This report presents performance benchmarks for AI-powered commercial lease abstraction based on Lextract's extraction pipeline. All figures reflect processing of standard US commercial lease formats (NNN, full service gross, modified gross) using AWS Textract OCR and Anthropic Claude Sonnet AI extraction.
Core Performance Benchmarks
Extraction speed: Under 3 minutes per lease
Lextract processes a standard commercial lease PDF from upload to structured output in 2 minutes 47 seconds on average. The pipeline has three stages:
- OCR processing (AWS Textract): 35–55 seconds depending on document length and scan quality
- AI extraction (Anthropic Claude): 80–120 seconds for 126-field extraction
- Output generation (JSON, Excel, Word, PDF): Under 10 seconds
For comparison, manual abstraction by a trained US-based paralegal averages 3.5 to 4.5 hours per lease. The AI pipeline is approximately 80x faster on standard commercial lease formats.
Field coverage: 126 structured fields per extraction
Lextract extracts 126 structured fields from each commercial lease, organized into 9 categories:
| Category | Field Count | Examples |
|---|---|---|
| Party Information | 12 | Landlord entity, tenant entity, guarantor |
| Financial Terms | 24 | Base rent, escalation, security deposit, TI allowance |
| Dates & Term | 14 | Commencement, expiration, renewal notice deadlines |
| Options | 10 | Renewal options, termination rights, expansion rights |
| CAM/Operating Expenses | 18 | CAM cap, base year, gross-up, audit rights |
| Permitted Use | 6 | Use clause, exclusivity, co-tenancy |
| Assignment & Subletting | 8 | Assignment consent standard, subletting rights |
| Casualty & Condemnation | 10 | Abatement rights, termination triggers |
| Miscellaneous Operational | 24 | Holdover rate, SNDA, estoppel obligations, surrender conditions |
For comparison, a typical paralegal manual abstract covers 40–60 fields. General AI tools (ChatGPT with a lease prompt) extract the fields you specifically request — typically 10–20 without a carefully engineered prompt.
Red flag detection: 20 automated checks
Lextract runs 20 automated red flag checks on every extraction. These checks identify high-risk provisions that tenants, lenders, and investors need to evaluate.
| Red Flag Category | Detection Rate | Notes |
|---|---|---|
| Uncapped CAM charges | 94% | Catches leases with no annual CAM increase cap |
| Missing tenant audit rights | 96% | Identifies leases with no CAM audit provision |
| Personal guarantee present | 99% | Near-perfect detection on explicit guarantees |
| Holdover rate exceeding 150% | 92% | Flags punitive holdover provisions |
| Management fee exceeding 5% | 95% | Identifies above-market management fee pass-throughs |
| One-sided termination rights | 91% | Flags landlord-only termination provisions |
False positive rate: 5–8% of flags triggered on provisions that are tenant-favorable or standard market terms. False negative rate: 4–8% on standard commercial lease formats for risk patterns that exist but are not flagged.
Accuracy Benchmarks by Document Type
Field-level accuracy is the percentage of extracted fields that match the ground truth value in the source document.
| Document Type | Accuracy Range | Key Variable |
|---|---|---|
| Typed NNN lease (standard) | 95–98% | Clean OCR, consistent field locations |
| Full service gross lease | 93–97% | Complex operating expense definitions |
| Modified gross lease | 92–96% | Variable structure across documents |
| Retail percentage rent lease | 90–95% | Breakpoint calculations, gross sales definitions |
| Ground lease | 85–93% | Complex cross-references, non-standard structure |
| Lease with 3+ amendments | 88–94% | Amendment hierarchy reconciliation |
| Low-resolution scanned PDF | 78–88% | OCR accuracy limits extraction |
| Handwritten annotations | 60–80% | Current OCR technology limitation |
AI vs. manual first-pass accuracy: Manual abstraction by trained US-based paralegals achieves 85–92% accuracy on a first pass before senior review. AI extraction on standard typed leases achieves 95–98% — higher than manual first-pass accuracy, though manual reviewed output achieves 97–99% after QA.
The practical implication: AI extraction is more accurate than manual on the first pass for standard commercial leases. For complex non-standard documents, human review remains advantageous.
Confidence Scoring Distribution
Every extracted field receives a per-field confidence score (0–100). This is the critical differentiator from manual abstraction and general AI tools, which provide no indication of extraction certainty.
Across standard commercial lease extractions:
- Score 90–100 (high confidence): Approximately 70–75% of fields. These fields match ground truth at 97–99% accuracy.
- Score 70–89 (moderate confidence): Approximately 18–22% of fields. Match ground truth at 88–94% accuracy.
- Score below 70 (review recommended): Approximately 5–10% of fields. Match ground truth at 72–85% accuracy.
Practical validation workflow: A reviewer focusing exclusively on fields scoring below 85 needs to verify approximately 8–15 fields in a 126-field extraction. At 3–5 minutes per field verification, complete validation of uncertain fields takes 25–60 minutes — compared to 3.5–4.5 hours for full manual re-abstraction.
Cost Comparison
At scale, the cost difference between AI and manual abstraction is substantial.
| Volume | Lextract (AI) | Manual (US paralegal) | Offshore Managed Service |
|---|---|---|---|
| 10 leases | $200 | $1,500–$3,000 | $400–$750 |
| 50 leases | $850 (10-pack rate) | $7,500–$15,000 | $1,500–$3,750 |
| 100 leases | $1,700 | $15,000–$30,000 | $3,000–$7,500 |
| 500 leases | $8,500 | $75,000–$150,000 | $15,000–$37,500 |
AI abstraction at scale is 10–20x cheaper than US-based manual abstraction and 2–5x cheaper than offshore managed services.
Including internal reviewer time (20–30 minutes at $75/hour per lease for confidence-flagged field validation):
| Volume | AI + Review Total | Manual Service |
|---|---|---|
| 50 leases | $850 + $1,250 = $2,100 | $7,500–$15,000 |
| 100 leases | $1,700 + $2,500 = $4,200 | $15,000–$30,000 |
The cost advantage holds even after accounting for internal review labor.
Processing Capacity
Lextract's extraction pipeline supports concurrent processing with no queuing delay for standard volumes:
- Single lease: Under 3 minutes from upload to download
- 10 leases: All processed concurrently, complete within 4–5 minutes
- 50 leases: Complete within 12–18 minutes
- 100+ leases: Complete within 25–40 minutes
For a 100-lease due diligence project, AI extraction delivers all 100 structured abstracts in under 40 minutes. Manual abstraction of the same 100 leases requires 350–450 hours of paralegal time — approximately 9–12 weeks for a single abstractor.
Key Benchmarks at a Glance
| Metric | Lextract (AI) | Manual Paralegal | Offshore Service | ChatGPT |
|---|---|---|---|---|
| Processing time | Under 3 minutes | 3.5–4.5 hours | 24–72 hours | 2–5 minutes |
| Field coverage | 126 fields | 40–60 fields | Custom (per client) | 10–20 (as prompted) |
| Accuracy (standard leases) | 95–98% | 85–92% (first pass) | 97–99% (reviewed) | Unverified |
| Confidence scores | Yes (per field) | No | No | No |
| Red flag detection | 20 checks (automated) | Manual | Manual | None |
| Cost per lease | $20 | $150–$300 | $30–$75 | Free |
| Data retention | Zero | Varies | Varies | Trains on inputs* |
*ChatGPT consumer plans may train on inputs. Enterprise plans with data processing agreements are an exception.
Methodology Note
Accuracy benchmarks are based on internal testing against a reference set of commercial lease documents with known ground truth values. Document types, scan quality, and lease complexity varied across the test set. Figures represent ranges across the full test set, not best-case performance. Actual accuracy on any specific lease may vary based on document characteristics.
These benchmarks are for standard US commercial real estate leases in English. Performance on non-English documents, foreign lease formats, or highly non-standard structures is not covered by these benchmarks.