Why we built block-level diff review instead of single-pass AI edits

Early Documatch prototypes looked like every other AI writing tool: paste a prompt, get a wall of text back, hope you caught what changed. Our first lawyer beta testers hated it — not because the suggestions were bad, but because they couldn't defend the document to a client.

The problem with single-pass rewrites

When an LLM rewrites a whole section at once, three things break for professional documents: numbering drifts, defined terms get silently swapped, and accountability disappears. Partners need to answer "who approved this clause?" — not "the model suggested it."

• Track changes in Word shows edits, but not AI reasoning or scoped intent.
• ChatGPT gives answers outside the document — copy-paste reintroduces formatting risk.
• Full-document regeneration hides which paragraphs actually changed.

Hunks, not hallucinations

We borrowed the mental model from code review. Every AI proposal lands as a discrete hunk: old text struck through, new text highlighted, accept or reject per block. Nothing touches the canonical document until a human approves it.

Selection-scoped edits came next. Highlight a liability cap, ask for a tweak, and the model cannot wander into indemnification. That constraint dramatically reduced "fix one paragraph, break three others" failures.

What we measure now

Teams on block-level review accept ~84% of proposed hunks on the first pass — but the 16% they reject would have been silent rewrites in a single-pass tool. Version checkpoints on every apply mean rollback is one click when a partner wants the prior language back.

If you're evaluating AI for client work, ask one question: can I reject a single bullet without undoing the whole section? If not, you're not reviewing — you're gambling.