Quick answer (TL;DR): Let your AI assistant draft anything — messages, documents, slide decks, ticket descriptions — but install one hard gate: nothing externally visible goes out without your explicit approval of the exact final content. Draft freely; send deliberately. This single rule lets you capture the speed of AI-generated drafts while keeping the judgment of a human who knows the room. Pair it with a verification step — always confirm the edit you asked for is the edit that actually landed — and you can delegate real deliverables without ever being embarrassed by one.
This is the rule that makes it safe to hand real work to an AI assistant, and it’s the backbone of any serious AI working method.
The tension: speed vs. judgment
AI assistants are phenomenal drafters. They’ll produce a first version of an email, a project doc, a deck, or a tricky reply in seconds — versions that would’ve taken you twenty minutes to start. That speed is the whole appeal.
But speed without a gate is dangerous. The assistant doesn’t know the room the way you do. It doesn’t know that this stakeholder is touchy about that topic, that this number is sensitive, that this phrasing will land wrong with that person. If generated drafts go out unreviewed, you eventually ship something that embarrasses you — and the time you saved is dwarfed by the cleanup.
The resolution isn’t to draft less. It’s to separate drafting from sending.
The review-before-send rule
State it plainly and make it a hard rule in your preferences:
Nothing externally visible goes out without my explicit approval of the exact final content.
“Externally visible” means anything other people will see: emails, messages in shared channels, tickets, comments on shared documents, decks for stakeholders. For all of it: the assistant proposes; you dispose. It drafts the exact final text; you read it; you approve it or change it; only then does it go out.
This gives you both halves of what you want:
- The speed of never starting from a blank page.
- The judgment of a human who knows the context, applied at the one moment that matters — right before it goes out.
The one sensible exception
There’s one place to relax the gate: your own private workspace. Your personal tracker, your scratch notes, your second brain files — the blast radius there is just you. Let the assistant act freely in your private space so you’re not bottlenecking trivial updates. The gate is specifically about things other people will see. Anything that stays inside your own four walls doesn’t need it.
Documents you iterate on: who owns the master?
A special case worth getting right: documents you’ll revise repeatedly over time — a deck you refine across weeks, a doc that goes through rounds. Here, decide explicitly who owns the master copy.
The approach that worked best for me: I kept one anchor file and made the edits myself, and the assistant gave me precise, located instructions — this section, change this to that — rather than spawning a brand-new version every time. This prevented “version sprawl,” where you end up with seven nearly-identical files and no idea which is current.
Then came the step that matters more than it sounds: after I applied the edits, the assistant pulled the file back and audited that each change had actually landed.
Why the verification step is non-negotiable
Here’s a subtle failure mode that will bite you if you don’t guard against it. When a human applies a list of edits by hand, the most common error isn’t making a wrong edit — it’s making a half-applied one. You meant to replace the old phrase with the new one, but the old fragment got left behind, and now the two collide into a garbled sentence. You don’t notice because you were moving fast.
Having the assistant verify the result — re-read the document after your edits and confirm each intended change is cleanly in place, with no leftovers — catches exactly this. It’s the cheapest insurance you’ll ever buy on a deliverable. The rule:
Draft freely, send deliberately, and always verify that the edit you asked for is the edit that landed.
A practical workflow for AI-assisted deliverables
Putting it together, here’s a reliable loop for producing real work with AI:
- Brief the assistant with the goal and the relevant context (which your second brain already holds).
- Let it draft the full deliverable — freely, no hesitation.
- Review the exact final content yourself, applying your knowledge of the audience and the room.
- Iterate — give specific feedback, get a revised draft.
- For iterated documents: make edits in your single anchor copy; have the assistant give located instructions, not new versions.
- Verify: have the assistant confirm the final state matches what you intended, with no half-edits.
- Then, and only then, send — through the gate of your explicit approval.
Most of this is fast. The gate adds seconds, not minutes, and it’s the difference between “AI saved me time” and “AI created a mess I had to clean up.”
The mindset: AI as a capable junior who needs sign-off
The cleanest way to hold this is to treat the assistant like a sharp, fast junior colleague whose work you’d never send out unread — not because they’re bad, but because you’re the one whose name is on it and who knows the politics. You’d happily let that junior draft everything. You’d also read it before it left the building. Same here.
That’s not a lack of trust in the AI. It’s ordinary professional accountability, applied at the right checkpoint.
Frequently asked questions
Can I trust AI to send emails and messages for me? Let AI draft anything, but apply a review gate: nothing externally visible should go out without your explicit approval of the exact final content. Draft freely, send deliberately. The assistant proposes; you dispose.
How do I stop ending up with dozens of AI document versions? Keep one anchor copy as the master and have the assistant give you precise, located edit instructions rather than generating a new file each time. This prevents version sprawl and keeps a single source of truth.
Why should AI verify edits it suggested? Because the most common human error when applying edits by hand is a half-applied change where old and new text collide. Having the assistant re-read and confirm each change landed cleanly catches these garbled leftovers before they ship.
Is there any work I can let AI do without reviewing? Yes — work in your own private workspace (personal trackers, scratch notes, your second brain files) where the only person affected is you. The review gate is specifically for anything other people will see.
This is Part 7 of a series on building a working method with AI. Next: the discipline that stops a helpful AI from becoming a flattering one.