Journal · The AI Architect · 2026-08-26
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 is not really about trust in the machine. It’s about trust in yourself under pressure. When you’re behind on email, or juggling three client threads, or trying to sound gracious to someone who annoyed you, the temptation is to let the draft go straight out the door. The AI wrote something reasonable-sounding. Why add friction? The honest answer is that fluency and correctness are different things, and the moments you’re most tempted to skip the read-through are exactly the moments you most need it.
Fluency is not the same as truth
AI-generated text has a particular danger: it reads as though someone confident and competent wrote it. That confidence is borrowed, not earned. The model doesn’t know your relationship with the client, doesn’t know that “following up as promised” will land wrong because you never actually promised anything, doesn’t know that the joke about deadlines will sting because this person just missed one. It produces plausible language, and plausible language is disarming. You relax your guard precisely because it sounds like something a careful person would write.
This is why the split matters more for outbound communication than for almost anything else you use AI for. A bad first draft of a report just costs you editing time. A bad message that already sent costs you a relationship, a correction, an apology. The stakes are asymmetric, so the safeguard should be non-negotiable, not optional.
Try this: for one week, before you approve any AI-drafted message, read it out loud. Not skim — actually voice it. You will catch things silent reading misses: wrong tone, a claim you can’t back up, a sentence that’s technically accurate but sounds evasive. Reading aloud slows you down just enough to reengage your own judgment instead of rubber-stamping someone else’s fluency.
Build the pause into your workflow, not your willpower
Relying on yourself to “remember to check carefully” is a weak plan, because the whole point of using AI for drafting is to move faster, and speed is the enemy of careful review. If the system depends on you staying vigilant every single time, it will fail exactly when you’re tired, rushed, or distracted — which is often.
Better to build a structural pause than a mental note. Some people set their AI tool to draft into a separate holding document rather than the send window itself, so there’s a literal extra step before it can go out. Others keep a short checklist taped near the screen: Does this say something I didn’t intend? Does it commit to anything? Would I be comfortable if this were read back to me in six months? A checklist sounds unglamorous, but it does the job willpower can’t — it doesn’t degrade when you’re busy.
Try this: pick one recurring type of message you send often — client updates, team check-ins, whatever it is — and write yourself three questions to ask before approving that specific type. Keep it short enough that you’ll actually use it every time, not just when you remember to.
The approval step is where your judgment still matters
There’s a quieter reason for the split beyond catching errors. Approving a message is where you decide what you actually think, not just what sounds good. AI can generate a version of your opinion, your apology, your pitch — but it can’t know which version is true to you until you tell it, and the moment you tell it is the moment you read the draft and say yes or no or not quite.
Skip that step often enough and something subtle happens: your outbound voice starts drifting toward the model’s defaults rather than your own. Not dramatically, not all at once, but steadily. The people you communicate with start hearing a slightly smoothed, slightly generic version of you. The split isn’t just quality control. It’s how you stay the author of your own words, even when you didn’t type the first draft.
I think about this distinction a lot when I write about AI as a tool rather than a replacement — the difference is almost always in who does the last read before something leaves your hands. In the book I go into more detail on how to structure that review step so it’s fast rather than burdensome, but the core habit is simple enough to start today: whatever it drafts, you read it before it sends. Every time, not just when you remember.
Go deeper. The full method is in The AI Architect. New here? Start with the free companion pack, or explore the series.