Liorant deployed a permission-aware document-intelligence system that turns scanned and digital lease contracts into structured, searchable information with citations back to the source.
The anonymous client manages leasing contracts across shopping centers and property portfolios. Legal, finance, administration and asset-management teams need the same facts — expiration dates, renewal conditions, amounts, obligations and specific clauses — from documents that include historical scans and image-based PDFs. Liorant deployed the system to make that portfolio searchable without asking any team to trust an answer it cannot verify.
Architecture was sized for portfolios ranging from roughly 50 to 1,000+ contracts.
Legal, finance, administration and asset-management roles.
Clauses, dates, amounts, obligations, renewals and conditions.
These are scale and scope figures, not a validated time-savings or ROI claim.
Contract knowledge that only lives in someone's memory — or in a folder no one has time to reread — is a liability, not just an inefficiency.
Legal, financial, administrative and asset-management teams needed to identify expiration dates, renewal conditions, amounts, obligations, clauses and other contractual terms across a large and growing portfolio of leasing contracts. Some of those contracts were historical scans and image-based PDFs that standard search tools can't read at all.
The deployed system brings scanned and digital leases into the same searchable portfolio. Authorized teams can ask portfolio-wide questions, retrieve structured facts and follow each answer back to the relevant contract and clause.
Executive implication: a contract question can start with a direct, cited answer instead of a document-by-document search, while legal and business reviewers retain responsibility for the final interpretation.
Important facts were spread across long digital files and historical scans, and teams searched them one document at a time.
The portfolio supports permission-aware search, structured extraction and answers linked to the underlying source.
Claim boundary: this case documents a live deployment and its operational change. No audited time-savings or ROI figure is available for publication.
A search bar over a shared drive looks like a solution until someone actually needs an answer they can rely on.
Many documents needed OCR before they could be searched reliably at all.
Teams needed answers across many contracts at once, not a document-by-document hunt.
Dates, amounts and obligations needed to come out as structured data, not as a paragraph someone still has to re-read.
Different teams legitimately needed different access rights to the same portfolio.
Every answer needed a clear path back to the exact source document and clause.
Five separate failures, one root cause: retrieval was never designed as a system.
Liorant built and deployed a controlled document pipeline that combines extraction, normalization, metadata, retrieval and generation. The system enriches each contract with information such as property, shopping center, tenant, contract type, portfolio, effective dates and document category.
The access model remains independent of the AI response. Retrieval applies authorization before any information reaches the model, so the system never relies on a prompt to enforce document permissions.
OCR brings historical scans into the same workflow as digital PDFs.
Parsing and metadata turn documents into searchable contract information.
The system filters results by the requesting user's authorized scope.
Each answer links back to the contract and clause a reviewer can verify.
"Which shopping-center leases include a CPI-linked rent escalation clause, and when do they next renew?"
The system returns a direct answer with a citation back to the specific clause and contract — not a list of documents to go read manually. Example is illustrative; no real client contract text is shown.
The executive case stands on its own. Technical and security teams can expand the deployed implementation pattern below.
Vector-store infrastructure is selected according to deployment requirements — candidates include Qdrant, Pinecone, Azure AI Search or FAISS, depending on the client's hosting and compliance constraints, rather than one fixed choice for every engagement.
Architecture and repository examples have been simplified and anonymized to protect client confidentiality while showing the implementation pattern. This is a representative structure, not a claim that every production repository used these exact filenames.
contract-intelligence/ ├── app/ ............... api · chat · extraction ├── ingestion/ ......... pdf · ocr · normalization · pipelines ├── documents/ ......... schemas · metadata · parsers ├── knowledge/ ......... contract_types · clause_taxonomy · business_rules ├── retrieval/ ......... chunking · embeddings · vector_store · hybrid_search · reranking · filters ├── extraction/ ........ clauses · dates · amounts · obligations · renewals ├── security/ .......... roles · permissions · access_filters ├── generation/ ........ prompts · answers · citations ├── evals/ ............. retrieval · extraction · groundedness └── tests/
The client deployed the system for a commercial real estate contract workflow. During design and evaluation, Liorant assessed scenarios ranging from roughly 50 contracts to portfolios of 1,000+, a target user base of approximately 15–30 people and six structured extraction categories: clauses, dates, amounts, obligations, renewals and conditions. The scenario figures describe evaluated scale; they do not represent the client's published production volume or a verified ROI result.
Permissions, metadata and source traceability were treated as retrieval architecture from the start, not bolted on as prompt instructions after the fact. The AI model only ever receives information the requesting user is already authorized to retrieve — which means access control doesn't depend on the model behaving correctly, it's enforced before the model sees anything.
Enterprise RAG becomes substantially more useful when retrieval, permissions, metadata and source traceability are designed together.
Yes. Liorant deployed the contract-intelligence system for an anonymous commercial real estate organization. Client identifiers and contract contents remain confidential, and the published scope figures distinguish evaluated scale from production volume.
No. It surfaces the relevant clause, date or amount with a citation back to the source document so a person can verify it quickly — it doesn't make legal judgment calls on the organization's behalf.
Yes. OCR converts scanned and image-based PDFs into the same searchable, structured format as newer digital contracts, so historical portfolios aren't left out.
No. Permissions and metadata filters are applied before retrieval, so the AI model only surfaces what the requesting user is already authorized to see — legal, finance, administration and asset-management roles can have different access.
No. It's the upper end of the range the architecture was evaluated against (roughly 50 to 1,000+ contracts), describing scale and scope — not a validated deployment size or a time-savings claim.
In a focused 30-minute session, we map one recurring document workflow, identify where search, extraction or verification creates friction and assess whether governed AI contract intelligence is worth pursuing.
You leave with an initial feasibility view, the access-control questions to resolve and the evidence needed for a business case — no slides, no pitch.