Hong Kong Property Developer: Document AI Cuts Lease Extraction Time by 82%

A Hong Kong-listed residential and commercial property developer with a portfolio of 34 active projects across Hong Kong and the Greater Bay Area was processing approximately 9,400 lease-related documents per year — tenancy agreements, renewal negotiations, fit-out approval requests, and end-of-tenancy inspections — through a 22-person leasing administration team.

The team spent an estimated 55% of its time on document extraction and manual data entry: pulling key terms (rent review dates, break clauses, permitted use definitions, landlord-tenant obligations) from lease agreements into the company's property management system. The remainder was spent on tenant communications, dispute resolution, and regulatory compliance work.

Two problems compounded the inefficiency: first, the company's lease portfolio used both English and Traditional Chinese documents, requiring bilingual staff on every extraction task. Second, the documents ranged from 8-page standard residential leases to 180-page commercial tenancy agreements with multiple schedules — creating significant variability in extraction complexity that the team had to manage manually.

The General Manager of Property Management engaged AIMenta after a lease renewal was delayed by six weeks due to a missed break clause date that had not been captured during document intake.

The engagement began with a two-week document audit across the 34-project portfolio. The audit produced three findings that shaped the solution design:

1. 71% of the 22 most critical lease terms (rent review dates, break clauses, option periods, fit-out obligations, permitted use) appeared in predictable structural positions across document types — suggesting high-confidence extraction was achievable with a structured approach
2. Traditional Chinese documents were handled by a different sub-team and represented 38% of the total document volume, but used consistent commercial real estate terminology (unlike consumer-facing Chinese where informal phrasing varies)
3. The 180-page commercial leases required extraction from multiple interdependent schedules — a challenge for chunk-based approaches, requiring full-document context

The solution design: a document intelligence pipeline combining OCR (for PDF/scanned documents), a layout-aware extraction model for structural parsing, and a fine-tuned extraction model for 22 critical lease terms — supporting both English and Traditional Chinese input.

A key design decision: rather than deploying fully automated extraction, the system presented extracted values with confidence scores to the leasing administration team, who confirmed or corrected each extraction before the data entered the property management system. Corrections fed a weekly retraining loop that improved the model on the company's specific lease portfolio.

The production system deployed as an internal web portal integrated into the company's existing property management platform (Yardi Voyager). Staff upload documents through the portal; the pipeline processes them and returns a structured extraction review form pre-populated with the extracted values.

The interface presents each extracted field with a confidence indicator (high/medium/low) and the source text highlighted in the original document for verification. High-confidence extractions (>88%) are presented for one-click confirmation; medium and low confidence require staff to review the source text before confirming.

The Traditional Chinese pipeline uses the same architecture with a Chinese-specialised embedding model (BGE-M3) for retrieval and a Chinese-capable extraction model fine-tuned on the company's specific terminology — including property-specific terms that standard Chinese NLP models do not handle correctly (e.g., the specific Chinese legal terminology for break clauses and fit-out obligations differs between HK and Mainland China conventions).

The system also generates automated lease summary sheets — a one-page document in the company's standard template summarising the 22 key terms — which are stored in the company's SharePoint for use by leasing managers, lawyers, and executive reviewers. These were previously produced manually, taking 45–90 minutes per lease.

Measured at six months post-deployment:

• Document extraction time per lease: 45–90 minutes (manual) → 8–12 minutes (review of AI extraction) — 82% time reduction
• Extraction accuracy on critical terms: 91.3% on first pass (measured against manual re-extraction of a 10% random sample)
• Traditional Chinese extraction accuracy: 88.7% (slightly lower due to smaller fine-tuning dataset; improving with each retraining cycle)
• Lease summary sheet production time: 45–90 minutes → automated within 3 minutes of extraction confirmation
• Leasing administration staff time on extraction tasks: 55% → 22% of total working hours
• Missed critical dates (break clauses, rent reviews) in the 6 months post-deployment: 0 (vs 3 in the same period prior year, one of which caused the six-week renewal delay)
• Staff satisfaction (post-deployment survey): 4.1/5 (was not measured pre-deployment; benchmark from comparable projects: 3.2/5)

The extraction model's performance has improved in each monthly retraining cycle. At the six-month mark, the high-confidence threshold (>88%) applied to 79% of fields on English documents and 71% on Traditional Chinese documents, reducing staff review burden while maintaining the human-in-the-loop verification step.