AI Lease Review: How It Works, What It Catches, and When to Use It

AI lease review uses machine learning and OCR to do what paralegals have done manually for decades: read a commercial lease and extract the material terms. The difference is 3 minutes instead of 3 hours, and a structured output instead of a filled-in template.

Here is how it works, what it reliably catches, and where human judgment is still required.

How AI Lease Review Works

The process has three stages for any commercial lease PDF:

Stage 1: OCR extraction. The PDF is scanned using optical character recognition to convert the document into machine-readable text. Purpose-built tools use layout-aware OCR (like AWS Textract) that preserves the spatial structure of tables, columns, and numbered clauses — not just raw text.

Stage 2: AI field extraction. A large language model reads the extracted text and maps it to a structured field schema. For each field — base rent, lease commencement date, CAM cap percentage, renewal option notice period — the model identifies the relevant lease language and extracts the value.

Stage 3: Confidence scoring and red flag detection. Each extracted field receives a confidence score indicating how certain the model is about the extraction. Separately, a rule-based engine checks the extracted fields against a set of red flag patterns and flags provisions that meet the risk criteria.

The output is a structured data set: 126 named fields (in the case of Lextract), each with a value, a confidence score, and any applicable red flag designation. Review time focuses on the 8–15 fields below the confidence threshold, not the full document.

What AI Lease Review Reliably Catches

AI performs consistently well on fields with clear, unambiguous values in standard lease formats:

Numeric fields:

• Base rent (annual, monthly, per SF)
• Security deposit
• Tenant improvement allowance
• Escalation rate (fixed percentage)
• CAM cap percentage
• Management fee percentage
• Gross-up percentage

Date fields:

• Lease commencement and expiration dates
• Renewal option notice deadline
• Rent commencement date
• Free rent period

Binary/categorical fields:

• Lease type (NNN, gross, modified gross)
• Personal guarantee existence
• Audit rights existence
• Assignment allowed with or without consent
• Renewal option existence

Party and premises:

• Landlord and tenant entity names
• Property address
• Rentable and usable square footage
• Pro rata share

On standard commercial leases (typed, well-formatted NNN, gross, and modified gross leases), accuracy on these fields runs 95–98%.

What AI Lease Review Misses or Flags Uncertain

Complex defined terms. Commercial leases often define "Operating Expenses," "CAM Charges," or "Gross Revenues" with multiple exclusions, carve-outs, and cross-references spread across different sections. AI extracts the primary definition but may miss exclusion language on page 48 that qualifies the provision on page 12.

Amendment hierarchies. When a base lease has multiple amendments — some superseding earlier amendments — the AI must understand which version of a provision controls. Current tools vary in how well they handle this. For heavily amended leases, human review of the amendment structure is important.

Ambiguous escalation language. CPI-linked escalations with complex calculation methodology, base index definitions, and cap/floor provisions require interpretation that AI sometimes misclassifies. The value gets extracted, but the formula may be partially wrong.

Percentage rent calculations. Retail leases with percentage rent tied to gross sales require extracting the breakpoint, rate, and definition of "gross sales" — often with lengthy carve-outs across multiple pages.

Non-standard provisions. Provisions with no precedent in the training data — unusual operating rights, complex co-investment structures, bespoke development obligations — may receive low confidence scores or be missed entirely.

For all of these, confidence scoring is the mechanism: the model assigns a low confidence score when uncertain, which directs human review to the specific fields that need it.

AI Lease Review vs. Manual Review: The Comparison

AI review does not replace manual review for complex or high-stakes leases. It replaces the first-pass extraction step, letting human review focus on the 8–15 fields that are uncertain rather than 126 fields that are routine.

AI Lease Review vs. ChatGPT Lease Review

The most common question from CRE professionals evaluating AI tools is: "Can I just use ChatGPT?"

ChatGPT can read lease language pasted into a prompt and produce a narrative summary. It is useful for explaining specific clauses or getting a quick overview.

What ChatGPT does not provide:

• Structured output: No fixed schema means no two extractions look alike. You cannot import a ChatGPT response into Yardi.
• Confidence scoring: No indication of which values are uncertain.
• Red flag detection: No automated check against risk patterns.
• Consistent field coverage: You get what you ask for; you do not get what you forgot to ask for.
• Context window limits: Long leases get truncated. You may miss provisions on page 65.
• Audit trail: No link between extracted value and source language.

Purpose-built tools like Lextract are designed for professional CRE workflows. ChatGPT is a general-purpose assistant. For occasional quick reads on non-critical provisions, ChatGPT is fine. For due diligence, rent roll verification, or portfolio review, a purpose-built tool is required.

See: Why ChatGPT is not enough for lease review.

When to Use AI Lease Review

Use AI as the primary first pass for:

• Standard commercial leases (NNN, gross, modified gross, shorter ground leases)
• Portfolio-level due diligence where volume requires speed
• Rent roll verification against a seller's representation
• CAM audit preparation (extract the CAM provisions, then compare to statements)
• Property management system onboarding (extract → export → import)

Add human review on top of AI for:

• Heavily amended leases (5+ amendments)
• Ground leases with complex rent structures
• Any lease where a specific provision will be negotiated or litigated
• Leases with non-standard operating rights or development obligations
• Any extraction with multiple low-confidence fields

Use human-only review when:

• The lease is in active litigation
• The extraction will be presented to a court, regulatory body, or lender with independent auditors
• The document contains primarily handwritten annotations or non-English provisions

For most commercial real estate portfolios, AI handles 90%+ of leases as the primary extraction tool. The 10% requiring deeper human review are identifiable by the confidence scores — not by re-reading everything.

How to Get Started

Lextract processes any commercial lease PDF in under 3 minutes for $20. Upload the document, receive 126 structured fields with per-field confidence scores and red flag detection, export to Excel, Word, PDF, or JSON.

No subscription. No implementation. No demo required.

Extract your first lease →