Two associates spend an afternoon comparing a contract draft against the version opposing counsel sent back, reading clause by clause to spot what changed. By the fourth review cycle, the redline history is broken, the formatting has shifted twice, and nobody is fully sure the indemnification language they are relying on is the version everyone actually agreed to. A missed change here is not a typo — it is a liability clause, a payment term, or a termination right that ends up in a signed contract nobody meant to sign. Contract comparison is the process of identifying every substantive difference between two versions of the same document, not just the differences a word processor's track-changes feature happens to flag.
What makes manual contract comparison error-prone?
Manual comparison fails most often at clause renumbering and defined-term drift, not at obvious rewrites. When one party inserts a new subsection, every clause after it shifts — Section 8 becomes Section 9 — and a word-processor redline shows that as a wholesale change to every renumbered clause even when the text itself is identical. Reviewers learn to skim past these false positives, and that habit is exactly what causes a real change buried in a renumbered section to get skimmed past too.
Defined-term drift is quieter still. A term like “Confidential Information” gets defined once, early in the document, and referenced dozens of times after. If a later draft narrows that definition — say, by adding a carve-out for information the receiving party already knew — every downstream reference to the term changes meaning without a single word changing at the point where it is used. A line-by-line read of the body text will not catch that; only checking the definition itself, deliberately, will. A contract with even five or six defined terms and three redline rounds routinely accumulates changes that never show up as a highlighted edit anywhere in the document.
How does grounded AI compare two contract drafts?
Grounded AI comparison works by retrieval, not by a visual diff. You upload both drafts as separate documents, then ask a scoped question of each — “What does the indemnification clause say in the March 3 draft?” and the same question against the March 10 draft — and get back an answer built only from that document's own text, with a citation to the exact page. This is a meaningfully different workflow from an automatic redline tool: you are asking targeted questions and comparing cited answers side by side, not waiting for software to highlight every character-level change. What it explains well is covered in an earlier piece on what grounded AI means for legal document research — the same retrieval-plus-citation mechanism that answers a research question also answers a comparison question, one clause at a time.
The advantage over a plain read is that the citation forces precision. If the assistant cannot find an indemnification clause in one of the two drafts, it says so instead of guessing — and that refusal is itself useful information, because it usually means the clause was renamed, merged into another section, or genuinely removed. A visual skim tends to assume a clause is “probably still there somewhere” and moves on; a grounded answer either finds it and cites the page, or tells you plainly that it did not. That distinction — a cited answer versus an honest refusal, never a guess — is the whole difference between a tool you can rely on for a live deal and one that quietly makes something up.
What is a comparison workflow that catches renumbering and defined-term drift?
The following sequence works for a two-draft comparison of any length, and scales to three or more rounds of redlines on the same matter:
- Upload both drafts as separate documents in the same matter, named clearly by date or round (“Draft 3 – March 3”, “Draft 4 – March 10”).
- Ask the same scoped question against each draft, one clause at a time — indemnification, termination, limitation of liability, assignment — rather than one broad “what changed” question.
- Read both cited answers side by side, noting the document name and page for each.
- Ask a dedicated question about each key defined term (“Where is Confidential Information defined, and what does the definition say?”) against both drafts, since body-text questions alone will not surface a narrowed definition.
- Treat a refusal on one draft and a clear answer on the other as a flag to investigate manually — it usually means the clause moved, was renamed, or was cut.
- Record the clause-by-clause findings in your own notes before circulating a summary, so the paper trail shows exactly what was checked.
This is slower than trusting a single automated redline pass, and that is deliberate: each answer is grounded in one document's actual text and cited to a page, so the reviewer is verifying a specific claim rather than trusting a highlight color.
Manual read vs. grounded question-and-answer
| Method | Catches | Misses | Time for a 40-page contract |
|---|---|---|---|
| Side-by-side visual read | Obvious wording changes | Renumbering noise, quiet definition drift | 1–2 hours |
| Track-changes redline | Character-level edits | Cross-references broken by renumbering; anything outside the diffed pair | 20–40 minutes, plus review of false positives |
| Clause-by-clause grounded questions | Specific clause language, defined-term drift, missing clauses (via refusal) | Changes in a clause you did not think to ask about | 15–30 minutes for the key clauses |
Should a firm trust an AI tool with a live deal draft?
This is the right question to ask before uploading anything, and the honest answer depends on what the vendor actually does with the text, not what it claims to do. Vorticel encrypts each firm's documents at rest with per-firm libsodium encryption, keeps every firm's data in its own tenant boundary, and does not use uploaded documents to train any model. Document text is sent to Voyage AI to generate search embeddings, and only the specific passages retrieved for a question are sent to Anthropic to draft an answer — the original file itself never leaves encrypted storage. If your firm's IT or compliance reviewer wants the full list of questions to put to any AI vendor before a live draft goes anywhere near it, the earlier post on questions to ask an AI vendor before uploading client documents covers that ground in full.
The honest caveat: this workflow answers questions about what is in the documents you uploaded. It does not draft the negotiation response, and it does not replace an attorney's judgment on whether a change is acceptable — it replaces the tedious, error-prone part of finding what changed so the attorney's judgment gets applied to the right passage the first time.
Doing this by hand on every draft of every matter does not scale past a handful of active deals, which is the point where clause-by-clause grounded search stops being a nice-to-have and starts saving real review hours. Vorticel's documentation covers how to structure a matter with multiple document versions so this workflow is fast from the first upload.
How do you get started with this on your own drafts?
Upload the two most recent drafts of one active matter, ask the indemnification and termination questions against both, and compare the cited pages. That single pass tells you within a few minutes whether the workflow catches what your current process misses. Start a 14-day free trial, no card required, and run it on a real contract you are already comparing this week.