Ben Miller, co-founder of Fundrise and CEO of RealAI, told Ben Lorica on the Generative AI in the Real World podcast how RealAI evolved from an internal data project into an AI-powered analyst tool. The central claim Miller advanced is that the competitive advantage lies less in a better language model and more in a proprietary, domain-specific data layer that general-purpose tools cannot easily replicate.
What RealAI does and who it serves
RealAI is built for professional, institutional real-estate investors who buy and manage properties worth tens of millions of dollars. Traditional real-estate analysts perform due diligence, build financial pro forma models, produce investment memos and manage assets post-acquisition — workflows that are still largely manual and spreadsheet-driven.
Miller emphasized that a typical apartment building emits a large volume of operational and tenant data (maintenance events, leasing flow, source of tenants, demographics), but the industry usually aggregates these raw events into sparse reports and Excel sheets, losing much of the original signal.
The data model: place, property, person
As Fundrise moved from a fintech investor platform toward vertical integration (owning and operating assets), the company discovered and collected two core types of data: time-series and comparative property/place data, and people-level data. Miller pointed out that legacy property management/accounting systems (examples mentioned: Yardi, Tratta, RealPage) are address/unit-centric, but typically lack a rich person-level data model.
To address that gap, Fundrise collected granular people data — who moves in and out, age, income indicators, device usage, and more — and integrated those attributes with property and place datasets. This fusion enabled new correlations and insights not available from traditional aggregated reports.
Technical architecture and deterministic analysis
Miller described a two-layer storage and compute architecture: an analytical layer where many computations are pre-computed nightly (Snowflake-based, with clustering and aggregations by city/block), and a separate transactional, low-latency store used for live API queries. Pre-computation of metrics (market growth, rent trends, clusters) enables fast, deterministic responses to user queries.
The user-facing interface is chat-based. A language model (RealAI has used Claude and moved some pieces to Vercel) interprets prompts and translates intent into queries, but it is not relied on to perform core financial calculations. Instead, Miller characterizes the LLM as an orchestrator that invokes tools (web search, calculator, or deterministic SQL/analytic queries). Factual values are pulled from the company’s databases to reduce hallucinations.
Models, distribution and platform strategy
Miller acknowledged the convenience of general-purpose models like Claude and ChatGPT and explained RealAI’s pragmatic distribution choice: they built a connector and a plugin to appear in the Claude marketplace so they can reach customers where they already work. Still, Miller argued that long-term defensibility comes from proprietary, cleaned, and pre-computed data rather than from any single model; even if models become open-source or specialized, the premium lies in the data layer and product integration.
He also discussed the broader ecosystem trends around fine-tuning, reinforcement learning, and pre-training for specialized models, but cautioned that open-source or new model releases do not automatically translate to better customer outcomes without the underlying data and product work.
UX: communicating uncertainty and credibility
On user experience, RealAI surfaces meta-information about data coverage and methodology alongside answers. Miller observed that users initially focus on validating data quality; once they trust the data, they are less concerned with methodological caveats. The product therefore includes information about how a metric was calculated (for example median vs. point-to-point growth) to help users interpret uncertainty.
AI’s impact on real estate demand and the labor market
Miller described how AI accelerates divergence in economic outcomes — a K-shaped effect — which has implications for real-estate investment. High-end, price-insensitive housing markets benefit from concentration of wealth, while affordable and workforce housing face political and regulatory pressure that constrains profitability. For investors, this suggests favoring affluent or luxury-oriented markets.
He also pointed to the vulnerability of middle-office and back-office white-collar roles (positions that had helped populate mid-sized markets such as Charlotte, Tampa or Orlando). Many of those jobs, which influenced where young professionals rented or bought housing, are exposed to AI-driven automation, diminishing demand in certain urban rental markets.
Regarding hiring, Miller painted a mixed picture: some roles are needed less, while other specialized roles expand. He said the current large capital investments in data centers (the “trillion-dollar a year AI CapEx” he mentioned) are temporarily supporting labor demand on-site, but he warned this is a multi-year build and that once the investment wave subsides there could be a deeper trough in white-collar employment. That trajectory also affects the traditional talent pipeline: with fewer entry-level positions, companies may need to accelerate turnover or raise expectations for mid-level employees.
Pricing and product maturity
On monetization, RealAI currently uses a hybrid of seat-based tiers and usage tokens: a seat purchases a bundle of tokens, so pricing mixes package and usage elements. Miller acknowledged they are still refining the right pricing model as adoption in real estate lags other sectors.
Takeaway
RealAI’s story underscores that for industry-specific AI products, the differentiator is often not solely the underlying LLM but the proprietary, pre-processed data and the deterministic computation layer that feeds the model. Fundrise’s approach — integrating granular people data with property and place analytics, pre-computing heavy metrics, and using the LLM as an orchestrator — aims to deliver credible, actionable analysis for institutional real-estate decision-making while navigating distribution choices and the broader labor-market effects of AI.
