When to Stop Using an AI Draft and Start Again

There is a trap in editing an AI draft that every lawyer falls into once. The draft is eighty percent there, so you start fixing it, a wrong date here, a clumsy clause there, and an hour later you are still fixing it, because the twenty percent that was wrong was structural, not cosmetic. Sunk cost keeps you polishing a document built on a premise that does not hold. Knowing when to stop editing and rewrite from a clean instruction is a skill in itself, and it usually saves more time than the editing ever would.

Distinguish surface faults from foundational ones

Some faults are local: a garbled sentence, a missing party name, a clause in the wrong order. These are worth fixing, because the draft's foundation is sound. Foundational faults are different, they infect everything downstream. If the draft has misidentified the legal basis, answered a different question than the one you asked, or built its logic on a fact that turns out to be wrong, then every well-written paragraph resting on that error is also wrong, however good it reads. Fixing sentences will not repair a broken premise.

The signs that mean start over

Stop editing and restart when you see any of these:

• The draft has invented facts more than once, a sign it is filling gaps rather than working from your material
• It has missed the controlling issue entirely, answering an adjacent question well
• Its structure obscures the decision you actually need made
• You have rewritten the same clause twice and it keeps drifting back
• You can no longer tell which parts you can trust without re-reading all of it

That last one is decisive. When verifying the draft costs as much as writing it fresh, the draft has stopped saving you time.

Diagnose why it went wrong before you retry

A restart is wasted if the second attempt fails the same way. Usually the fault traces back to the instruction: the model was not told the governing law, was given facts you were still confirming, or was asked a question so broad it had to guess. Read the draft as evidence of what your prompt actually communicated, then fix the brief, not just the draft. A better instruction is almost always cheaper than a third round of edits.

Preserve the rejected draft as context

Do not simply delete a draft you are abandoning. Keep it, with a short note on why it failed, invented the limitation period, missed the arbitration clause, misread the client's objective. This record does two things: it tells the next person why the current version exists, and it stops you or a colleague from regenerating the same flawed output next week. A rejected draft is a lesson about the instruction, and it is worth keeping for that alone.

Write the corrected brief, then begin clean

Restarting does not mean starting from nothing. You now know far more than you did the first time: what the model gets wrong here, which facts it needs stated plainly, which issue it keeps missing. Fold that into a corrected instruction, name the law, supply the confirmed facts, flag what must stay open, and generate again from verified material. The second draft, built on a brief informed by the first, is usually the one worth editing.

The instinct to salvage is a good one in most of legal work, but with AI drafts it can quietly cost you an afternoon. Treat the model as a fast drafter whose output you are free to discard, not a colleague whose effort you owe respect. Sometimes the fastest route to a good document is to close the bad one, write a sharper brief, and start again.