Journal · The AI Architect · 2026-07-31
For anything that sends, split it: let your AI prepare it, then you read and approve before it goes out.
For anything that sends, split it: let your AI prepare it, then you read and approve before it goes out.
The tension isn’t really about trust in the machine. It’s about trust in yourself under time pressure. You know you should read the email before it sends. You know you should double-check the client message. But at 4:47 on a Friday, with six more things waiting, the temptation is to let the draft become the send. That gap — between what you intend to review and what you actually review when you’re tired — is where the trouble lives.
Separate the writing from the releasing
The core move is architectural, not moral. Don’t rely on willpower to catch mistakes when you’re rushed. Instead, build a structure where the AI’s output physically cannot leave your hands without a separate, deliberate action from you. Drafting and sending should never be the same step, even when the tool makes it easy to combine them.
Try this: for one week, change your workflow so every AI-drafted message goes to a holding place first — a drafts folder, a notes doc, a Slack message to yourself — instead of directly to the recipient. Then, on a second pass, you move it out manually. Notice how often you change something in that second look. That number tells you how much this gap is actually protecting you.
Read it as if someone else wrote it
There’s a specific failure mode worth naming: when you’ve been part of drafting something, even loosely, you start reading it as “mine” and skim for tone instead of checking for accuracy. The fix is a small trick of perspective. Read the AI’s draft the way you’d read an email a new hire sent you for approval — someone competent, but who doesn’t know everything you know, and who occasionally gets confident about things that are wrong.
This means checking three things every time, not just vibes:
- Facts and numbers — did it get a date, a price, a name right, or did it smooth over something it wasn’t sure about?
- Commitments — did it promise something on your behalf that you haven’t actually agreed to?
- Tone for this specific person — a draft that sounds fine in general can land wrong for someone who’s frustrated, grieving, or already skeptical of you.
Try this: before you approve anything that goes to a client, a boss, or anyone outside your closest circle, say out loud (or type) one sentence answering “what could go wrong if this exact message was misread.” If you can’t answer quickly, that’s a sign you haven’t really read it yet — you’ve read past it.
Make the pause procedural, not personal
If the pause depends on you remembering to be careful, it will erode the first week you’re overwhelmed — which is exactly when you need it most. So make the split a rule about the tool or the process, not a habit you’re hoping to maintain through discipline alone.
Some ways to do this concretely: set up your email client or app so AI-drafted messages default to a draft state, never auto-send. Use a naming convention — every AI-prepared file gets tagged “DRAFT — needs eyes” until you’ve personally cleared it. If you work with a team, make it a stated norm: anything AI-assisted going external gets a second human look before it goes, no exceptions for urgency. Urgency is precisely when people skip steps, so the rule has to hold especially then.
Try this: pick the one channel where a mistake would cost you the most — client emails, public posts, anything with your name on it publicly — and put a hard technical block in place this week. Not a reminder. An actual obstacle, like requiring a manual export or a second app, that makes sending without reading nearly impossible.
The reason this matters isn’t that AI writes badly. Often it writes better than you would at the end of a long day. The reason is that speed removes friction, and friction is sometimes the only thing standing between a good draft and a real mistake. Splitting the process — draft, then a genuine pause, then approval — puts a small, deliberate cost back into the system, right where you need it.
I go into more of the specific setups for this — how to build the pause into different tools, what to do when you’re working with a team instead of solo, and how to tell the difference between healthy caution and just being slow — in the fuller chapter in the book. This is meant as a starting point, not the whole map.
Go deeper. The full method is in The AI Architect. New here? Start with the free companion pack, or explore the series.