Context
LexCentrix is a legal platform with three portals: a request portal for the business, a workspace for the legal team, and an admin portal where the rules live. Each contract in the workspace opens in one dialog with tabs for details, parties, versions, history, an AI summary and the AI clause review. Before this feature, a lawyer checked every incoming contract against the company's approved wording by reading it line by line.
The challenge
- Approved wording differs by contract type (NDA, employment contract, lease agreement) and by country and office, so one checklist does not fit all.
- Some clauses are mandatory, some are optional, and some are restricted: wording the company does not want to accept at all.
- Comparing a contract with the approved text by eye is slow and easy to get wrong, especially when a clause is reworded but still looks similar.
- A legal team will not trust an AI verdict it cannot check against the document.
My role
I reviewed the feature end to end as a product person and tester: how a lawyer reaches it, what the review shows, where it fails, and how it behaves for each contract type and organisation. I documented it in the platform user manual, wrote the test cases and bug reports, and built the interactive prototype linked above to walk stakeholders through the flow without touching the live system.
The review never just says “this clause is wrong”. It puts the approved wording and the contract wording side by side, says in one sentence how they differ, and warns when the quoted text could not be found in the document. The AI points; the lawyer decides.
Approach
- A clause library that the legal team owns. In the admin portal, every clause has a kind (standard or restricted), a contract type, an organisation (country and department), a mandatory or optional flag and a version. Reviews compare against the library for the contract's type and office.
- One tab, three actions. The AI Clause Review tab holds a contract type selector, a Detect with AI button for contracts nobody has labelled, and a Review button that becomes Re-review once a result exists.
- Always labelled as AI. A banner on the tab reads “AI-generated, verify against source; not legal advice.”
- Results grouped by what the lawyer must do. Findings sit under “Standard clause comparison” and then “Deviations from the approved wording”, each with a count, so the first thing a reader sees is how much needs attention.
- Side-by-side comparison. Each finding shows the approved wording and the contract wording in two columns, with the clause name, version and a one-line explanation of the difference.
- Traceable runs. The result header records what was compared against: contract type, organisation, number of clauses, clause version and the time of the review.
Try it
The prototype below uses sample contract text. Press Detect with AI or Review, open the groups, and mark findings as checked.
Open the prototype in its own tab →
Outcomes
Key decisions & learnings
- Show the evidence next to the claim. A finding without the approved text beside it forces the lawyer to go and look it up. Two columns removed that step.
- Warn when the AI quotes something it cannot find. The tab tells the reader to check the document before relying on a quote. That one line is what makes the rest of the output safe to use.
- Make the library the single source of truth. Putting kind, mandatory flag, organisation and version on each clause keeps the AI to rules the company has actually approved.
- Label it as AI, every time. The banner stays on the tab in every state, empty, running or finished.