Compare document versions with a native difference check, reconcile changed figures to exact source locations and use AI only to explain verified changes.
Direct answer: Preserve both versions, run the document application's own comparison where available, then check every relevant numeric field against the original and revised source locations. Use AI to organise or explain those verified differences, not as the only detector of changed figures. If a number affects cost, timing or an obligation, approve it individually before relying on the summary; no prompt guarantees that a change will be found.
“Summarise the changes” invites prioritisation. That can be helpful for understanding a revision, but it is a poor substitute for accounting for every changed price, quantity or unit. A summary may accurately describe the general direction while omitting the one figure that changes your decision.
I recommend deterministic document comparison first and AI explanation second. The exception is a format with no useful native comparison, where you must construct a side-by-side record manually before asking AI to interpret it.
Applies to: two identified versions of the same document. The specific comparison controls described below are documented for Word desktop on Windows, not a claim of identical behaviour in web or Mac versions.
Use the revision reconciliation pass
The revision reconciliation pass is an editorial method for making each consequential difference traceable. It separates version identity, change detection and interpretation so that a plausible explanation cannot conceal missing evidence.
Create a record with the field label, original value, revised value, units, source location and verification status. Use the document's section and table labels as well as page numbers, since pagination can change between versions.
For example, “Table B, delivery allowance” is more stable than “page four”. Record whether £45 became £55 and whether the charge remains per delivery. A change from a monthly allowance to an annual one can matter even if the printed number is unchanged.
This applies the focused knowledge-work workflow to one bounded risk: ensuring a generated explanation remains attached to the precise evidence it describes.
Establish which versions you are comparing
Save copies with clear identifiers before running any comparison. Confirm which is the agreed baseline and which is the proposed revision. A filename containing “final” is not evidence that it is the authoritative copy.
Check dates, document owners and any existing tracked changes. If one file includes proposed edits and the other includes accepted edits, decide what state you intend to compare. Do not silently accept outstanding changes in the source document to simplify the task.
Keep confidential documents within approved software and storage. Before uploading to an AI service, verify permission and the exact account's processing terms. For a demonstration or prompt trial, use invented extracts rather than a live contract or customer price schedule.
If the two documents cover different products, periods or scopes, stop calling the exercise a revision comparison. First align the meaning of their fields. Otherwise differences may represent incomparable offerings rather than edits to the same agreement.
Run the native comparison without overwriting evidence
In supported Word desktop versions on Windows, Microsoft's legal blackline comparison can show differences in a third document. Open Review > Compare, identify the original and revised documents, and choose the comparison options. Keep results in a new document: Microsoft warns that selecting either source as the output location modifies it. Read any tracked-change prompt carefully before proceeding. Microsoft's document comparison instructions.
This tool identifies document differences; it does not decide whether a changed figure is commercially acceptable or mathematically consistent. Inspect tables, captions and notes that affect the quantities you care about.
If you have PDFs or scanned pages, establish whether their text and table layout are reliably extractable. Compare the visible pages as well as extracted text. A copied column without its heading can turn a valid number into misleading evidence.
For a small schedule, manually checking every relevant numeric field may be faster than configuring another service. Use the simplest method that gives you an auditable record. Do not skip the check because the AI's summary looks complete.
Give AI a bounded interpretation task
Once you have a candidate change record, ask for a table that preserves the original and revised values exactly, separates direct observations from implications and marks any unresolved mapping. Do not ask it to decide which changes are “important” before you have accounted for the numerical fields.
A useful plain-language instruction is: “Explain the effect of these supplied changes. Keep each original and revised value with its unit and source label. Do not infer missing figures. Mark an implication as uncertain when the supplied material does not establish it.” This is an editorial prompt, not a guarantee of compliance.
Check the output against both documents. Read numbers digit by digit where necessary, including signs, decimal places, percentage marks and currency. Distinguish a price increasing by five pounds from increasing by five per cent.
If the AI merges several changes into a single sentence, keep the individual source rows available. A concise explanation should sit above the evidence, not replace it. When a row is omitted or its unit changes, correct the record before circulating the summary.
Reconcile eleven changes across four tables
Consider two fictional event-supply schedules with four tables. Your source comparison identifies three changed numeric fields in Table A, four in B, two in C and two in D. The expected total is 3 + 4 + 2 + 2 = 11 changed fields.
Suppose an illustrative AI summary accounts for nine. That leaves 11 minus 9 = 2 unaccounted-for changes. This is not an AI accuracy benchmark; it is a completeness check against your own known revision record. The two omissions remain open regardless of how convincing the nine explanations sound.
One changed line might increase quantity from 12 to 15 and unit price from £8 to £9. The original total is 12 × £8 = £96. The revised total is 15 × £9 = £135, a difference of £135 minus £96 = £39.
An explanation that mentions only the £1 unit-price increase misses the quantity change. Applying £1 to the original 12 units gives £12, understating the total change by £39 minus £12 = £27. Both fields must stay attached to the line item.
Now inspect any unchanged totals too. If the revised document still prints £96, you have found a consistency problem, not merely a wording change. Flag it to the document owner rather than deciding silently that the AI-calculated £135 must be the intended contractual amount.
All values here are invented arithmetic examples. For an actual quotation, verify whether taxes, minimum charges, rounding or separate fees affect the total before drawing a financial conclusion.
Separate editorial understanding from approval
A completed change record can tell you what differs and where. It does not automatically authorise accepting the revision. Keep an explicit status such as checked, unresolved or approved by the responsible person.
For each unresolved item, state what is missing: an unclear unit, a shifted table row, an inconsistent total or an unidentified baseline. This is more useful than a broad warning to double-check everything.
Do not let the AI choose which document governs a disagreement. Ask the document owner for clarification and retain their response with the comparison. Where the interpretation concerns legal obligations, use qualified advice for the relevant jurisdiction and contract.
Finish the comparison in a controlled session
- Spend ten minutes confirming version identity and saving untouched copies.
- Run the native comparison or create a manual side-by-side record. List the numerical fields relevant to your decision before asking for a summary.
- Use one focused review session to reconcile each changed field, check affected totals and inspect the AI explanation against the sources.
- Send unresolved items to the responsible owner with precise locations. Approve or reject the revision only after those answers arrive.
Stop if the baseline cannot be established or if a consequential value remains ambiguous. For a deadline, provide the verified subset and the outstanding questions instead of describing the entire document as checked.
Related guides
Frequently asked questions
Can I ask AI to compare screenshots of the two pages?
You can use suitable approved images as an aid, but do not assume every number or table relationship will be read correctly. Preserve full headings, units and page context, and inspect the source images yourself for consequential fields. If the images are unclear, obtain a better source rather than asking the model to guess. A screenshot comparison can help identify candidate differences, but its output still needs reconciliation against the original values. For a small number of pages, a careful side-by-side visual check may be simpler than processing repeated corrected uploads.
What if the revised document changes the order of rows?
Match rows by their meaning or stable identifier, not simply by position. A moved item is not necessarily a changed price, and two similar labels may refer to different scopes. Record the label, unit and relevant context before pairing values. If you cannot establish a reliable match, mark it unresolved and ask the owner. Do not let the assistant merge near-duplicate items solely because their names look alike. Once the mapping is confirmed, compare the quantities and totals within each matched item and separately note genuine additions or removals.
Does a matching total prove that no important number changed?
No. Different changes can cancel out in the total while altering quantities, rates or responsibilities. A schedule could keep the same overall amount while moving cost between items or reducing what is supplied. Inspect the fields that affect the decision, not only the final sum. Use totals as a consistency check after reviewing the individual differences. If the commercial question is whether the same service remains included, compare scope and units alongside price. A matching total is useful evidence about arithmetic, but it is not evidence of unchanged meaning.
Should I accept tracked changes before comparing?
Only after deciding which document state you intend to compare and preserving untouched sources. Pending tracked changes can make the baseline ambiguous, so follow the application's documented workflow and read its prompts carefully. Do not accept edits in a live source merely to get a cleaner-looking output. If the owner intends a comparison between two proposed states, establish that explicitly and work with copies. Keep a note of the choice so another reviewer can reproduce the comparison. When you are unsure which edits were actually approved, ask the owner before interpreting differences.
Is it enough to check only the numbers mentioned by the AI?
No, because that approach cannot reveal a number the AI omitted. Build the review scope from the original documents and your decision requirements, then see whether the generated output accounts for it. Check the relevant tables, notes and numerical conditions even when the summary says there are no significant changes. If the scope is too large for one session, divide it into identified sections and track completion. Do not report a complete review when only selected rows were inspected. The evidence for completeness must exist independently of the explanation being checked.
How should I share the finished comparison with a colleague?
Provide the concise explanation together with the change record, source version identifiers and any unresolved questions. Make clear whether each consequential item has been checked or approved, since those are different states. Share documents only through the appropriate access-controlled process and avoid attaching confidential originals unnecessarily. A colleague should be able to locate a changed value without recreating the AI conversation. If you later receive a replacement revision, label it as a new comparison rather than silently overwriting the existing record and leaving earlier conclusions attached to different evidence.
Sources and verification
- Microsoft: compare document differences using legal blackline, checked 11 September 2026 for supported Windows desktop controls, new-document output and source-modification warnings. No hands-on comparison was performed for this article.
- The supplied parent was read in local publication files after public retrieval failed. Internal paths are retained without claiming independent confirmation of live publication. The schedules and reconciliation example are fictional.
This article is practical guidance. Apply it in proportion to your tools, evidence, risks, and responsibilities.
