Case Study · Compass · 2021–2022
End-to-End Transaction Management
Folding an acquired transaction platform into Compass and designing offer management for both sides of a real estate deal, from first offer to close, plus an applied-AI feature that flags missing signatures and inputs.
- Disciplines
- Product Design · AI Interaction Design · User Research · Platform Integration
- Categories
- Proptech · Applied AI · Computer Vision · Platform · Workflow · M&A Integration
22%
Lower transaction processing costs (FY23)
Most US markets
Buy-side offer management rollout
Buy + sell
Both sides of the deal
Context
Compass acquired Glide, a real estate transaction startup, in July 2021. Transaction management was the missing piece in the Compass platform, and my team folded Glide's product into Compass's broader real estate ecosystem.
My role
I owned design for offer management, the new product at the center of the integration, covering both the buy and sell side of a deal.
Goal
Offers get written and run in wildly different ways. The goal was to streamline the process for agents on both the buy and sell side, across the building blocks of an offer:
- MLS and property search
- Form fill and e-signature
- Document reading through OCR
- Offer terms management
- Email composition and notifications
- Brokerage commissions and compliance workflows
Approach
I mapped the full offer journey across pre- and post-contract for both the buy and sell side, then worked through the bi-directionality between Compass's internal tooling (Business Tracker) and Glide's external transaction tooling. From there I designed the buy-side offer flows:
- Start an offer: add a property and open an offer against it.
- Add documents: upload, transfer, or pull from form libraries.
- Add an offer for compliance only: log an accepted offer to kick off compliance and commissions.
- Submit an offer: a guided wizard covering parties, key terms, and an email composer.
Applied AI
The clearest test of the platform's applied-AI ambition was missing input detection: a machine-learning and computer-vision feature that reads a flattened PDF and flags the signatures, dates, and initials a document still needs. That work usually falls to transaction coordinators, who scan hundreds of pages by hand, where a single missed field can carry legal consequences.
AI that is sometimes wrong cannot speak with one voice. Before designing a single screen, I wrote three principles the feature had to honor:
- Transparency and grading: surface how confident the model is, and make that confidence part of the interface instead of hiding it behind a single yes or no.
- Consent: agents turn detection on themselves; the system never acts on their documents by default.
- A feedback loop: every correction an agent makes trains the model, so accuracy compounds the more the feature is used.
Those principles became an interface keyed to the model's own confidence. A low-confidence detection asks the agent a question and stays out of the way. A high-confidence one states a finding and offers a resolution, with a dismiss path for false positives and a lightweight accuracy rating that feeds the loop.
Outcome
Buy-side offer management shipped to most US markets. The platform-integration and applied-AI work it was part of cut transaction processing costs 22% entering FY23, and adoption was tracked through a HEART framework, measuring happiness, engagement, adoption, retention, and task success, to guide the next round of iteration.