Workspace Document Management
Zillow Workspace
document management
A full document system for real estate transactions — AI-assisted intake, a compliant signing workflow, and a centralized documents hub, built natively into Zillow Workspace.
over two years
real estate transaction
in H1 2025 alone
adopted from zDocs
Problem Statement
Agents lived in documents. Zillow had no native way to manage them.
A real estate transaction generates 30+ documents — listing agreements, purchase and sale agreements, disclosures, addendums. Agents and transaction coordinators juggled them across Dotloop, email, and external e-sign tools. None integrated into Zillow. Without a native document system, Zillow couldn't be the system of record for transactions — and every workaround an agent built was time not spent serving clients.
"Help movers transact faster on their journey to their dream home by enabling efficient, secure and seamless digital document experiences throughout the home purchase, financing and leasing process."
Issues for Agents & Transaction Coordinators
Three problems that made documents painful
Anchored every design decision below to a concrete agent or TC need rather than a hypothetical one.
Project Scope
Five interlocking workstreams over two years
This wasn't a single project. It was a suite of connected design challenges that each addressed a different layer of the same underlying problem: how do you make document management a native, integrated capability rather than something agents leave the platform to do?
The org-level framing behind the roadmap: each workstream traces back to one of these six capability areas.
Workstream 01 — Document Execution v1
Zillow's first audit-compliant signing workflow
Document Execution v1 was the highest-stakes design on the roadmap. It had to satisfy Legal requirements for audit integrity while remaining usable for non-technical consumers — home buyers who would receive a signing request with no Zillow account and no prior experience with e-signatures. The core insight: Action Codes were a Legal requirement that also happened to produce a better consumer experience than requiring account creation.
Every state a document can occupy, and the actions that move it between them. This diagram — used directly with Engineering — became the source of truth for what "status" meant across the signing experience.
Workstream 02 — Customer Ops Form Library
One form, one source of truth — across every association and brokerage
The foundational design principle: a single authoritative form per scenario — no copies, no local customizations, no version sprawl. When an association updates a form for a regulatory change, every agent in that association sees the update automatically. I designed a three-tier inheritance model (Association → Organization → Workspace) and a Supportal management interface for Zillow's Customer Ops team.
Agents downloaded forms from association websites and stored local copies. When an association updated a form, agents had no reliable way to know. Version sprawl was the norm — compliance risk was constant.
Customer Ops publishes forms through Supportal with full version history — publish, schedule, revert, replace base doc. Agents see only the forms they're entitled to. Version confusion eliminated.
"Designed a centralized form management system for Customer Operations, enabling internal teams to manage, update, and reuse forms independently. Reduced manual duplication, increased compliance, and improved operational efficiency."
— ProEx Talent Profile, Sahithya Baskaran (Manager)Workstream 03 — Document Intelligence (L1–L3)
AI that proposes. Agents who decide.
Unlike Listing Kit's fully automated approach, Zillow's Document Intelligence was designed around a core principle: AI proposes, humans confirm. At every step of the processing pipeline — split, classify, extract, assign — the agent reviews and approves before anything is committed to the transaction record.
Agents could hit document processing from four different surfaces — chat, a listing sidebar, the documents tab, or a dedicated hub. Mapping every entry point end-to-end was what surfaced how much the confirmation and classification steps needed to be shared, reusable components rather than one-off screens per surface.
The technical backbone under the AI review panel below: document intelligence classifies the file, catches type mismatches and composite packets before they cause confusion, then extracts and maps field data — with a human review step gating every commit to the engagement record.
Key Iterations
The design decisions that changed direction
After AI finished processing a document, the interface showed a "Reviewed" status. This implied the document requirement was fulfilled — but AI completion and document fulfillment are two different states. Agents were confused about whether they still needed to act.
Changed to "Uploaded" — reflecting what's actually true about the document, not what the AI did. Action stays "View" because agents can still edit and prepare after upload. A small label change with significant clarity implications.
Early designs required agents to manually re-split rejected files using a drag-and-drop page organizer. Added significant complexity to the MVP flow for an edge case that could be handled more simply.
Simplified to a binary accept/reject model. Rejected splits go to Unassigned Files where agents deal with them later. The MVP stays clean without sacrificing data integrity.