft was too formal → add a real example of how you actually talk.
- Bad draft promised weekend delivery → add "we don't deliver at weekends."
Do that for two weeks and the bins mostly stop. That's the whole trick, and it's why this gets better with use instead of staying flat.
Why a person, and not a setting
You could imagine an AI that checks its own work. It'll tell you it did. But it will also tell you, with total confidence, that a draft is fine when it isn't — the same confidence it uses when it's right. There's no tell. (Thirty-second proof: the mirror test.)
So the check has to be someone with something at stake. That's you. It's twenty seconds, and it's the reason you can use this stuff on real customers at all.
→ Try the Reply Drafter, free — a real draft to run this on.
→ What the Gate caught — what this looks like when the AI is the one being stopped.