AI System for Explaining Changes Between Contract Versions

The problem

In day-to-day operations, formatting differences can obscure substantive changes. When the old and new contracts, clauses, definitions and version metadata are spread across different screens, documents or team records, the issue may not be noticed in time.

This fragmented data delays preparation of the clause-level difference report and forces staff to search for supporting records again by hand. Before deciding, the contract lawyer must complete missing fields and verify each inconsistency individually.

A common exception is that a change in numbering or formatting may not reflect a change in content. If this context is not checked, an incomplete or incorrect automation could treat a minor difference as a risk or miss a substantive change, so the output cannot be acted on directly.

What we set out to improve

• Share of changes the lawyer verifies as substantive differences
• Share of outputs corrected or rejected by the contract lawyer
• Time between the source record and the review output

How the system works

1. 01 Define the review scope and the authorised contract lawyer role
2. 02 Collect the inputs: old and new contracts, clauses, definitions and version metadata
3. 03 Check the source identifier, date, version and data freshness
4. 04 Divide the files into clauses
5. 05 Match differences in text, definitions and references
6. 06 Check the exception: a change in numbering or formatting may not reflect a change in content
7. 07 Place low-confidence or conflicting records in a separate human review queue
8. 08 Write the source identifier, rule, output and timestamp to the audit trail
9. 09 Have the contract lawyer approve, correct, defer or reject the output

Methods we used

• Structural document comparison and definition-reference checks
• Source, format, version and required-field checks for the old and new contracts, clauses, definitions and version metadata
• A missing-data alert instead of an estimate when inputs are absent or conflicting
• Separate handling of exceptions and the main rule
• Joint retention of the source identifier, rule result, clause-level difference report and human decision

Where people stay involved

Data and security

Access to the old and new contracts, clauses, definitions and version metadata is limited to the minimum permissions needed for the task. Identifiers and sensitive fields are masked where possible, and only the necessary section is sent to the model rather than the full raw file. Access, output and the contract lawyer's decision are logged; production use does not begin until the organisation approves the retention period and data location.

Who this suits

Frequently asked questions