Let AI propose. You approve. The small workflow rule that keeps you in charge

You asked the AI to clean up the draft overnight. This morning the draft is beautiful, and it is no longer yours. Something about clause seven feels different, but you cannot say what changed, because everything changed a little. Your choices are to diff the whole thing line by line, or to sign work you have not really read.

That is the silent-apply trap, and it has nothing to do with how good the AI is. Even a flawless rewrite severs the link between you and the text you are responsible for. The person who signs owns every line, and ownership you cannot trace is ownership in name only.

Propose-don't-apply means the AI's output lands as a visible, attributed suggestion: a tracked change, a comment, a highlighted diff with an accept button. Nothing becomes part of the document, the record, or the outgoing email until a human says yes to that specific change.

The power is in the granularity. Accepting or rejecting one small change takes seconds and keeps your mental model of the document intact. You are never asked to trust a whole rewrite; you are asked, dozens of small times, whether this exact edit is right. Courts sanctioned lawyers over unverified AI work, including the $5,000 fine in Mata v. Avianca, precisely because the output skipped human review on its way to mattering.

Full automation with a final review sounds equivalent, and it is not. A finished document hides its changes; auditing it means re-reading everything with no map of what moved. Attention fades, the deadline leans on you, and unreviewed text slides through. Every professional has skimmed page nine.

Proposals reverse the economics. Each change arrives small, in context, at the moment you are already thinking about that passage. Rejecting a bad suggestion costs one click instead of an argument with a finished draft. And the AI's misreadings surface early, while they are cheap, instead of compounding through a document built on a wrong assumption in paragraph two.

The dividing line is authorship and accountability. Work a human signs, files, sends, or defends: contracts, pleadings, client emails, board minutes, published posts. There, insist on propose-don't-apply, because the signature is yours and the judgment must be too.

Mechanical, reversible, delegated work is different: reformatting, transcription cleanup, sorting, renaming, summarizing for your own reading. Silent application is fine where a mistake is cheap and undoing is easy. The skill is noticing which kind of task is in front of you before the AI starts, not after.

This pattern only survives if the tool makes it effortless. A chatbot that outputs a rewritten wall of text has already applied; pasting it over your draft is the silent-apply trap with extra steps. Look for surfaces where suggestions are a first-class object: tracked changes, comment threads, accept and reject controls on each edit.

That is how Tatsulok treats collaboration on work that is yours to own: the AI drafts and suggests, the suggestions stay visibly suggestions, and every claim it makes carries its citation so the accept decision is an informed one. The AI is capable of doing more, and it deliberately asks first, because the point is not to replace your judgment but to make exercising it fast.

A drill for this week, no signup needed: next time an AI rewrites something you will sign, do not paste the result. Ask it to list its changes one by one, each with a reason. Accept or reject each from the list. Notice how different that feels from receiving the finished wall, and how much you catch.