To compare contract versions with AI, first preserve the two authoritative files and create a deterministic document comparison. Then use AI only to organize clause-level changes and draft neutral descriptions. Verify every ledger entry against both originals, and send consequential or unclear changes to a qualified lawyer. AI output is not legal advice.

This workflow is narrower than summarizing a document with AI. A summary describes one document; a version comparison must prove what changed between two identified texts. The general AI guide covers prompt and review basics. Here, the evidence is the old wording, the new wording, and an exact source location.

Key Takeaways

Keep immutable copies and explicit version labels before comparison.

Use a deterministic diff as the baseline; do not ask AI to remember both files unaided.

Record additions, deletions, replacements, moves, and formatting-only changes separately.

Describe changes neutrally instead of interpreting their legal effect.

Verify every entry against both originals.

Escalate liability, money, dates, termination, indemnity, and governing-law changes.

What authorization and confidentiality checks come first?

Contracts may contain names, signatures, pricing, trade secrets, security details, personal data, and negotiation positions. Confirm who may access each version and whether the chosen AI service is approved for that information. The ABA's guidance on generative AI highlights duties that can include competence, confidentiality, communication, and supervision when lawyers use these tools.［2］ Even if you are not a lawyer, those risks explain why “paste the full contract into a chatbot” is a poor default.

Use the least information necessary. Prefer an approved private workspace, local redaction, or clause excerpts with stable identifiers. Remove signatures, contact details, account numbers, hidden comments, tracked-change identities, and unrelated schedules when they are not needed. Do not upload a complete contract merely for convenience.

Write a comparison header before processing anything:

Field	Record
Matter or project	Neutral internal identifier
Old version	Filename, date, and checksum or document-system ID
New version	Filename, date, and checksum or document-system ID
Authorized scope	Included clauses and excluded attachments
Tool boundary	Approved environment and retention setting
Review owner	Person accountable for verification

This header prevents a later draft from being mistaken for an authoritative version.

How do you freeze the originals and build a deterministic baseline?

Copy both authoritative files into a read-only comparison workspace. Give them unambiguous labels such as supplier-agreement-v3-received and supplier-agreement-v4-proposed. Keep the originals unchanged. If the files are scans, retain the scans and treat any OCR text as a derivative that needs page-by-page checking.

Run a deterministic document comparison before using AI. Microsoft Word's legal blackline option compares two documents and produces a third document showing differences while leaving the source files unchanged.［1］ Other approved diff tools can serve the same purpose if they preserve additions, deletions, and locations reliably.

Check the baseline for common traps:

headers, footers, footnotes, tables, and appendices omitted from extraction;
numbering changes that make unchanged clauses look moved;
formatting changes mixed with substantive edits;
scanned text recognition errors;
comments or tracked changes that are not part of the agreed text;
an attachment referenced in one version but missing from the package.

The diff is evidence of textual change, not a conclusion about legal meaning.

Align clauses before asking AI to classify changes

Split the comparison by stable units: clause number, heading, schedule, table row, or paragraph anchor. Do not use arbitrary token windows that separate a defined term from the clause that uses it. When numbering changed, record both the old and new locator.

Define a small vocabulary before prompting:

Change type	Meaning
Added	New text appears only in the new version
Deleted	Old text has no equivalent in the new version
Replaced	Wording changed within the same logical clause
Moved	Substantially identical text has a new location
Reference changed	A cross-reference, definition, or attachment link changed
Formatting only	No verified wording change
Uncertain	Alignment or extraction is ambiguous

Use “uncertain” rather than forcing a confident label. A moved paragraph can also contain a replacement; record both facts when necessary.

Build a contract change ledger

The ledger is the primary deliverable. It must let another reviewer reproduce each observation without trusting the model.

Clause	Old text	New text	Type	Neutral summary	Review question	Source location	Verified by
4.2	Exact excerpt	Exact excerpt	Replaced	Notice period changed from stated old value to stated new value	Does this require legal or operational review?	Old p. 6; new p. 7	Initials/date

Quote only enough text to prove the change, subject to your confidentiality rules. Preserve punctuation, defined terms, numbers, and qualifiers exactly. A neutral summary says “the notice period changed from 30 to 15 days,” not “the supplier can now terminate unfairly.” The second sentence is an interpretation and may depend on other clauses and applicable law.

Include unresolved alignment problems as rows. Silence is dangerous: a blank entry can look like “no change” when it actually means “not reviewed.”

Compare contract versions with AI for bounded organization

Provide the model with the schema, allowed change types, paired excerpts, and strict instructions. A useful prompt is:

You are organizing a textual comparison, not giving legal advice.
For each paired excerpt:
1. preserve the clause locators;
2. classify only the visible textual change;
3. quote the minimum old and new text needed to prove it;
4. write one neutral factual summary;
5. list a verification question without answering it;
6. mark uncertain alignment as UNCERTAIN.
Do not infer enforceability, risk, intent, fairness, jurisdiction, or negotiation strategy.
Return the contract change ledger columns exactly.

Process manageable batches and keep clause context. Ask the model to flag inconsistent defined terms, changed cross-references, unmatched schedules, and numbers that differ. Treat each flag as a search lead, not a finding.

NIST's AI Risk Management Framework emphasizes governing, mapping, measuring, and managing AI risk across context and use.［3］ Applied here, that means defining who may use the tool, what evidence goes in, how errors are measured, and who accepts the final record.

How do you verify every row against both originals?

Review the ledger from top to bottom. For each row:

open the cited old location and confirm the excerpt exactly;
open the cited new location and confirm the excerpt exactly;
confirm that no nearby sentence, footnote, definition, or table cell changes the context;
confirm the change type and neutral summary;
mark the reviewer and date;
leave the row unresolved if any source cannot be located.

Use a second comparison method for high-impact sections when practical. Search separately for changed numbers, dates, currency symbols, percentages, “shall/may,” exceptions, and negations. AI can omit a small word that reverses meaning. Follow the same source-first discipline used to fact-check AI answers.

Do not accept “no other differences found” from a model as completeness evidence. Completeness comes from the deterministic baseline, scope accounting, and human reconciliation.

Which changes need qualified legal review?

Create a mandatory professional-review queue for additions, deletions, or ambiguity involving:

liability limits, exclusions, warranties, or remedies;
fees, payment dates, interest, credits, or price adjustment;
renewal, suspension, termination, notice, or survival;
indemnity, insurance, audit, or compliance obligations;
intellectual property, data rights, confidentiality, or security;
governing law, venue, arbitration, or dispute procedure;
employment, consumer, tax, privacy, or regulated obligations;
any text the reviewer cannot align confidently.

Label these rows “needs qualified legal review.” Do not label them “safe,” “standard,” “enforceable,” or “high risk” unless a qualified reviewer makes that determination in the proper context. The Solicitors Regulation Authority warns that misuse of AI can produce false or misleading material and stresses checking outputs.［4］

Keep decisions separate from the comparison

The ledger answers “what text changed?” A decision record answers “what will we do?” Keep them separate so neutral evidence is not overwritten by negotiation comments.

A decision record can include the ledger row ID, business owner, legal reviewer, decision, rationale, required redline, deadline, and approval. Limit access because it may contain privileged or strategically sensitive material.

Version the ledger itself. If a fifth contract version arrives, compare version four with version five in a new run; do not silently replace the old baseline. Preserve a chain that shows which pair was compared.

Add a reconciliation summary that accounts for the complete comparison scope. Record how many diff blocks were reviewed, how many became ledger rows, how many were confirmed as formatting-only, how many were merged because they belonged to one clause, and how many remain unresolved. The counts do not establish legal significance, but they expose omissions between the deterministic output and the human ledger. Have the reviewer explain every difference between the baseline count and the ledger count. For example, one renumbered section may generate several mechanical diff blocks that become one moved-and-modified row; that consolidation is acceptable when documented. Conversely, a diff block must never disappear because it looked trivial to the model. Reconcile attachments separately so a missing schedule is not hidden behind the main agreement's totals.

Final quality checklist

Both authoritative versions are frozen and clearly identified.
The deterministic comparison covers the stated scope, including attachments.
Every ledger row cites exact old and new locations.
Formatting-only and uncertain changes are not mixed with verified replacements.
Numbers, dates, definitions, cross-references, and negations received explicit checks.
Each AI-generated row has a named human verifier.
Consequential changes are queued for qualified legal review.
Confidential material stayed inside an approved boundary; review AI privacy risks before sharing excerpts.
The comparison record and the business or negotiation decision record remain separate.

FAQ

Is an AI contract comparison legal advice?

No. It is an information-organization aid. Legal meaning, enforceability, jurisdiction, and negotiation decisions require a qualified lawyer.

Can I skip the deterministic diff?

No. The diff establishes the textual baseline. AI may organize it, but should not be the only method used to detect changes.

Should I upload the complete contract?

Only if you have authorization and an approved environment. Prefer minimum necessary excerpts, redaction, and controlled retention.

Can AI decide whether a change is risky?

It can flag categories for review, but it should not decide legal risk. Context elsewhere in the agreement or applicable law may change the answer.

How do I compare scanned contracts?

Retain both scans, create OCR derivatives, and verify every quoted excerpt against the page image. Record uncertain recognition instead of correcting it silently.

What if clause numbers changed?

Record both old and new locators and compare the text and surrounding definitions. Use a moved-plus-modified classification when both occurred.

Who should verify the ledger?

A person authorized to access both versions and competent to check the text. Consequential or ambiguous rows should go to qualified legal counsel.

What should the final output contain?

A verified change ledger, unresolved questions, reviewer identities, scope notes, and a separate record of any approved decisions.

Sources

Microsoft Support, “Compare document differences using the legal blackline option” — https://support.microsoft.com/en-US/Word/compare-document-differences-using-the-legal-blackline-option
American Bar Association, “Formal Opinion 512: Generative Artificial Intelligence Tools” — https://www.americanbar.org/content/dam/aba/administrative/professional_responsibility/ethics-opinions/aba-formal-opinion-512.pdf
NIST, “AI Risk Management Framework” — https://www.nist.gov/itl/ai-risk-management-framework
Solicitors Regulation Authority, “Warning notice: The misuse of artificial intelligence” — https://www.sra.org.uk/solicitors/guidance/misuse-ai/

Sources checked 31 August 2026.

How to Compare Contract Versions with AI—Not Legal Advice