AI Agent for Document Version Comparison
A redline shows that words moved. It does not tell you that a renumbered clause now points at a different schedule, or that the other side accepted your tracked changes and then edited the clean copy. This agent compares two versions of a contract, policy or specification, finds every change including the untracked ones, explains what each one does, and flags the edits a reviewer has to read.
What is an AI document comparison agent?
An AI document comparison agent is a governed software worker that compares two versions of a document and explains how they differ. It parses both files, pairs their clauses, runs a deterministic diff that also catches untracked edits, describes what each substantive change does with the old and new wording quoted, and hands the result to a reviewer.
What it does
What it is not
The redline only shows the edits someone chose to track
Negotiated documents pass through several hands, and any of them can switch off tracking, accept changes in bulk or paste in a clean section from another draft. By the fourth round nobody can say with confidence how the file differs from the version the business last approved, so reviewers either re-read the whole document or trust the markup they were sent.
Untracked edits slip through
A returned draft can carry changes made with tracking switched off, and a reviewer reading only the markup will never see them.
Renumbering hides real changes
When clauses move or are renumbered, a line-by-line compare reports whole sections as new and buries the one sentence that changed.
Small words carry the risk
Shall becomes may, or a cap moves from fees paid to fees payable, and a word-level redline shows the edit without saying why it matters.
Comparisons leave no record
Nobody notes which two versions were compared or what was concluded, so the same check is repeated from scratch at signature.
A comparison that says what changed and why it matters
Detection
Compares The Text, Not The Markup
Untracked edits included.
Both versions are parsed into clean text with their headings and clause numbers, paired clause to clause, and compared with a deterministic diff. Because the comparison runs on the content rather than on the tracked changes, an edit made with tracking switched off is found exactly like a marked one.
- Clauses paired before comparison
- Moved and renumbered text recognised
- Untracked edits reported like tracked ones
- The raw diff kept as evidence
Every edit found
Explanation
Each Change Described In Plain Words
What it does, not just where.
For every difference the agent writes a short note on its effect: an obligation that became discretionary, a notice period that shortened, a defined term widened so that every clause using it now reaches further. The old and new wording sit beside each note, so the reader can check the description against the text.
Side by side
Triage
Substantive Changes Flagged First
Cosmetic edits grouped apart.
Changes are sorted into substantive, editorial and formatting groups, and the substantive ones are tagged by subject, such as liability, payment, term, data or scope. The reviewer opens a short list of edits that need a decision, with punctuation fixes and renumbering collected at the end.
Reviewer decides
How the AI Document Comparison Agent runs a task
- STEP 01
Parse both versions
Each file is converted to clean text with its headings, numbering and tables preserved, whether it arrives as a Word draft, a PDF or a scanned execution copy. Scanned pages are read with optical character recognition and marked, so a misread character is not mistaken for an edit.
Word and PDF parsingOCR for scans - STEP 02
Pair the structure
Clauses in the old and new version are paired by content as well as by number, so a section that moved from 9 to 11 is compared with itself instead of being reported as one deletion and one addition. Clauses with no partner are listed as genuinely added or removed.
Clause pairingMove detection - STEP 03
Run the diff
Paired clauses are compared with a deterministic text diff that lists every insertion, deletion and replacement exactly. The diff, not the language model, decides what counts as a change, so nothing is dropped because a summary judged it too small to mention.
Text diffChange register - STEP 04
Explain and classify
The model reads each change in its context and writes a note on its effect, following any defined term the change touches through the rest of the document. Changes are then grouped as substantive, editorial or formatting and tagged by subject, with both wordings kept beside every note.
Plain-language notesMateriality grouping - STEP 05
Hand to a reviewer
The change register, an HTML side-by-side view and a short memo go to the named reviewer, who confirms or corrects each classification. Neither source file is edited, accepted or rejected, and the record keeps which two versions were compared, by whom and when.
HTML comparisonReviewer sign-off
Systems the AI Document Comparison Agent connects to
Document sources
Parsing and comparison
Inputs, outputs and runtime
- Ingests
- Original and revised versionsWord, PDF or scanned copiesLibrary or folder locationReviewer and matter details
- Produces
- Change registerPlain-language change notesHTML side-by-side viewWhat-changed summary memoUntracked edit list
- Triggered by
- Draft returned by counterpartyExecution copy receivedPolicy revision submitted
- Human oversight
- A reviewer confirms every substantive change
- Models
- Open-weight LLMs you host, plus a deterministic diff
- Typical latency
- Minutes for a typical agreement
- Deployment
- On-premise or private cloud
- Data residency
- Both versions are processed inside your network
Where an AI document comparison agent earns its place
Returned Counterparty Drafts
Compare the draft that came back against the one you sent, so edits made outside tracked changes surface before signature.
Execution Copy Check
Confirm that the signature version matches the last agreed draft word for word, and list anything that does not.
Policy Revision Notes
Produce an accurate what-changed note for a revised internal policy, ready for the owner to circulate with the new version.
Specification Change Review
Show engineering and suppliers which requirements were added, removed or tightened between two specification releases.
RFP Addendum Tracking
Read each tender addendum against the original pack and list the requirement changes the bid team now has to answer.
Regulatory Text Updates
Compare a consultation draft of a rule with its final text and highlight the provisions that changed in between.
AI Document Comparison Agent vs chatbots and SaaS copilots
Asking a model what changed between two long drafts produces a fluent summary of the changes it happened to notice, and the risk in a negotiated document usually sits in the edit nobody mentioned.
| Generic chatbot | SaaS copilot | VDF AI | |
|---|---|---|---|
| How changes are found | Model reads both texts | Summary of the open file | Deterministic diff of the content |
| Untracked edits | Often missed | Depends on the prompt | Found like any other change |
| Moved clauses | Inconsistent | Varies by product | Paired and compared in place |
| Explanation | Free-form summary | General summary | A note per change, wording quoted |
| Materiality | Not separated | Not separated | Substantive changes flagged first |
| Record kept | Chat history | None | Versions, register and reviewer |
| Where drafts are read | Vendor service | Vendor tenancy | Inside your own perimeter |
Governance and controls
A comparison is only useful if the reviewer can trust that it is complete, so finding changes is left to a deterministic tool, describing them is left to a model, and both results stay on the record side by side.
Deterministic change detection
The diff, not the model, lists changes
Source files untouched
Neither version is edited or accepted
Wording quoted
Old and new text beside every note
OCR passages marked
Scanned text flagged for checking
Reviewer recorded
A named person confirms each class
Requester access respected
Files read with the user’s own access
Evidence it leaves behind
What reviewers notice after rollout
Who runs the AI Document Comparison Agent
Commercial lawyer
Receives the returned draft with every difference from the version sent already listed, including two edits made outside tracked changes, and spends the review on the clauses that moved rather than proofreading forty pages for surprises.
Policy owner
Publishes a revised policy with an accurate what-changed note drafted from the actual comparison of both versions, so staff read three paragraphs about the changes instead of working through the whole document again.
Bid manager
Gets each tender addendum compared against the original pack soon after it arrives, with requirement changes listed by section, which keeps the compliance matrix current without a late-night manual read-through.
Questions about the AI Document Comparison Agent
What is an AI document comparison agent?
It is an agent that compares two versions of a document, such as a contract draft, a policy or a specification. It finds every difference whether or not it was tracked, explains in plain language what each change does, and separates the substantive changes from the editorial ones for a reviewer.
How is an AI document comparison agent different from a generic chatbot?
Pasting two drafts into a chatbot gives you a summary that can skip changes and cannot be checked. This agent runs a deterministic diff first, so every difference is listed, and only then uses a model to explain each one, with the old and new wording quoted beside the explanation.
Can an AI document comparison agent run on-premise on contract and policy data?
Yes. Draft agreements, unreleased policies and product specifications are among the most sensitive files a company holds, and the parsing, the diff and the explanations all run on servers you control, with models you host.
What does an AI document comparison agent produce, and in what format?
A change register listing every difference with its location, type and old and new wording; a plain-language note on each substantive change; an HTML side-by-side view; and a short memo naming the two versions that were compared and when.
Where does an AI document comparison agent fit in a governed AI programme?
It finds and explains changes; people decide what they mean. Accepting a change, negotiating it back or approving a policy revision stays with the lawyer or document owner, and grading a draft against your playbook belongs to the contract review agent.
How is this different from the AI Contract Review Agent?
The contract review agent reads one draft against your playbook and grades each clause against your approved positions. This agent compares two versions of the same document and reports what changed between them. The two often run in sequence: the comparison shows which clauses the other side touched, and the playbook review then grades only those.
Can it find changes that were not tracked?
Yes, because it never relies on the tracked-changes markup. Both files are reduced to their text and compared directly, so an edit made with tracking switched off, a section pasted in from another draft, or a change accepted before the file was returned all appear in the register like any other difference.
Does it compare PDFs and scanned documents?
It does. Text is extracted directly from digital PDFs, and scanned pages are read with optical character recognition. Passages read from an image are labelled as such in the register, because a misread character can look like an edit, and the reviewer is asked to check those against the original page.
Can it compare more than two versions of a document?
It compares two at a time, which keeps every register unambiguous. For a document that has been through several rounds, the agent compares consecutive pairs and can also compare the latest draft with the last version your side approved, which is usually the comparison that matters at signature.
Is there a free AI document comparison tool to try?
You can start on the VDF AI Cloud Free plan from this page and try the agent on documents that are not confidential. Live contract drafts usually belong on the on-premises edition, where the files never leave your servers, and an on-premises deployment can start with a three-month platform trial.
See every change between two drafts, tracked or not
Watch the AI Document Comparison Agent compare a returned draft with the version you sent.