What are the risks and challenges of using AI in real estate?

Why does AI in real estate carry more risk than in other industries?

Real estate combines three characteristics that raise the risk of AI relative to the average enterprise use case. Transactions are high-value and infrequent. A single AI-driven error on a $2M sale or a corporate lease can produce material financial damage, and the client has no volume of similar transactions to average out any single mistake.

The industry is heavily regulated across multiple overlapping frameworks. Fair housing law (HUD, state civil-rights acts), securities law (SEC for REITs and syndications), consumer-protection law (FTC), data-protection law (GDPR, CCPA, state analogues), and licensing regulations for agents and brokers all apply simultaneously. AI outputs that are legal in a marketing context can trigger enforcement in a housing or securities context.

Client trust is the actual product. Real estate is a relationship business built on fiduciary duty. Clients discovering that an agent's advice was generated by an unvetted AI, particularly if that advice turns out to be wrong, damages the relationship in ways that generic industries do not face. Consumer trust in real estate is already the second-lowest of any regulated profession, per Gallup's annual Honesty and Ethics survey, and AI missteps compound the problem rather than reversing it.

These three factors together mean that AI risks in real estate need a higher baseline of caution than in, say, an e-commerce or productivity setting. Not to avoid AI. To use it inside guardrails that actually protect the operator and the client.

For the underlying framework of where AI genuinely helps real estate operators, see our AI in real estate blog post and our AI solutions service page.

What is the hallucination problem, and why does it matter for property data?

Language models like ChatGPT, Claude, and Gemini generate answers by predicting the most probable next token, not by looking up verified facts. Vectara's Hallucination Leaderboard tracks factual accuracy across major models and finds hallucination rates ranging from around 3 percent for the best-performing models to over 27 percent on longer, more open-ended tasks.

In real estate, this shows up as confident but wrong outputs. An AI-drafted listing that lists 2,400 square feet when the actual is 2,100. A comparable-sales analysis that cites transactions that never happened. A tenant email that quotes an HOA fee off by $150 per month. A market update that references a data point invented from correlations rather than pulled from a source. Each of these looks polished and confident, which is precisely the problem: humans review confident writing less carefully than obviously rough drafts.

The controls that reduce hallucination risk are well-defined. First, keep every high-stakes fact behind a source citation. If the AI cannot cite where a number came from, treat the number as a draft. Second, use retrieval-augmented generation (RAG) architectures that ground the AI in your own MLS feeds, PMS data, or CRM records rather than the model's training data. Third, mandate a human review of every AI-generated document that goes to a client, an investor, or a regulator before it leaves the office.

How does AI create fair-housing and discrimination risk?

This is the highest-consequence risk category in real estate AI, and the one most operators underestimate. The US Fair Housing Act, and analogous laws in the UK, EU, India, and most operating markets, prohibit discrimination in housing on the basis of protected classes. Algorithmic tools that produce discriminatory outcomes are covered even when the operator did not intend discrimination and even when the protected class was not an explicit input.

Three real-world examples make this concrete. In 2024, the US Department of Justice filed a civil antitrust complaint against RealPage alleging that its algorithmic pricing software coordinated multifamily rents across landlords, with fair-housing implications embedded in the pricing outcomes. HUD has published guidance treating tenant-screening algorithms and targeted-advertising algorithms as subject to Fair Housing Act enforcement, including a 2023 charge against Meta for the ad-delivery algorithm's demographic filtering. Multiple state attorneys general have opened investigations into algorithmic tenant-screening vendors for disparate-impact outcomes.

The specific risks in an AI-enabled real estate workflow are: AI-generated listing copy that inadvertently encodes preferences (steering language toward or away from protected classes), targeted advertising that filters delivery on demographic proxies (postcode, ethnicity-correlated interests), tenant-screening AI that weighs criminal-record data or eviction data in ways HUD has flagged as discriminatory, and pricing AI that surfaces different offers to different demographics.

The controls: audit every AI-touched pricing, screening, and advertising output for disparate impact quarterly. Do not use consumer proxies (postcode, name origin, language preference) as targeting inputs on housing-related campaigns. Document the human decision-making review step on every AI-generated tenant-screening or pricing output. Prefer vendors that publish their fair-housing testing methodology.

What data can and cannot go into a public AI tool?

The default assumption should be that anything typed into a public AI tool (free ChatGPT, free Claude, free Gemini) may be used to train future models, retained in company logs, or exposed in a breach. IBM's 2024 Cost of a Data Breach Report puts the average global breach at $4.88M, with data pasted into unauthorised AI tools (so-called "shadow AI") now identified as an emerging cost driver.

The framework agencies should apply.

Never paste into a public AI tool. Full client PII (name, address, phone, email, SSN, government ID), signed contracts, financial statements, mortgage pre-approvals, tenant applications, offer letters, seller net-sheet numbers, tenant income and employment data, corporate deal terms under NDA, or any data covered by GDPR, CCPA, HIPAA (if applicable), or state analogue.

Safe to paste into a public AI tool. Public listing data (already published on MLS, portals, or the agency website), generic market commentary, general educational content ("explain HOA fees"), high-level campaign strategy without client names, and drafts where all identifying data has been redacted.

For anything sensitive, use enterprise-tier AI. ChatGPT Enterprise, ChatGPT Team, Claude for Enterprise, Anthropic's zero-retention API, and Azure OpenAI or AWS Bedrock deployments all offer contractual guarantees that inputs are not used for model training and that data residency, retention, and access controls can be configured to meet GDPR, CCPA, and industry-specific standards. Cost is materially higher than the free tier, but so is the compliance footing.

For an agency operating at scale, standing up an enterprise-tier AI deployment with clear usage policies is the single highest-impact risk-reduction step. Without it, agents will paste client data into free tools regardless of what the internal policy says. For CRM implementation and data-flow work, see our CRM implementation service.

What is vendor lock-in, and how does it affect AI workflows?

The AI vendor market in 2026 is unstable in ways real estate operators do not fully price in. OpenAI, Anthropic, Google DeepMind, and Meta all deprecate models on cycles ranging from 6 to 18 months. A workflow built on GPT-4 in early 2024 needed migration to GPT-4o in mid-2024 and to newer models in 2025. Each migration surfaces subtle behaviour changes that break workflows in unexpected ways.

Vendor lock-in shows up in three forms in real estate AI. Data lock-in: the agency's CRM, MLS integration, and prompt library are configured for a specific vendor's format, making it expensive to switch. Workflow lock-in: agents have learned a specific vendor's tools and productivity drops during migration. Contract lock-in: multi-year enterprise contracts that lock in pricing at rates that look bad against next year's market.

The controls that reduce lock-in risk. Prefer vendors offering an OpenAI-compatible API (multiple providers now offer this so switching costs are low). Store your prompt library, prompt-evaluation results, and RAG-grounding data outside any single vendor's system, so a migration means changing an API endpoint, not rebuilding the whole workflow. Negotiate contract exit terms upfront, including a data-export SLA. Budget for annual re-evaluation of model performance across the top three or four frontier vendors.

For teams building AI into a wider PropTech stack, this is one reason architecture matters more than any specific model. Our PropTech platform architecture guide covers the underlying pattern.

What are the cybersecurity risks specific to AI in real estate?

Beyond generic cybersecurity risk, AI introduces three real estate-specific attack surfaces.

Prompt injection. A malicious buyer or seller drops instructions into an email, a listing enquiry form, or a document that gets fed to an AI. The AI reads the malicious instruction as if it were legitimate and executes it (leaking data, forwarding messages, generating misleading content). The OWASP Top 10 for LLM Applications lists prompt injection as the number-one risk, and real estate is a natural target because agents routinely feed inbound emails and documents to AI tools.

Deepfake fraud. AI-generated voice cloning, video, and text impersonation is now cheap and convincing. Real estate is a high-value target for wire fraud (the FBI's IC3 report has flagged real estate wire fraud as a growing category for several years running), and deepfakes materially raise the ceiling of what fraudsters can attempt. A cloned agent voice instructing a title company to redirect closing funds is the archetypal 2026 scenario.

Third-party AI vendor breach. An agency's AI vendor gets breached and every client conversation, every uploaded document, and every prompt logged over the past year is exposed. This is a growing category of incident. The agency's contract with the vendor determines liability, but the reputational damage is the agency's regardless.

The controls: prompt-injection testing on every AI workflow that accepts external input, wire-fraud verification protocols (out-of-band callback on any funds instruction), vendor due diligence including SOC 2 Type II or ISO 27001 evidence, and cyber-insurance coverage that explicitly includes AI-driven incidents.

Which regulations apply to AI in real estate right now?

Regulatory frameworks are still developing, but 2026 already has several enforceable rules operators cannot ignore.

United States. HUD guidance on algorithmic tenant screening and targeted advertising. FTC Section 5 on unfair or deceptive AI practices. State-level laws in Colorado (SB 24-205 on high-risk AI), California (multiple bills), and New York City (Local Law 144 on hiring algorithms, with analogues emerging for housing). SEC guidance on AI use in investment advice, relevant for REITs and syndications.

European Union. The EU AI Act, which entered into force in phases from 2024 through 2027, classifies certain AI systems used in real estate (particularly tenant-screening and credit-scoring uses) as "high-risk," triggering documentation, human-oversight, transparency, and conformity-assessment obligations. GDPR continues to govern data-processing risk in parallel.

United Kingdom. No dedicated AI Act, but the ICO has published guidance on AI and data protection, and the Equality Act 2010 applies to any AI system producing discriminatory outcomes in housing.

India and Middle East. India's DPDP Act 2023 governs personal data. UAE and Saudi Arabia have both published national AI strategies with regulatory frameworks emerging.

The practical controls: appoint someone on the team to track regulatory developments in every operating jurisdiction, maintain a documented AI use policy that references the applicable regulations, and build a bias-testing and audit cadence that would satisfy an EU AI Act high-risk classification even in markets that do not currently require it. Building to the highest standard now saves rebuilding to it later.

What happens when agents over-rely on AI?

The most subtle risk, and the one most operators dismiss until it costs them a deal. When AI drafts every email, every market update, and every offer letter, agents lose the muscle memory for the work. Junior agents in particular never build the skills that make senior agents valuable: reading a client's real concerns behind the stated ones, drafting an offer that anticipates the counter-move, structuring a difficult conversation.

The pattern is documented across professions. A 2024 MIT study on programmers using GitHub Copilot found productivity gains on routine tasks but skill degradation on non-routine tasks. Similar patterns are showing up in medical, legal, and financial services where AI assistants have been deployed at scale. Real estate is not immune.

The controls: require agents to draft the first version of high-stakes communications (offer letters, listing presentations, negotiation emails) before AI is consulted. Use AI for editing, formatting, and enrichment rather than for first-draft generation on judgment-heavy work. Track the ratio of AI-generated to human-generated content in the CRM and cap it at a level that preserves skill development. Rotate junior agents through AI-free deals periodically so they build the underlying capability.

What are the hidden costs of running AI in real estate at scale?

The demo cost of AI looks low. A ChatGPT Plus subscription is $20 per user per month. Enterprise deployments budget in $30 to $80 per user per month all in. What most operators underestimate is the total cost when AI is genuinely embedded across the workflow.

Token and inference costs at scale. A brokerage running RAG-based property matching, listing description generation, tenant-screening automation, and client-email drafting can burn through 50 to 150 million tokens per month per 50 agents. At current API pricing, that is $2,000 to $9,000 per month in raw model spend, on top of licence fees.

Data engineering and integration cost. AI is only as good as the data it grounds on. Building the MLS-to-CRM-to-RAG pipeline that makes AI outputs accurate typically runs $30,000 to $120,000 in initial engineering plus ongoing maintenance.

Human review layer cost. The controls this post recommends (human review on high-stakes outputs, fair-housing audit, prompt-injection testing) are not free. A mid-size brokerage typically budgets 20 to 60 hours per month of trained-reviewer time.

Rebuild cost on model deprecation. Every 12 to 18 months, the underlying model changes and workflows need re-testing. Budget 40 to 120 engineering hours per major migration.

The honest all-in cost of AI for a mid-size brokerage in 2026 is more like $8,000 to $30,000 per month than the $500 the free-tier demo suggests. Operators who budget for the sticker price and not the total cost end up scaling back after year one.

How do we know if the AI is producing value versus looking impressive?

Attribution is where most AI investments quietly fail. The AI feels productive. Agents report faster email drafting, quicker market summaries, better-formatted listing descriptions. But when the broker actually looks at closed deals per agent, cost per acquired client, and lead-to-close conversion rate over the six months since AI deployment, the numbers have not moved.

This happens because AI improves the visible cadence of work without improving the underlying pipeline economics. Drafting emails faster does not close more deals if the follow-up SLA was the actual bottleneck. Generating better market summaries does not lift lead quality if the source mix was the actual constraint.

The measurement framework: define the specific business metric each AI use is meant to move before deployment (closed deals, CPQL, response time, listing conversion rate, agent retention, hours saved). Baseline the metric for 90 days before AI goes live. Measure the same metric 90 days after. If the metric has not moved, the AI is a cost, not an asset, and either the deployment or the use case needs revisiting.

Operators who run this discipline routinely find that two or three of their AI use cases produce real returns and the rest are neutral or negative. Concentrating spend on the two or three that work outperforms a spread-bet across ten.

For attribution modelling more broadly, see our blog on lead leakage in real estate.

How does AI affect client trust, and what disclosure is needed?

Consumer sentiment on AI in real estate is mixed and shifting. A 2024 NAR consumer survey found that 62 percent of home buyers were comfortable with AI helping their agent find properties, but only 24 percent were comfortable with AI writing the offer letter, and just 11 percent were comfortable with AI making pricing recommendations without agent review. The lower the perceived human oversight, the lower the trust.

The controls that preserve client trust while using AI. Disclose that AI is used, at what stage, and with what level of human review. State this in the buyer or seller representation agreement in plain language. Never present an AI-generated document to a client as if a human wrote it unaided. If AI has been used, name it (draft prepared with AI assistance, reviewed and edited by the agent).

Client-facing disclosure is also increasingly a regulatory requirement, not just a courtesy. EU AI Act transparency provisions require disclosure of AI-generated content in many contexts. Colorado SB 24-205 requires notification to consumers when high-risk AI systems make decisions about them. Even where disclosure is not yet mandated, it is the direction of travel across regulated professions.

The agencies that adopt disclosure early build client trust rather than eroding it, because the alternative (getting caught silently using AI on high-stakes documents) is the fastest way to damage a decade-long client relationship.