Real estate AI has two faces. One is the low-risk, high-value productivity layer — better listings, faster search, automated tour scheduling. The other is the regulated zone — tenant screening, valuation, lending — where fair-housing law imposes strict constraints on what automation can actually decide. Shipping proptech that works means knowing which face you are building for.
The safe automation wins
Listing generation from facts
Structured property data (bedrooms, sqft, neighborhood, features) into compelling listing copy. LLMs do this well, agents accept the output, time-to-market drops from 2 days to 20 minutes. One of the highest-adoption AI features in real estate because the failure mode is obvious (bad copy) and recoverable (rewrite).
Photo processing and quality control
Auto-tagging photos (kitchen, master bath, backyard), flagging low-quality shots, re-ordering for best first impression. Computer vision is mature here; value is in speed and consistency.
Search refinement
Users describe what they want in natural language: 'quiet neighborhood with good schools, walkable to coffee shops, under $700K.' LLM parses into structured filters, runs against MLS, presents results. This is where consumer real estate apps compete hardest in 2026.
Tour scheduling and coordination
Calendar coordination across agent, buyer, and seller. Historically a back-and-forth phone game; now handled by agent copilots that propose slots and confirm automatically. Low-risk, high-value productivity.
The regulated zone: be careful
Automated Valuation Models (AVMs) are fine to use — they're a mature product category — but they must be explainable and the user must understand they are estimates. LLMs can add narrative ('this property is priced above comps because of the recent kitchen remodel and the premium lot position') but cannot replace the AVM's statistical core. Regulators and lenders have specific requirements around comparable selection, adjustment methodology, and confidence intervals.
Tenant screening is the highest-risk use case. HUD, state law, and city ordinances impose strict rules on what factors can be used and how. Automated decisions that deny housing based on AI scoring are legal minefields. Our guidance: AI can flag applications for review and highlight specific items in public records, but the decision is always a human one with a documented reason that references allowable factors.
Lending and risk scoring similarly require regulator-explainable models. LLMs can assist loan officers — pulling relevant documents, summarizing financials, drafting condition letters — but the underwriting decision runs on traditional models with audit trails.
What we deploy
Agent copilots are the sweet spot. A real estate agent spends most of their day on admin: scheduling, emails, CMAs, listing copy, buyer follow-ups. An AI that handles these tasks while the agent focuses on client relationships has clear ROI and avoids the regulated zones.
Consumer-facing search-and-discover products compete on UX more than on AI quality these days. Features users notice: natural-language search, conversational refinement ('show me more like this one but cheaper'), saved-search alerts with contextual reasoning, commute-time overlays.
Investor tools — identifying promising properties for flip or rental — are a narrower market but a willing-buyer one. Aggregate listings, zoning, permits, neighborhood trajectory data; score opportunities; explain reasoning. Avoid fair-housing sensitive features (targeting specific demographics, area avoidance based on protected-class proxies).
Common mistakes we see
Overreaching automation. A startup builds a 'fully automated tenant screening' product, gets sued. Disastrous unit economics on chatbot-only lead handling — real estate sales close with humans; AI is for qualifying and routing, not closing. Ignoring MLS data rules — MLS agreements have strict rules on data use and display; ignore them at your peril. Underinvesting in mobile. Most agent and consumer real estate work happens on mobile; AI features that only work on desktop barely get used.