12,000 rent rolls per year arriving in 4 different channels, each in a different format. One extraction-to-insight pipeline — from first scoping call to executive-facing demo in under 3 weeks.
Company:
Global Commercial Real Estate Leader
My Role:
AI Product Manager, Enterprise Solutions (Unframe AI)
Year:
2026
Techstack
Document Extraction Agents · Inbox Processing · NER Pipelines · Validation Agents · CRE Intelligence Layer · Yardi · MRI Software · Argus Enterprise · Box · Google Drive · Outlook · DocuSign · Salesforce · Snowflake · Looker
A global commercial real estate leader was receiving 12,000 rent roll documents per year across 4 inbound channels: Outlook attachments, Box, Google Drive, and DocuSign-executed docs. Each document was a different format — brokers don't standardize their spreadsheets, property managers don't follow templates, and Excel variations across 50+ brokers meant every document required manual analyst attention.
The extraction process was entirely manual: analysts opened each document, copied data into Argus Enterprise or Yardi, and verified field by field. At 12,000 documents per year, this was a material analyst capacity constraint — and every delay in extraction was a delay in the portfolio intelligence that senior leadership needed for capital allocation decisions.
STEP 1 — Inbox Processing and Triage
The pipeline started at ingestion — before extraction, documents needed to be received, deduplicated, and routed. Built an inbox processing agent that monitored all 4 inbound channels (Outlook, Box, Google Drive, DocuSign), classified each incoming document by type, and tagged it with property and broker metadata derived from the email subject, sender, and filename patterns. No manual triage, no missed attachments buried in inbox threads.
STEP 2 — Extraction Agent for Non-Standard Documents
Standard OCR fails on broker spreadsheets because the schema changes document to document. Built an extraction agent using NER pipelines that learned field locations dynamically: it identified the rent roll fields by semantic meaning (tenant name, lease expiration, base rent, square footage) rather than fixed cell coordinates. Audited rent roll formats across property types and geographies to establish the real range of document variation the agent needed to handle.
STEP 3 — Validation Layer
Extracted data is only useful if it's accurate. Built a validation layer that ran every extracted record against analyst-defined quality criteria: are lease dates internally consistent, does base rent per square foot fall within the range for this property type and market, are tenant names normalized against the master tenant list? Fields that didn't pass validation were flagged for analyst review with the specific failure reason, not a generic error.
STEP 4 — Commercial Intelligence Layer
With clean, validated rent roll data flowing automatically into Argus Enterprise, Salesforce, and Looker, the portfolio became queryable in real time. Senior leadership could see current occupancy, expiring leases, and renewal probability across the entire portfolio without waiting for a weekly analyst report. The same data fed the capital allocation models that previously ran on stale, manually-updated spreadsheets.
RESULTS
12,000 rent rolls per year — full inbox-to-insight pipeline automated
4 inbound channels unified: Outlook, Box, Google Drive, DocuSign-executed docs
Manual extraction eliminated — agent-validated against analyst-defined criteria before every downstream use
Time from first scoping call to executive-facing working demo: under 3 weeks
Downstream systems fed directly: Argus Enterprise, Salesforce, Looker — no manual re-keying
