Yes. An AI system can help compare an amended contract with its signed version, identify changed clauses and flag missing attachments. It should produce a verifiable change-review pack, not a legal verdict. Each finding needs evidence from the relevant version, clear page references and a route for human review. The proposed workflow below keeps interpretation and acceptance with the responsible professional.
Confirm which documents belong in the comparison
Start by having the contract owner identify the signed baseline and the proposed amendment. A filename such as “final signed” is not enough to establish which document governs the comparison.
Ask the owner to confirm the parties, agreement description, signing date and complete set of schedules. Record whether the new document is a full replacement contract, a marked-up draft or a short amendment referring to selected clauses. These require different comparisons. A short amendment should not be reported as deleting every clause it does not repeat.
Create a document inventory before extracting text. Include the original files, their page counts, attachment names and where each attachment appears. Keep separate records for attachments embedded in a PDF and files delivered separately.
Proposed intake rules should stop comparison when the baseline is disputed or the wrong parties appear. The output can state “baseline unresolved” and list the documents needing confirmation. It must not silently choose the most recent file.
This is a scoped document processing use case: prepare evidence for a decision rather than make the decision.
Extract wording without losing the original pages
Keep the original documents unchanged and connect extracted passages back to their pages. OCR, which converts scanned text into machine-readable text, is a preparation step rather than proof that the wording was read correctly.
Source: Microsoft’s Document Intelligence overview describes models for extracting text, tables and document structure. Its contract model extracts agreement and party details. Those capabilities do not establish a native, complete contract-version comparison process.
For this proposed workflow, retain the PDF page number, printed page label where present, clause heading and exact extracted passage. A reference should distinguish “PDF page” from a page number printed in the footer, particularly when cover sheets change pagination.
Inspect scans for cropped margins, faint wording, handwritten additions and broken tables. If a notice period is unreadable, record the uncertainty and request a better scan. Do not reconstruct it from surrounding clauses.
Before selecting a cloud service, the technical and security owners should confirm regional availability, account access, retention arrangements and approved handling of confidential contracts. The supplied overview alone does not establish suitability for a particular South African business.
Match clauses by content, not only numbering
Use clause numbers as clues, then confirm matches through headings and wording. Renumbering can make an unchanged clause look deleted and a relocated clause look newly added.
Build an inventory of clauses and subclauses for both documents. Apply a text comparison first, then use AI to propose matches where headings, numbering or layout differ. Keep original punctuation, figures and units in the evidence, even if a working copy removes repeated headers or formatting noise.
Proposed change categories are: wording changed, added, deleted, moved, formatting only and unresolved match. Do not collapse these into one “different” label. A moved clause with altered wording should carry both observations.
Source: OpenAI’s file search documentation describes semantic and keyword retrieval from uploaded files, with file citations. Retrieval can help locate a related passage, but a retrieved passage is not proof that every clause was compared. File citations also do not replace a page map built from the originals.
Require an inventory outcome for each clause, including unchanged clauses. This makes gaps visible instead of hiding them behind a short summary.
Check attachments and cross-references separately
Treat the attachment inventory as its own comparison, because an unchanged contract body can still refer to a different or absent schedule.
Extract references to annexures, schedules, exhibits and documents incorporated by reference. Match each reference to an actual supplied item. Compare titles, version labels, dates and relevant content, rather than relying on filenames alone.
Use precise absence language. “Schedule B was not found in the supplied amended pack” describes the available evidence. “Schedule B has been removed from the agreement” makes a contractual inference that may not be justified.
If an amendment says the existing schedules remain unchanged, list those schedules as retained by reference and confirm that the reviewer has access to them. Do not flag them as deleted merely because they were not attached again.
Duplicate attachments need a separate exception. If two files share a title but contain different wording, preserve both and ask the owner which belongs to the amendment. If the files are byte-identical, a technical duplicate check can identify that fact, but the inventory should still explain how the duplicate was handled.
Build evidence-backed findings, not legal conclusions
Write each finding as an observable difference followed by a review question. “The proposed wording changes the notice period” is appropriate. “The new clause is enforceable” is a legal conclusion and belongs with a qualified professional.
For every changed passage, include the baseline excerpt, amended excerpt, references to both originals and the change category. For an addition or deletion, cite the available passage and state that no counterpart was found within the confirmed comparison scope.
Source: OpenAI’s structured outputs documentation describes responses constrained to a supplied JSON schema. This can support consistent fields in a change register. A correctly shaped record still needs evidence checks: schema compliance does not establish that a quotation, match or conclusion is correct.
Keep factual findings separate from professional notes. The reviewer can record commercial concerns, legal questions or negotiation instructions without making them appear to be machine-verified facts.
A proposed rule is to leave any unsupported finding unresolved. Never fill a missing citation with a plausible page number or convert uncertainty into “no change”.
Route exceptions to the person who can resolve them
Send each exception to someone with authority over the missing information or decision. More model processing is not a substitute for an owner confirming the document set.
Use the following proposed routing rules:
| Exception | Expected human handling | Comparison treatment |
|---|---|---|
| Signed baseline disputed | Contract owner confirms the baseline | Hold the overall comparison |
| Passage unreadable | Document owner supplies a clearer copy | Leave the affected finding unresolved |
| Clause match ambiguous | Reviewer checks both passages | Preserve candidate matches |
| Referenced attachment absent | Owner confirms or supplies the attachment | Mark the pack incomplete |
| Competing attachment versions | Owner identifies the intended version | Do not select automatically |
| Changed payment or liability wording | Responsible professional assesses consequences | Record the difference without acceptance |
Distinguish preparation from actions such as sending a revised contract, accepting terms or changing payment instructions. Those actions require appropriate human judgement and separate authority.
The AI agents versus automation comparison is useful when deciding whether this needs a fixed pipeline or flexible retrieval. For vocabulary, see the custom AI agent glossary entry. The name of the system does not change who must make the contractual decision.
Evaluate coverage before relying on the pack
Evaluate the proposed process against a reference comparison prepared by a competent reviewer. Measure whether it finds the known differences, preserves evidence and exposes uncertainty, rather than whether its summary sounds convincing.
Choose authorised examples covering clean digital files, scans, renumbered clauses, changed tables and incomplete attachments. Include unchanged documents so the evaluation can reveal false findings. Ask the reviewer to mark expected changes before examining the system’s output.
Track missed changes, incorrect matches, false change reports, quotation errors, broken page references and missed attachment exceptions. Separate extraction failures from matching failures so the team knows what to fix.
A proposed release condition is that unresolved evidence gaps remain visible and no finding without a checked reference is presented as verified. The responsible professional should set any further tolerances according to the consequences of a mistake.
Record reviewer effort as well as machine output. The process may help organise review, but vendor capability does not prove reduced errors or saved time. For implementation planning, the custom AI agents workflow resource can inform the wider design without replacing this evaluation.
Reusable change-review pack checklist
Use this proposed template for each comparison, keeping incomplete items visible rather than treating them as passed checks.
Pack identity
- Agreement: [name and parties].
- Signed baseline: [file, signing date and owner confirmation].
- Amendment: [file, date and document type].
- Scope: [full replacement or selected amendments].
Evidence inventory
- Preserve original files and record page counts.
- List embedded and separate attachments for each version.
- Map extracted passages to PDF pages and printed labels.
- Record unreadable pages and competing document versions.
Change register fields
| Field | Required entry |
|---|---|
| Finding | Unique reference and change category |
| Baseline | Exact excerpt, file, PDF page and clause |
| Amendment | Exact excerpt, file, PDF page and clause |
| Evidence gap | Unreadable, absent, ambiguous or none identified |
| Review question | Specific question without a legal conclusion |
| Human handling | Responsible person, verification note and next action |
Coverage and handover
- Give every baseline clause a comparison outcome.
- Check new clauses, cross-references and attachment versions.
- Verify excerpts and references against original pages.
- List unresolved exceptions and their owners.
- Separate factual differences from professional interpretation.
- Record the responsible professional’s response separately from AI findings.
Completion rule: evidence preparation is complete only when coverage is checked and every exception is resolved or explicitly carried forward. Completion does not mean contractual acceptance.
Hypothetical walkthrough: a clear change and incomplete evidence
A useful pack reports a clear change normally while leaving incomplete evidence unresolved. All names, wording, page references and figures in this example are hypothetical.
Suppose a South African supplier provides a signed service agreement and a proposed full replacement. The contract owner confirms both files and the original schedules.
The signed agreement states “30 calendar days” in clause 8 on PDF page 6. The replacement states “45 calendar days” in clause 9 on PDF page 7. The headings and surrounding wording support the match despite renumbering. The expected finding contains both exact excerpts and references, classified as wording changed and renumbered. The reviewer verifies the pages, then assesses the commercial and legal consequences. The system does not recommend acceptance.
The replacement still refers to Schedule B, but no Schedule B appears in its supplied pack. The expected exception is “referenced attachment not found”, not “schedule deleted”. The owner must confirm whether the original remains applicable or provide the replacement.
Meanwhile, two files labelled Schedule A contain different service descriptions. The pack lists both as competing versions. The owner identifies the intended attachment, after which comparison resumes. Until then, the reviewer can assess the notice-period change, but the pack remains incomplete and should not be presented as a complete comparison.
Frequently asked questions
The central distinction is between finding evidence and deciding what that evidence means.
Can it compare a short amendment rather than a replacement contract?
Yes, within a confirmed scope. Map each amendment instruction to the signed clause it names, then show the existing passage alongside the proposed change. Keep clauses outside that scope marked as not amended by the supplied document, rather than deleted. If the instruction points to a clause number that does not exist, the reviewer must resolve the reference before the system proposes a match as verified.
What if pagination changes between the signed and amended versions?
Use separate page references for each original. Do not assume that a clause remains on the same page after new cover sheets or paragraphs are added. Show the file name, PDF page number, clause heading and printed page label where available. The reviewer should be able to open either original and locate the quoted wording without relying on an automatically generated summary.
Does a signature image prove that the baseline is authentic?
No. This comparison workflow does not establish authenticity, signing authority or execution validity. It can record that a signature area is present, but the contract owner or responsible professional must confirm the baseline using appropriate records and judgement. If authenticity is disputed, hold the comparison as an authoritative change-review pack. Any preliminary observations should remain clearly limited to the supplied files.
Decide whether this workflow fits your business
Use this approach when the team needs traceable differences and attachment checks before professional review. A conventional document comparison may be enough for clean, similarly formatted files; AI is worth evaluating where matching and evidence organisation are harder.
If your business needs help defining this workflow, explore Symaxx’s AI automation services and get in touch to discuss the document set, exception handling and review boundary. Start with authorised examples and an agreed evaluation, not an assumption that the system can accept contract changes.