Two versions of a contract go in. The agent runs the diff, resolves the edge cases, verifies its own work, and returns a summary of what actually changed — with citations into both versions. The lawyer's review is preserved as the final check. Not replaced.
Two versions of a 148-clause agreement, side by side, late in the evening. Hours of reading to find what changed — and the material change is the one buried in boilerplate.
The agent runs the diff, handles the edge cases, and verifies its own output. The lawyer reads a cited summary and makes the call. The review is preserved, not replaced.
The counterparty's revised draft lands at 19:40. One hundred forty-eight clauses. Somewhere between version 7 and version 12, things changed — and the redline is a wall of tracked edits, most of them cosmetic.
The old routine: a senior associate reads both versions side by side into the night, hunting for the changes that matter. Renumbered clauses. A definition quietly broadened. A cap moved two points.
It's careful work, and it's exhausting work, and it has to be done again for the next draft. The people trained to exercise legal judgment spend their evenings doing mechanical comparison.
The comparison is exactly the kind of work a machine should do first — if the machine's output could be trusted.
Redline tools list every edit — 312 tracked changes, most of them formatting and renumbering noise. What matters is buried in between: a definition broadened so it reaches three other clauses, a liability cap moved two points, a cross-reference that silently shifts meaning because the clause it pointed to moved.
A bare language model can summarize the differences, but it can't be trusted alone. It can miss a change, or report one that isn't there — and in a contract, either is unacceptable. And case files carry confidentiality duties. They cannot be pasted into a consumer chatbot running on someone else's cloud.
So the comparison stays what it has always been: a human reading two documents side by side, for hours, under deadline.
Inside your firm's application, the comparison runs as a pipeline: a deterministic diff, model judgment only on the genuinely ambiguous cases, and a verification pass before anything reaches the lawyer. Case files never leave the perimeter — on-premise where the matter requires it.
Hours of side-by-side reading collapse into minutes spent on a cited summary. The associate opens the five material changes, checks the cited clauses, and spends the evening on strategy instead of comparison. The next draft costs minutes too.
The lawyer's review isn't a fallback bolted on for comfort — it's a step in the workflow, by design. The system's job is to make that review fast and complete: every claim cited, every citation resolving to real text in both versions, every judgment the model made recorded and inspectable.
And the case file never left the building. The pipeline runs inside the firm's perimeter — on-premise where the matter requires it — so the confidentiality answer is architectural, not contractual.
Document comparison is the workflow that proves the pattern: deterministic where it can be, model judgment where it must be, self-verified, human-decided. The same foundation carries due diligence bundles, clause libraries, precedent retrieval — modules of the firm's own application, not a stack of vendor tools. The knowledge stays in-house.
What took hours of side-by-side reading arrives as a verified, cited summary read in minutes.
The diff is deterministic and exhaustive — including changes hidden by renumbering and moved text.
One human decision on every comparison. The lawyer's review is preserved as the last check.