Comparing documents is rarely about finding every changed character. It is about finding the changes that matter: a new obligation, a different date, a removed caveat, a revised number, or an assumption that no longer holds. AI can make the first pass much faster when you give it a comparison framework and insist on evidence for every conclusion.
Decide what “different” means before uploading
First name the two sources clearly: for example, “2025 supplier agreement” and “2026 supplier agreement”, or “forecast version 3” and “forecast version 4”. Then define the comparison criteria. A contract review may focus on liability, data handling, termination, pricing, and renewal. A policy review may focus on obligations, owners, deadlines, and exceptions. A spreadsheet review may focus on inputs, formulas, totals, and periods.
Without that framing, a comparison can drown you in harmless wording changes and overlook a material shift in one sentence.
Ask for a structured difference table
Start with a table instead of a prose summary. It creates a shared shape for the first pass and makes the review easier for a colleague.
| Field | What to request |
|---|---|
| Topic or clause | The exact item being compared. |
| Document A | The relevant wording, value, or section. |
| Document B | The relevant wording, value, or section. |
| Change | Added, removed, narrowed, broadened, or changed. |
| Significance | Why the change may matter, stated cautiously. |
| Evidence | File name and page, section, sheet, or cell reference. |
For a contract, you might ask: “Compare only commercial and operational changes. Quote the relevant wording and label legal interpretation as a question for counsel rather than a conclusion.” For a spreadsheet, you might ask for the cell range, worksheet, units, and periods behind each changed figure.
Separate detection from interpretation
The first job is to detect the difference. The second is to explain why it could matter. Keep those jobs separate.
For example, “the notice period changed from 30 to 60 days” is a factual difference. “This increases operational risk” is an interpretation that depends on the organisation, process, and contract context. Ask the model to label interpretations, assumptions, and questions instead of presenting them as facts in the source.
This distinction makes the output more honest and more useful. A reviewer can accept the factual comparison while deciding whether the implication applies to their situation.
Use targeted follow-ups after the first pass
Once you have a table, ask focused questions:
- Which changes are likely material to the stated decision?
- Which sections contain different terms that use similar wording?
- Which values changed without an obvious explanatory note?
- What information is missing to decide whether a difference is important?
- Which three source passages should a human review first?
For versions with many changes, work section by section. That is usually clearer and more reliable than asking for an entire legal agreement, financial model, or technical specification to be resolved in one answer.
Keep source references with the handoff
When the comparison will be shared, preserve the links or precise references to the original material. Ask for a final memo with three parts: material differences, items requiring owner review, and questions that remain open. Include the date and version of each source so the comparison does not become misleading after another revision arrives.
Try this prompt:
Compare the two attached documents using the criteria below. Return a table with the exact relevant text from each file, the material difference, and a source reference for every row. Separate factual changes from possible implications. Do not provide legal, financial, or policy conclusions that are not directly supported by the documents; list those as review questions.
AI is a strong first-pass comparison assistant. It is not a replacement for the domain expert who must interpret a material change, confirm a calculation, or accept the final decision. Used with a clear scope and a reviewable evidence trail, it can give that expert a much better place to start.
