“It is only a formatting update—could you take a look?” often arrives with a long contract and a new version. Page numbers and headings look unchanged. The meaningful edit may be “shall” becoming “may,” thirty days becoming ten, or an auto-renewal clause moving into an appendix. Version comparison is a strong use case for AI, as long as it is treated as a locating tool rather than legal advice.
This workflow is for business, procurement, sales operations, and project owners. It helps AI find differences, group them by business area, and produce questions for legal review. It does not determine legality, enforceability, or legal risk. Important contracts, cross-border transactions, personal-data terms, IP clauses, and disputes still require qualified professional review.
Make the files comparable first
Do not upload two scanned PDFs and ask what changed. Confirm the source, date, file name, and page count. OCR scanned files and sample-check money, dates, company names, and appendix numbers. A Word export compared with a scan can create a great deal of noise from line breaks, headers, and numbering.
Create a record for every version: version label, received date, sender, archive path or hash, included appendices, and OCR check status. Quotes, data-processing agreements, order forms, and schedules are often sent separately. Confirm which files belong to this comparison before asking AI to work on the wrong set.
Use a conventional diff first, then AI for structure
Word tracked changes, PDF comparison, or a text diff prove which characters changed. AI is better at turning those changes into a readable checklist. Provide both old and new clauses, clause number, page, one paragraph of surrounding context, and the business areas you care about: payment, term, renewal, termination, liability, confidentiality, data, IP, delivery, and disputes.
Ask for evidence, not conclusions
You are a contract-version checking assistant and do not provide legal advice. Compare each old and new text pair. List only changes directly supported by the text. Group them by payment, term and renewal, termination, liability, confidentiality and data, IP, delivery and acceptance, dispute resolution, or other. For every item return clause number, old text summary, new text summary, change type, and a question that business or legal should confirm. Do not judge legality, fairness, or legal consequences. Mark insufficient evidence for human review.
ChatGPT or Claude can structure the comparison. Assess data permissions before upload. Unsigned contracts, contacts, bank details, internal pricing, and confidential materials should not be put into a public service. Use an approved workspace and redact according to internal policy.
Separate changes that need action from changes that need understanding
Formatting and spelling can be handled together. Put payment terms, dates, auto-renewal, termination, liability caps, indemnity, data roles, security-notice duties, audit rights, ownership, licences, confidentiality, and governing law in the priority queue. AI may flag an item for attention, but should not call it “low risk.” Risk depends on the transaction, counterparty, law, and your ability to perform—details not contained in two clauses.
Hand off an issue list, not a long summary
Every item should contain the clause, change, source location, owner, required approver, due date, and final decision. “Auto-renewal changed from 12 to 36 months” should lead to questions: Is a three-year commitment acceptable? Is there a cancellation window? Do pricing and budget cover it? If a definition changes, search all uses of the term. AI can suggest affected clauses; a person must confirm the meaning actually changed.
Run a reverse coverage check
Using this difference table and both contract tables of contents, identify high-attention areas that may not be covered: payment, term, renewal, termination, liability, confidentiality, data, IP, delivery, governing law, and appendices. Do not add legal conclusions. Return only clause numbers that require human review, why they may have been missed, and the original text location to inspect.
The purpose is to find omissions, not let the same model grade itself. Open the source text for every item. Check footers, tables, appendices, signature pages, online terms linked from the document, and differences between language versions.
Pre-signing checklist
- Are both versions, all appendices, and dates complete and traceable?
- Were numbers, names, clause references, and tables sample-checked after OCR?
- Does each high-attention change have an owner, source evidence, and explicit decision?
- Are AI classifications clearly separated from business or legal conclusions?
- Does the final signed version match the reviewed version, including schedules and signature pages?
The time saver is not asking AI to say “the contract is fine.” It is turning easily missed differences into a traceable work list. If document fields are messy, begin with the AI spreadsheet cleanup workflow to normalise dates, IDs, and fields before comparison.