Structured extraction labels for an enterprise contract platform

An enterprise contract-lifecycle platform serving in-house legal teams needed labelled training data for structured extraction across 28 contract types – party identification, clause classification, key-date tagging, and renewal-condition flags – at a quality bar that justified replacing manual review.

Free-text variation in commercial agreements was breaking the platform's previous extraction approach, with field-level accuracy stalling around 91% on the messier contract families.

We assembled an annotation team led by paralegals with corporate-law backgrounds and built a custom labelling tool that highlighted ambiguous clauses for second-pass review. Schema design was iterated jointly with the platform's ML and legal-product leads, with edge-case rulings captured in a versioned playbook.

Every document went through automated pre-OCR, primary annotation, and a sampling-based QA layer that audited 10% of completed work blind.

Delivered 90K labelled contracts at 99.2% field-level accuracy on the validation set, lifting the platform's production extraction model from 91% to 97.6% on a customer-blind benchmark.

Their largest customer cut average contract-review time per document from 41 minutes to 9 – the cornerstone metric in their renewal pitch and a direct lever on a A$3.4M ARR retention.