What AI Is Actually Doing in Real Estate Right Now (vs. the Hype)
Real estate has a complicated relationship with AI. On one side, the headlines: billions in projected market growth, sweeping claims about disruption, tools promising to replace agents, appraisers, and analysts overnight. On the other side, the reality most operators are actually living: a Zillow price estimate that’s wrong by 15% on a unique property, a chatbot that can’t answer a basic question about a listing’s HOA fees, and a pilot program that never made it to production because the data infrastructure wasn’t ready.
Both sides exist. The hype isn’t entirely wrong — AI is producing real, measurable results in specific parts of the real estate industry. But the results are narrower, more specific, and more dependent on data quality than the market projections suggest. The organizations getting genuine value from AI right now are the ones who understand the difference between where AI works and where it doesn’t — and have stopped trying to use it everywhere just because it’s available.
This post is that honest map. What’s working in production at scale, what works in narrow conditions but gets overstated, what has been tried and failed, and where the open problems are that good AI implementation could actually solve. Not a forecast — a ground-level account of where things actually stand in 2025.
The Numbers: What They Mean and What They Don’t
The market size figures for AI in real estate vary so dramatically across sources that they’re almost meaningless without context. Some reports cite the market at $2.9 billion in 2024. Others put it at $222 billion, or $402 billion, depending on what they’re measuring — whether it’s purpose-built real estate AI tools, the broader software market that includes AI features, or the total addressable market of real estate services that AI could theoretically touch.
The number that’s actually useful is this one: 82% of real estate agents in the US are now using AI tools in their business, according to RPR’s 2026 survey of NAR members. That figure isn’t about what AI could do — it reflects what’s happening today, in practice, across the agent population. And when you look at what those agents are actually using AI for, the picture becomes specific very quickly. Writing tools are dominant — 78% use AI for listing descriptions, marketing copy, and emails. Chatbots and AI assistants come next at 47%. Market analysis and pricing tools are used by 39%.
What those agents are not doing with AI, in large numbers, is using it for pricing decisions, compliance-sensitive conversations, or complex market interpretation. Not because the tools don’t exist — they do — but because 63% of agents cite accuracy as their primary concern with AI output, and 49% cite compliance risk. The adoption is wide. The trust is narrow. And that gap between adoption breadth and trust depth is the most important thing to understand about where AI in real estate actually is right now.
Morgan Stanley’s analysis of 162 REIT and CRE firms with $92 billion in combined labor costs projects $34 billion in AI-driven efficiency gains for the real estate industry by 2030. That’s a significant number. It’s also a five-year projection, not a current result — and it reflects potential efficiency gains if AI is implemented well across the industry, not the results firms are achieving today. The gap between projected potential and current reality is where most of the AI work in real estate actually needs to happen.
What’s Working: AI in Production at Scale
These are the applications where AI is generating documented, measurable results in real estate operations right now — not in pilots, not in press releases, but in production at organizations that are running at meaningful scale.
AI leasing agents for multifamily are the clearest proof point that AI can handle a complete real estate workflow end to end. EliseAI — which raised at a $2.2 billion valuation in 2025 — operates as a voice and text agent that handles inbound leasing inquiries, qualifies prospects, schedules tours, processes maintenance requests, and follows up on application status around the clock. The key design decision that makes this work is scope discipline: EliseAI is not a general real estate AI assistant. It’s a purpose-built leasing agent with deep integration into property management systems, trained on leasing-specific conversations, and constrained to the workflows it handles well. Multifamily operators using it report meaningful reductions in leasing staff workload and measurable improvements in response time. This works because the problem is the right shape for AI: high volume, repetitive, bounded, and well-defined enough that the AI can handle it reliably without human oversight on every interaction.
Automated valuation models (AVMs) for residential property are genuinely mature and genuinely useful within their limitations. HouseCanary, CoreLogic’s Cotality, and Zillow’s Zestimate all produce valuations that are accurate enough to be operationally useful for residential properties in markets with good comparable transaction density — sub-3% median error rates in normal markets, per HouseCanary’s published performance data. Six federal agencies finalized AVM quality control rules in June 2024, which is a signal of regulatory maturity rather than concern: regulators don’t write rules for things that aren’t being used at scale. The limitations are real and important — AVMs break down on unique properties, in thin markets, and for commercial assets where comparable data doesn’t exist at the same density as residential. But within those constraints, AVMs are production infrastructure for iBuyers, lenders, and portfolio management firms.
AI-powered lease abstraction and document processing is the use case where large language models have added the most tangible value to commercial real estate workflows. Platforms like Kira Systems (now part of Litera) and Luminance extract key clauses, critical dates, rent escalation schedules, tenant options, and insurance requirements from dense commercial leases that would otherwise require hours of attorney review per document. The output isn’t perfect — complex or unusual clause structures still require human review — but for standard clause types in institutional-grade leases, the accuracy is high enough that legal teams are using it in production. The efficiency gain isn’t marginal: a portfolio review that previously took three weeks of attorney time is running in days. That’s a real business outcome.
Tenant screening automation has reached a level of sophistication that property managers are genuinely relying on. Snappt’s AI document fraud detection — which identifies manipulated pay stubs and bank statements with claimed 99% accuracy — addresses a specific, high-value problem that manual screening was handling poorly. Traditional screening couldn’t reliably detect sophisticated document manipulation. AI pattern recognition can. The application is narrow and specific, which is exactly why it works.
Behavioral search ranking in consumer real estate platforms — the AI layer that learns from what listings a user views, how long they spend on each, what they save, and how they refine their search — is working quietly and invisibly in Zillow, Redfin, and Realtor.com. Users don’t experience it as “AI.” They experience it as search results that feel more relevant the more they use the platform. The fact that this AI application is invisible is a signal of its maturity: it’s doing its job well enough not to get in the way.
What Works Narrowly but Gets Overstated
These applications are real and genuinely useful — but they’re being marketed as broader solutions than they actually are, which creates adoption disappointment when the reality doesn’t match the pitch.
AI chatbots for lead qualification work when they’re purpose-built with real estate domain knowledge, connected to live inventory, and integrated into a CRM so that the conversation history follows the lead into the agent’s workflow. They fail — visibly and consistently — when they’re generic LLM wrappers with no live data, no real estate context, and no integration layer. The technology gap has essentially closed; the implementation gap hasn’t. Most brokerage chatbot deployments are underperforming not because chatbots can’t work in real estate, but because they were implemented as a quick add-on rather than a purpose-built workflow tool. The distinction matters because the conversation about “AI chatbots in real estate” is dominated by the underperforming implementations, which makes the well-implemented ones harder to recognize and learn from.
AI writing tools for listing descriptions and marketing are genuinely useful for generating first drafts quickly. They’re being oversold as replacements for skilled real estate copywriting. An AI-generated listing description for a three-bedroom ranch house in a standard subdivision is probably good enough to publish with light editing. An AI-generated description for a $4 million historic property in a competitive luxury market is going to require significant rewriting, because the tone, the vocabulary, and the judgment about which features to emphasize require market knowledge and writing skill that AI doesn’t bring to the task. Agents who treat AI writing tools as draft generators get real value. Agents who treat them as finished output generators get descriptions that look competent and convert poorly.
Dynamic rent pricing algorithms — Yieldstar, LRO — are technically working: they optimize prices using market data, occupancy patterns, and competitor rates. The controversy around these tools is not about technical failure. RealPage faced a DOJ antitrust investigation in 2024 for alleged algorithmic price coordination across competing landlords. The legal and ethical questions about algorithmic rent pricing are separate from the question of whether the technology works — it does — and any property manager evaluating these tools needs to engage with both dimensions, not just the performance claims.
AI for investment deal screening helps analysts process more inbound volume faster. DealMachine’s Alma, Dealpath’s AI layer, and similar tools can run initial filters, pull comparable data, and surface preliminary risk flags more quickly than a junior analyst working manually. What they can’t do is make the judgment calls that experienced underwriters make on complex assets: reading the story behind the numbers, evaluating the quality of the operating team, assessing market dynamics that don’t show up cleanly in the data. AI is augmenting investment analysis in CRE, not replacing it. The tools that position themselves as the former are more credible than the tools that position themselves as the latter.
What Has Been Tried and Hasn’t Worked
The honest failures — where AI has been applied in real estate, hasn’t produced the promised results, and the gap between promise and performance is structural rather than a matter of implementation quality.
AI for CRE capital markets prediction is the clearest failure category. NAIOP’s 2025 research is direct: recent AI advances have not meaningfully improved predictive analysis in real estate investing. The core problem is data, not technology. Commercial real estate transactions happen through unstructured documents — emails, PDFs, phone calls, handshake deals — rather than through electronic platforms that capture structured data. The training data that would let an AI model predict CRE capital market movements with genuine accuracy simply doesn’t exist in the form AI needs it. What exists is a lot of structured data about residential transactions and a much thinner, less structured record of commercial ones. Until the data layer improves, CRE prediction models will continue to be confidently wrong on the edge cases that matter most.
General-purpose AI assistants applied to real estate workflows have a poor track record. JLL’s 2025 survey of over 1,000 senior CRE decision-makers found that only 5% of organizations are achieving real results from AI, with 95% still searching for their breakthrough. The pattern is consistent: a firm deploys a general AI assistant (or builds a custom GPT, or stands up a chatbot), discovers it doesn’t have the domain-specific knowledge and workflow integration to be useful in real real estate contexts, and the initiative stalls. The pilot produces impressive demos. The production deployment doesn’t make it. The problem isn’t AI capability — it’s that real estate workflows require real estate domain knowledge, live data integration, and workflow context that general-purpose AI tools don’t bring.
AI replacing agent judgment in negotiations has not been demonstrated in any serious deployment. The structured components of negotiation — comparable analysis, market timing data, pricing trend context — benefit from AI support. The judgment components — reading a seller’s motivation, understanding when a relationship is at risk, knowing when to push and when to hold — haven’t yielded to AI. This isn’t a temporary limitation waiting for a better model. It’s a reflection of what negotiations actually require, and honest AI vendors in this space are positioning their tools as decision support, not decision replacement.
Automated valuation for unique or complex commercial assets consistently underperforms. A 500-unit multifamily community in a market with many comparables: AVMs are useful. A custom hospitality asset, a medical office building, a mixed-use development in a submarket with three comparable transactions in the last three years: the comparables don’t exist at the density the model needs. Income approach valuations for complex CRE assets require human judgment about cap rate selection, tenant credit quality, lease structure, and market positioning that no current AVM handles reliably. This is widely known and widely underreported because AVM vendors focus their marketing on the residential use cases where their accuracy is genuinely strong.
Where the Open Problems Are
The gaps where the pain is real, the demand is demonstrable, and current AI solutions are inadequate — these are the most interesting parts of the landscape.
Maintenance workflow AI — the gap between a tenant submitting a maintenance request and a vendor arriving with the right scope of work and the right parts — is largely unautomated. The data for AI dispatch exists: issue type, property history, vendor availability, past performance, parts inventory. The orchestration layer connecting these into an automated dispatch and communication workflow doesn’t exist at scale in any current platform. This is a tractable problem that nobody has solved well.
Lease compliance monitoring at portfolio scale — tracking rent escalation triggers, option exercise windows, insurance certificate renewals, HVAC maintenance obligations, and co-tenancy clause thresholds across hundreds of commercial leases — is currently done in spreadsheets or with expensive manual review. The documents exist. The AI extraction capability exists. The ongoing monitoring workflow connected to the extracted data doesn’t.
Construction draw review — reviewing hundreds of pages of draw requests, lien waivers, and completion certifications for development projects — is high-stakes, time-consuming, and structured enough for AI to help significantly. No purpose-built AI solution exists for this workflow.
Fair housing compliance monitoring — detecting patterns in listing language, tenant screening decisions, and communication that might indicate fair housing risk — is a problem where the regulatory stakes are high enough to create real demand and the technical approach is achievable. It’s largely unsolved at the platform level.
Conversational AI for investor relations — letting LPs ask natural language questions about their investment and get answers drawn from actual fund data, capital account records, and portfolio documents — is technically achievable with current retrieval-augmented generation approaches. The implementation in investment management platforms is almost entirely absent.
AI for HOA and community management is an enormous underserved category. Document management, resident communications, violation tracking, financial reporting, rule enforcement — all of these are high-volume, repetitive, and data-rich enough for AI to help significantly. Vantaca’s acquisition of HOAi in November 2024 at a $1.25 billion valuation signals that the market is beginning to recognize this, but the implementation depth is still early.
The Data Problem That Explains Everything
The single most important insight from any honest analysis of AI in real estate is this: the results correlate almost perfectly with data quality. The applications that work — residential AVM, multifamily leasing automation, tenant screening fraud detection, lease document extraction — all operate on high-volume, structured, relatively clean data. The applications that fail — CRE prediction models, general portfolio AI assistants, automated CRE valuation — fail on thin, unstructured, inconsistent data.
Real estate as an industry has a data problem that predates AI by decades. 61% of real estate companies still rely on legacy systems, and 60% still use spreadsheets for reporting. The industry’s data is fragmented across MLS boards with different schemas, property management systems with different data models, county assessor records with different formats, and deal histories that exist in emails and PDFs rather than in structured databases. AI models need clean, voluminous, well-structured training data. Most real estate organizations don’t have it yet.
This isn’t a reason to wait on AI. It’s a reason to treat data infrastructure as the prerequisite for AI investment — because the firms getting results from AI are, without exception, the ones who solved their data layer first. The AI strategy question isn’t “which AI tool should we deploy?” It’s “is our data in a shape that AI can actually use?” That question, answered honestly, determines more about AI outcomes in real estate than any other single factor.
What This Means for Real Estate Operators and Builders
For operators running property management companies, investment firms, brokerages, or development businesses: the most productive frame for AI right now is specific use cases with well-defined data requirements, not platform-wide transformation. Start with the workflow where you have the cleanest data, the most clearly defined process, and the highest volume of repetitive decisions. That’s where AI delivers disproportionate value quickly. Maintenance dispatch, lease renewal outreach, tenant communication triage, document extraction for portfolio reviews — all of these are tractable starting points.
For builders and product teams designing real estate software: the open problems in the landscape above are product opportunities, not research problems. The data exists. The AI capability exists. What’s missing is the purpose-built application layer that connects them to specific real estate workflows with the right data model, the right integration layer, and the right human oversight design. The firms building that layer — purpose-built for specific real estate use cases, not general AI assistants applied to a new vertical — are the ones producing the results that 95% of the market is still searching for.
The gap between 82% agent adoption and 5% of firms achieving real results is not a gap in AI capability. It’s a gap in implementation depth. That gap is where the real work is, and where the real opportunity is.
This is the first post in our series on AI in real estate. Over the next several posts, we’ll go deep on specific use cases: what the technology actually requires, what the best implementations look like, and where the remaining open problems are.
If you’re building a real estate platform and thinking through where AI belongs in your product roadmap, let’s talk through the specific use cases that would have the highest impact given your current data infrastructure.