Lease Intelligence at Scale

JLL manages lease data across 175 legal jurisdictions and more than 45 languages — the foundation for every strategic, financial and operational decision across its global real estate business. To modernize how that data is captured, governed and accessed, JLL partnered with DealSumm and Databricks. Built on the DealSumm platform and powered by Genie and AI/BI, the solution transforms complex lease documents into accurate, traceable and queryable records, giving business users governed, conversational access to portfolio-level insights at scale.

Bridging the gap between lease complexity and business accessibility

Every high-stakes decision across JLL’s five business lines starts with one thing: accurate lease data. It underpins pricing strategy, property management, occupier advisory and portfolio planning for clients around the world. But JLL handles leases across 175 legal jurisdictions in more than 45 languages, each with its own rules, formats and client requirements. The margin for error is zero. Inaccurate lease data leads to flawed advice, missed financial commitments and poor outcomes for clients.

For years, extracting that data meant relying on thousands of tenured subject-matter experts to manually review documents, each a specialist in their business line and region. Extracted information typically landed in spreadsheets or siloed files, disconnected from the downstream systems where it was needed. The process was thorough, but it could not scale.

“The data was always there, and the models were strong, but accessing that information in a meaningful way was still a challenge for our teams,” said Andrew Ray, Head of Transformation Portfolio Services, Business Lines at JLL. “Business users needed a simpler way to interact with lease data without relying on technical teams.”

Early investments in machine learning helped automate portions of the workflow, but classical ML had limits. It could handle structured fields, but it could not reason through legal nuance, cross-reference amendments scattered across hundreds of pages or manage the complexity of multi-document packages spanning addenda, exhibits and translations. Human experts were still left sorting through noise at precisely the moments when speed and accuracy mattered most. The gap was not in data availability. It was in the ability to interact with that data meaningfully, once abstraction was complete.

A purpose-built platform on Databricks

To close that gap, JLL partnered with DealSumm, whose platform was built specifically for lease abstraction. DealSumm brings the specialized architecture to translate JLL’s deep domain expertise into high-performance AI workflows. Databricks provides the intelligence and governance layer that enables the entire system to operate at a global scale.

DealSumm’s platform delivers extraction across more than 2,500 data points per lease, with full traceability: every extracted value is cited back to its precise source in the original document. Because lease agreements leave no room for error, a human reviewer remains in the loop throughout, validating outputs before data is finalized. The platform also supports multi-cycle abstraction. When an amendment takes effect, the system tracks the change and identifies only the specific data points that have been updated, eliminating the need for full re-abstraction. Data privacy and security are preserved by design.

For every session, the pipeline spins up a dedicated AI endpoint that is deleted after use and never reused, so no client data is retained, trained on or shared across sessions. Because each endpoint has access only to the data provided during that specific session, the system does not require the strict guardrails or advanced hallucination protections that broader, persistent models typically need.

“Genie enables users to interact with finalized lease data through a governed, conversational interface,” said Jonathan Bauman, CTO at DealSumm. “Instead of navigating reports or submitting requests, users can ask questions directly and receive answers grounded in approved lease content.”

At the center of the architecture is a Supervisor Agent-style orchestration layer that acts as a workflow conductor, dispatching specialized agents to handle summarization, clause interpretation, translation and CAM term analysis within a unified pipeline. Databricks AI Search surfaces relevant content across massive, unstructured lease collections regardless of how a clause is worded or where it appears in a document. Advanced reranking ensures the AI prioritizes the best supporting evidence when answers are ambiguous or spread across multiple documents, producing grounded and explainable outputs for reviewers.

Model Serving endpoints enable rapid deployment of improvements across the platform. At the same time, MLflow LLM judges provide systematic quality assurance, evaluating every output for legal and business relevance rather than just completeness. An elastic Retrieval-Augmented Generation system, purpose-built for legal document complexity, allows DealSumm to process entire multi-document packages, including amendments, addenda and exhibits spanning thousands of pages as a single coherent knowledge source.

Once human review is complete, finalized lease content is integrated into a governed knowledge base that Genie queries to ensure all responses are grounded in validated information. Unity Catalog enforces access control, lineage and data security across every interaction. Finalized data flows automatically into JLL’s downstream systems, including Yardi and MRI, through Delta Sharing, eliminating manual hand-offs and the risk of transcription errors.

Productivity, accuracy and insight at portfolio scale

The impact of this approach is visible across every stage of the workflow. AI accuracy across the lease field extraction now exceeds 92% before human review, up from 65% with prior approaches. Review time per lease dropped from 130 minutes to 85 minutes, team throughput doubled and cost per completed abstract decreased by 30%, all without increasing the overall cost structure.

“Now, instead of waiting on reports or digging through documents, our teams can ask a question and get a clear answer immediately,” said Andrew. “That speed changes how decisions get made across the business.”

But the more significant shift is in how lease data is used after abstraction. Through Genie, business users can query individual leases or entire portfolios in natural language, asking questions such as “Which leases in Germany renew in 2027 with CPI-linked rent escalations?” and receive precise answers grounded in approved, governed content.

Users can ask follow-up questions to go deeper on lease terms, legal conditions or cross-portfolio trends, without waiting on technical teams or working through static reports. Because BI and reporting tools are built on the same governed dataset, teams can also navigate the full scope of data collected per lease from multiple angles, viewing activity at the property, portfolio or jurisdiction level.

“What used to be a final deliverable is now something we can continuously learn from and build on,” Andrew added. “Genie turns that data into something the entire business can use.”

Reviewed leases are no longer static outputs. They are governed knowledge assets, queryable, auditable and accessible, with the platform on track to scale across every jurisdiction and language in which JLL operates. Coverage has expanded without adding cost or risk, and the foundation is in place for deeper self-service capabilities and portfolio-level intelligence as the platform continues to evolve. By combining agentic AI with Genie and AI/BI, JLL transformed lease abstraction from a document processing task into an insight-driven system that connects data directly to business outcomes.