CASE STUDY · OFFERWELL

Deciding what should — and shouldn't — become AI.

Offerwell is building AgentOS, a transaction platform for real estate agents. I wrote the AI product strategy: which problems genuinely warranted a model, which interaction model each one deserved, and what to give away to win the market.

AI PRODUCT STRATEGY·INTERACTION MODEL·CONCEPT DESIGN·2024

The three-part filter applied to every proposed AI feature

Every proposed feature had to clear all three before it earned engineering time.

01 TRIGGER

One button, and a compliance problem disappears.

Fair Housing violations in listing copy are a real liability and every agent writes them by accident. Generated descriptions get screened for steering language, familial status, protected-class proxies, and unverifiable claims. I argued this should ship free to every agent on the platform — compliance is the wedge, not the upsell.

01 · Trigger-based

02 PROMPT

An offer written from what actually won.

The agent describes the constraints in plain language; the system drafts terms informed by patterns across accepted offers in that market. The reasoning panel matters more than the output — rent-back beat price in most wins, which is not what most agents assume. Patterns only: the system never surfaces the terms of an individual competing offer.

02 · Prompt-based

03 CHAT

The report contradicts its own summary.

A 41-page inspection buries a $2,000 sewer line under fourteen cosmetic findings. Conversation is the right interaction model here because the question is comparative and open-ended — scoring one report is useful, comparing every property on the table is the actual job.

03 · Chat-based

Strategy, not shipped product. The framework outlived the roadmap.

Concept design — created to pressure-test the strategy, not shipped to production.