Quick answer: Give AI both complete versions, require section-level source references, separate wording changes from actual obligation changes, and verify every material finding against the originals. The goal is not a polished summary. It is a change log that a human reviewer can audit.

Why a simple “summarize the differences” prompt is risky

Long documents often contain repeated language, moved sections, exceptions, tables, footnotes, and definitions that change the meaning of later clauses. A short summary can hide precisely the details that matter most. Treat comparison as a controlled review process, not a one-shot writing task.

1. Confirm the documents before comparing them

Record the title, owner, version date, approval state, and effective date for both files. Check that neither document is truncated and that appendices, tables, and referenced definitions are present. If one version is incomplete, stop and obtain the authoritative copy.

2. Build a section map

Ask AI to list the section hierarchy for each version. Match sections by purpose, not only by number, because a revision may move or rename content. Mark unmatched sections as added, removed, or requiring manual review.

3. Compare material changes in categories

• New or removed obligations.
• Changed owners, approvers, or escalation paths.
• New deadlines, thresholds, limits, or exceptions.
• Changed definitions that affect other sections.
• New required records, notices, training, or system steps.
• Wording changes that appear editorial but may alter meaning.

4. Require traceable evidence

For every finding, capture the old section, new section, short exact excerpts, a plain-language description, and a confidence label. If the model cannot locate both passages, mark the item Needs manual review instead of filling the gap.

5. Separate change from interpretation

“The deadline changed from 10 to 5 business days” is a document fact. “This creates legal exposure” is an interpretation. Keep them in separate columns so subject-matter experts can review conclusions without losing the underlying evidence.

6. Turn confirmed differences into an impact assessment

For each material change, identify potentially affected roles, systems, forms, training, customer communications, reporting, and dependencies. Use “potentially affected” until the responsible owner confirms the impact.

7. Create an owned action plan

Each action should contain one deliverable, one owner, a due date, dependencies, and a completion check. Avoid vague tasks such as “update everything.” Name the exact page, template, workflow, or training item.

8. Perform a human verification pass

Open both originals and verify all high-impact rows. Check definitions, exceptions, cross-references, dates, amounts, permissions, and responsibilities. Legal, regulatory, employment, privacy, safety, and contractual interpretations require an appropriate qualified reviewer.

Recommended change-log columns

• Change ID and category.
• Old section and excerpt.
• New section and excerpt.
• What changed.
• Possible impact.
• Confidence and open questions.
• Reviewer decision.
• Action, owner, due date, and verification.

Final takeaway

AI is valuable here because it reduces mechanical comparison work. Reliability comes from completeness checks, source references, explicit uncertainty, and accountable human review—not from how confident the generated summary sounds.