Comparing versions of documents with AI: finding changes

For compare versions of documents with AI You need two identified copies, a method to detect differences and references that allow you to control them. The useful output is not a summary of the most recent document. It is a list of changes between A and B, with the text involved, the possible effect and what remains uncertain.

An offer, procedure or agreement may change to small but important points. Here we describe a verifiable flow for an SME, with official sources consulted on 9 October 2026 and hypothetical scenarios. The AI Guide for Companies helps to place it; the analysis of business documents It addresses the reading of a document in general.

What two versions are you really comparing?

Before reading the content, build a pair of comparisons. It records title, provenance, date of receipt, identification and order of versions. The name "definitive" is not enough: it may have been reused several times. If you don't know which copy is the previous one, the system must ask for a confirmation instead of inferring it from the name.

It also identifies the perimeter. A main agreement with three attachments is not the same thing as the main file alone. If a price table changes to an annex, the comparison must include it or state that it has not been examined. This limit must appear at the top of the report, not hidden after the conclusions.

The Customer Document Collection can prepare the package, but availability does not prove the equivalence of the copies. Two files with the same name may have different content; two PDFs with different names may be a new scan of the same original. The pair must be fixed before the review.

First the extraction, then the differences

For a digital document you can read text and structure; for a scan you need an extraction from the page. In both cases, it controls tables, headers, notes, and block order. A text read out of sequence can create a false difference or hide a condition that belongs to the previous line.

The Microsoft documentation on Document Intelligence functions describes options for extracting content and structure. It is a reference for the reading phase, not a guarantee that every PDF will be interpreted without error. The comparison must maintain the link to the original pages.

If an area is unreadable, report "not comparable" with the reference. Do not ask the model to complete it based on the previous version. It could just erase the change you're looking for. For an important document it is more useful to request a readable copy than to deliver a seemingly complete list.

What should the A/B report show?

A line must represent a controllable edit. "More favorable conditions" is too general if it does not indicate to whom, in what clause and on the basis of which text. The report should allow the operator to accept the report, correct its interpretation or mark it as a false alarm.

Keep the model notes and the person's decision separately. If the person responsible determines that a variation is irrelevant, record the reason without rewriting the source. A colleague must be able to understand later on why that change was accepted or called for new negotiation.

Text difference and change of meaning

Text comparison helps to see inserted, deleted, or replaced words. It is also useful when a paragraph has been moved. The semantic comparison proposes an interpretation: does it change who has to perform an activity, the time of delivery, a condition or the perimeter of the service? These are later questions, with a margin of error of their own.

A clearer rewrite can leave the content unchanged. On the contrary, a brief denial can change a condition. Avoid a system that only reports paragraphs with many different characters. The operational relevance does not coincide with the size of the variation. Table lines also require attention.

The techniques of Checking documents with Claude Opus 5.5 They can help to read internal inconsistencies. In this flow the question is different: what has changed between two copies? A coherent document within it may, however, contain a change that the company does not intend to accept.

Numbers, units and referrals deserve dedicated control

For amounts, percentages and quantities, a structured verification is added to the explanation. It records value, currency, unit and condition to which it belongs. An hourly rate and a price per intervention are not comparable just because they are in the same cell. The report must show the base change, not just the number.

Check dates and durations in context. One deadline can be calculated from one event, while the other is a fixed date. The model must not convert both into an invented day if the event is not known. If it is necessary to monitor the term after approval, the flow of Contract Deadlines is based on verified data.

Referrals are a frequent source of error. A clause may retain the same text but point to a modified attachment or renumbered section. Check if the reference exists and if the linked content is available. Report the broken link without automatically giving it legal meaning.

How automatic does the review have to be?

At first the system prepares the report and the person checks for the relevant changes. You can provide filters to focus on amounts, times, or tasks, but the report must declare the filter applied. A selective view should not be presented as an integral comparison of all parts.

The human review of tools in n8n is a possible component of the confirmation step. It does not make the content of the report correct. The person must see the exact couple and the proposal to approve, especially if the decision initiates a registration or other activity.

A change to the sources after the review must make the previous decision outdated. The project should prevent the approval of the A/B comparison from being reused for A/C. The link to the version is part of the operational data: it prevents the system from applying a decision to a file that arrived later.

Hypothetical scenario: a typography in Pesaro

In this hypothetical scenario a printer receives a new offer for a consumable. The main file looks similar to the previous copy, but the packaging and a surcharge have changed in the table. The comparison reports values and units in the two versions, referring to the original lines.

The purchasing department checks that the difference depends on the quantity per package and a new delivery condition. The model can prepare applications for the supplier, without deciding that the offer is convenient. A missing document remains visible as a limit. The economic review uses the confirmed data.

The pilot includes copies with broken tables and notes at the bottom of the page. It measures major changes lost, false alarms due to extraction and control time. The test numbers must be collected on the test: it is not assumed that the system will find every variation just because it produces a long report.

Hypothetical scenario: a service company in Trento

A second hypothetical situation concerns a procedure agreed with a client. The new version moves an activity from the customer to the supplier and changes the point where the confirmation takes place. Many words remain the same. The textual comparison identifies the substitutions and the proposed interpretation highlights the change of responsibility.

The Operations Manager shall verify whether the new step is sustainable and whether the said Annex contains up-to-date instructions. The decision is not to sign automatically. The system prepares a list of questions and records which changes have been accepted, with reference to the examined copy.

The sample also includes purely editorial variations, to avoid a report full of noise. The measure distinguishes ability to detect change and correctness of its explanation. A system can find the right point and misinterpret it: the two performances must be evaluated separately.

A correct report must declare coverage

Define which parts have been compared and at what depth. The main body, attachments, and tables may have different reading paths. The report must state what is complete, what is excluded, and what requires a new copy. A "comparison completed" title is misleading if an important part has remained unreadable.

For a long document, keep a map of the sections examined. If you divide the work, check the references that cross the boundaries between parts. A condition in the final section can modify a statement read at the beginning. The review must recompose the context before presenting a conclusion on the overall variation.

Predict a route to contest a report. The operator may indicate that a passage has been matched to the wrong section or that the model has misinterpreted a sentence. Keep this fix as part of the pair report, with a note on the reason. Do not treat it as a modification to the original text.

When the counterpart sends a third version, it creates a new comparison and makes previous reports recognizable. The decision made on the A/B pair must not appear as an endorsement of copy C. It may be useful to compare B/C for recent changes and A/C for the overall perimeter, clearly stating the pairs used.

In the pilot, enter at least one document with an excluded section and verify that the limit remains visible in the view used by the handler. Also check an exported report: Header, copy identity, and review status must travel with reports. A list extracted without these references may become unintelligible when shared internally.

If the report is corrected, make the revision recognizable and keep the reason. The recipient must be able to distinguish an updated report from the copy already read.

How to choose the model and delimit the pilot?

Choose documents of one type, with pairs controlled by those who know the process. Prepare a list of expected changes and non-comparable points. In the test, include tables, attachments, renumbers, and imperfect scans. Quality should not be measured only on short, clean copies.

Among current candidates you can consider Claude Opus 5.5 or GPT-6.1 Sol, verifying the result on the sample. The availability of a model for complex tasks is not proof of accuracy on specific contracts, offers or procedures. No proprietary benchmarks are hired here.

The budget must include extraction, pair preparation, access, preservation of reports and human review. A long document may take several steps; the project must hold the cover and references together. It also counts how much it costs to open and control sources: a report without useful links can slow the person down.

Cases to be included in the comparison sample

• Paragraph moved without changes in meaning.
• A negation or subject modified into a short sentence.
• Table with equal price but different unit or quantity.
• Attachment quoted in new copy but unavailable.
• Unreadable page and internal reference renumbered.

What causes a seemingly good confrontation to fail?

The first mistake is asking for "recapture differences" without identifying copies. The second is to accept page references that do not lead to the reported passage. The third is to verify only what the model has found. Measuring omissions requires a couple already controlled by a person.

Accesses must also be defined. Those who can read document A may not be authorized to copy B or attachments. The guide on AI knowledge base and permissions treats this boundary. The report may contain information as confidential as the sources and must be protected accordingly.

From file pair to usable review

To start, it carries two anonymized real copies, the attachments and a list of changes that the manager considers important. The pilot must produce a verifiable report, with visible limits and a version-related decision. L’Olivia Document Automation can connect this step to the larger process after the test.

With the service AI agents you can design reading, comparison and review on the sources of the enterprise. The first useful choice concerns the type of document and the person who can control it. From there, cover, cost and criteria are defined to decide whether to extend the flow.