How Real Estate Agencies Will Use AI and Automation in 2027

In 2027, real estate agencies and brokerages that sell to buyers from other countries will use AI on two sides of the business, wherever the office and the property happen to be: internally, where clients see the results but not the systems, and in front of clients, on the agency's website and wherever buyers now search with AI.

Internally, to grow the business, agencies will use AI in three layers:

• Layer 1, a smart CRM, captures, remembers, scores, matches and follows up, so the agent works from a useful record rather than another inbox. This is the automation layer, built into the agency's system of record.
• Layer 2, a private AI team, is a set of company-specific AI agents that research properties, prepare due diligence and handle repeatable back-office work for named people to review.
• Layer 3, the guardrails, decides what those machines may read, write, say and send under the rules that apply in each market.

None of the three sells a home. Together, they give a broker more time, better information and a client who feels looked after. They also decide how good the client-facing side can be: a website assistant is only as accurate as the records and inventory behind it, and a buyer's personal AI agent needs readable, verifiable facts from the agency before it can make a dependable recommendation.

This article is for a realtor handling a Canadian buyer in Miami, an estate agent taking a London instruction, a sales team in Dubai Marina and a three-person agency in the Algarve, as much as it is for the Spanish coastal agencies we know well. It explains why so many tools bought in 2025 and 2026 were quietly switched off and how the three internal layers work on a normal Tuesday. It then turns to the client-facing side: what buyers should see on your website and what changes when they hand their search to a personal AI agent. It ends with what to build first.

Writing better listing descriptions or letting ChatGPT draft Instagram posts has become the price of admission, like having a website, rather than a strategy. What follows is about everything after that.

Where agencies actually are with AI in October 2026

Agents have adopted AI personally faster than their firms have adopted it operationally. The 2026 NAR technology report, based on 1,200 agents, found that nearly half used AI daily or weekly and only 12% did not use it at all. Zillow's 2026 Agent Trends Survey similarly found that nearly half used tools such as ChatGPT, Gemini or Claude daily. Yet NAR's leading reported uses were listing descriptions at 75%, social posts at 56%, and emails and follow-ups at 52%. That is useful individual assistance. It does not mean the inventory, CRM and transaction process can think together.

Brokerage leaders know the gap exists. A Delta Media survey reported by Mann Report, covering more than 100 leaders, rated AI's importance at 7.12 out of 10 in 2026, with 53% planning to adopt or expand administrative or back-office AI and 50% planning to adopt or expand agentic tools. The enthusiasm is no longer confined to an adventurous agent with a browser tab.

Buyers moved even faster. NAR reports that 55% of buyers use generative AI at least monthly, chiefly for affordability, education and neighbourhood research. In the UK, Estate Agent Today reported a survey of 500 UK homeowners who had bought in the previous twelve months, in which 15% said they used an AI tool during the purchase. Portals are meeting them inside the assistant: Zillow launched live sale and rental search in ChatGPT, idealista followed with natural-language search, comparison and summaries, and Rightmove became the first UK portal app in ChatGPT. A buyer can now begin with a sentence about a school, a commute and a dog, instead of a row of filters.

The first "agent" many clients consult will therefore be a model. The human agency still owns the difficult parts: understanding an anxious family, noticing the unstated objection, checking the source, arranging access and carrying a transaction through. But it may enter the relationship later, after an assistant has framed the shortlist and named the firms worth calling.

The buyers also arrive from everywhere at once, which is the part a domestic playbook underestimates. Henley and Partners and New World Wealth projected that a record 142,000 millionaires, people with at least 1 million USD in liquid investable wealth, would relocate internationally in 2025, with the UAE, the US, Italy, Portugal and Greece among the projected net beneficiaries. NAR says foreign buyers purchased 67,100 existing homes worth 45.3 billion USD in the US from April 2025 through March 2026, and foreign buyers took a record 15.98% of Spanish home purchases in the second quarter of 2026, led by British, Dutch and German buyers. On one sales desk in Miami, Marbella or Palm Jumeirah, that means passports, currencies, beneficial owners, tax positions and languages from several continents before lunch, and the rules underneath those files keep moving: Spain's residential Golden Visa ended on 3 April 2025, while the proposed tax on non-EU non-resident buyers that generated alarming headlines stalled in Congress in March 2026. Buyers still ask about both on the first call. If three colleagues give three answers, the business has a knowledge problem before it has an AI problem.

All of this is happening while most corporate AI work still fails to become useful operations. Fortune reported MIT NANDA's finding that 95% of AI pilots delivered no discernible savings or profit uplift. S&P Global found that the share of companies abandoning most AI initiatives before production rose from 17% to 42% in a year. Gartner expects more than 40% of agentic AI projects to be cancelled by the end of 2027. Bought, tried, quietly switched off. Many owners we meet are somewhere in that loop, and understandably reluctant to explain a second failed tool to the partners.

Key takeaways

Agents already use AI personally, buyers increasingly begin their search inside an assistant, and we expect many to hand parts of that search to a personal AI agent from 2027. The surveys say nothing about how many agencies have connected AI to the CRM, the inventory and the transaction, and in our experience few have. International files make that gap expensive, because generic automation breaks on the second language, currency or legal system. In our view, the advantage goes to firms that put AI inside the CRM, give a private AI team controlled work and keep both supervised through clear guardrails. Buyers should see the result on the website: homes shown as they are, and facts their own assistant can read and check.

Why the 2025 wave of "AI for real estate" tools ended up switched off

Somewhere in 2024 the internet decided every agency needed an AI. By 2025 the supply arrived: chat widgets that promised to qualify leads, voice bots that promised to book viewings, WhatsApp flows that promised to nurture "on autopilot", and newly minted experts who had never waited outside a notary's office with a missing document. We watched the demonstrations too. They were good. Demonstrations always are, because they have one buyer, one budget, one language and no history.

Then the tool meets the kind of enquiry every international desk knows. Picture a Belgian couple with two budgets, one if the house in Ghent sells and another if it does not. They have a dog, a firm view on morning sun, a father-in-law who needs to walk to a pharmacy, and a preference for Jávea unless Moraira offers a flat garden. A human smiles, asks three questions and moves on. A generic flow forces them through a form, creates duplicates, or recommends a property reserved on Friday. The edge case was never a bug. It was the job.

Now picture a desk in Dubai. An Indian buyer asks about an off-plan apartment, pastes two payment schedules into WhatsApp, and wants to know whether the quoted amount includes every charge. They also ask whether their sister can attend the handover. A bot trained on brochures confidently paraphrases an old schedule. A responsible system retrieves the current project file, marks what it can verify, and assigns the questions about charges and authority to the right human. It knows where its competence ends.

We have seen voice bots cheerfully book sold properties because nobody connected them to live inventory, and messaging flows greet a client of eight years as a new lead. The failures usually come from three places:

• Not integrated. The widget does not know the CRM, the CRM does not know the portal feed, and the calendar does not know the agent is showing a villa in Puglia. Nobody defined who owns the lead after the bot promises a call. A wire in an automation tool does not supply permissions, exception handling or an audit trail.
• Brittle by design. Happy-path rules buckle under two languages in one message, a price in pounds, fourteen unit types or a reply to an old campaign. Every break consumes a partner's evening and weakens the team's willingness to use the system.
• Not compliant. Passports appear in consumer chatbot accounts, conversations move into free tools without a processor agreement, and callers are not told that the voice is artificial. The exact legal duties depend on jurisdiction; Layer 3 separates them rather than pretending one checkbox covers the world.

The deeper problem is that the buyer experiences one agency while the technology sees a collection of channels. The buyer who emailed from a work address, sent a spouse's preferred listing on WhatsApp and called from an airport expects the team to know it is one conversation. If the system cannot resolve identity carefully, preserve context and show uncertainty, adding more automation increases confusion.

Ownership was often missing too. Sales assumed marketing ran the chatbot. Marketing assumed the CRM supplier handled routing. The supplier regarded the custom fields as the client's responsibility. When an enquiry disappeared, everybody could explain why it belonged to somebody else. Useful automation has a named owner, a failure queue and a measure that appears in management meetings.

None of this means an agency should build every component itself. Fortune's account of the MIT NANDA research says purchased tools succeeded about 67% of the time against 33% for internal builds. Buying can be sensible. The mistake is buying a finished promise when what the agency needs is a component fitted into its data, process and accountability.

We wrote a longer post-mortem in why real estate AI implementations fail. The short version is that the model was rarely the whole problem. What was missing was everything around it.

The strategy for 2027: AI prepares, people sell

Here is the thesis this article rests on: AI will not replace the agent who sells a home to a family that has visited the country twice. People making a consequential, emotional purchase abroad want a capable person in the room. What AI will decide is which agent reaches them first with five suitable properties, a sourced answer about the tourist licence and a follow-up that arrives while they are still on the flight home.

The goal is capacity and client experience. We want more qualified viewings per agent, first useful answers measured in minutes rather than hours, long nurturing cycles that do not fall silent, and sellers who receive a proper update without an agent starting from a blank document every Friday. Test every proposed automation against both sides of that goal. If it saves an internal click while making the buyer repeat their story, it has moved the wrong number.

That means leaving judgement where it belongs. A model can notice that the Canadian couple in Lisbon repeatedly saves buildings with lifts, though they never put a lift in the brief. The agent decides whether to ask about mobility, ageing parents or simply heavy luggage. A model can rank matches. The person reads the room, understands why a compromise may work and takes responsibility for the recommendation.

The opposite mistake is expensive too. Owners who regard AI as a passing fashion will discover the cost through ordinary lost enquiries, not a dramatic event. The agency across the road answers a Dutch message in Dutch at 23:40, using verified inventory, and has a sensible viewing pencilled in before the first laptop opens. It does not need a more charming team. It only needs the operating system to let that team arrive prepared.

Layer 1: the smart CRM, which is where most of the money is

Most of what follows already exists in some commercial product. Compass, for example, added a conversational assistant in July 2026 that updates records, drafts emails, creates searches and suggests owners likely to sell. An international agency needs the same principle fitted to its languages, jurisdictions, offers and inventory. "AI inside the CRM" means that the system understands the work around the record, not merely that a button can rewrite an email.

Intake and routing in five languages

A lead arrives from idealista, Kyero, a referral, a portal, a property fair or the agency site. Before a human sees it, the system reads the complete message, detects its language, looks carefully for an existing relationship and extracts a working brief: source, budget and currency, timing, buyer type, places and constraints. It routes by language, area, expertise and current workload, then drafts an honest first response naming the colleague who will follow up.

In practice, the hard part is identity and exceptions. The same buyer may use a work address, a spouse's phone and a transliterated surname. Aggressive deduplication can merge two people; timid deduplication gives three agents the same buyer. Good intake proposes a match with evidence and sends ambiguous cases to a short human queue. It never hides the merge.

Measure the time to a useful reply, not an instant acknowledgement. Then inspect contact within the first day, reassignment volume, duplicate rate and the share of records with a usable brief. If the first reply arrives quickly but the agent still has to ask which country the buyer means, the clock has flattered the system.

Every conversation becomes structured memory

Calls, messages, emails and the voice note recorded after a viewing become a concise chronology: what the client liked, what killed the property, changes in budget, promises made and who else influences the decision. The next colleague reads a story rather than "called, no answer". In our experience, that is when a CRM stops feeling like management's filing cabinet and starts helping the sales floor.

Memory needs discipline. A summary must distinguish a buyer's statement from an agent's inference, preserve important wording and link back to the source. Corrections should be easy. Without those details, a mistaken summary becomes more dangerous each time another workflow relies on it.

The useful measures are completeness, correction rate and retrieval. Can an agent find the last objection before returning a call? Does the Friday seller report use actual viewing feedback? Track whether staff correct summaries, because silence can mean accuracy or that nobody reads them.

Scoring that reads the conversation, not the click

Old scoring counted opened emails. Context-aware scoring reads intent. A buyer asking about school catchments, flight times and dog rules may deserve a different next step from somebody who downloaded a brochure. The model proposes a score, the evidence and an action. The agent can override all three.

What breaks is false precision. A score of 82 looks scientific even when the CRM lacks financing status or the buyer shares an email address with a partner. Use broad operational categories, show missing information and never let a score quietly restrict service. Fairness and automated-decision controls belong in Layer 3.

Measure whether high-priority records actually become qualified conversations and viewings, then review the overrides. Overrides are not agent disobedience; they are training material. If Italian-speaking agents reverse the model more often, examine the language and routing before blaming the team.

Matching that learns from the viewing

Matching moves beyond bedrooms and price. It can use "too close to the road", "sea visible from the kitchen", "proper garage" and "no tourist rentals in the building". When a Lake Como listing arrives, the CRM identifies relevant mandates, explains each fit and drafts a note in the buyer's language.

The usual failure is stale or embellished inventory. No model can rescue an incorrect availability field, a missing service charge or a view described more generously than the photographs support. Matching should expose uncertainty and the listing's last verification date. The agent chooses who receives a call and checks that the property is genuinely available.

Watch shortlist acceptance, viewing-to-offer movement and the reasons agents reject suggestions. Also look for sameness. A matcher that repeatedly selects the agency's easiest stock rather than the buyer's best fit can be commercially convenient and strategically dreadful.

Nurturing across a buying cycle that can run for years

International buyers often visit, disappear, sell something at home and return with a changed budget. The CRM keeps the relationship warm with material tied to the brief: a verified market note, a suitable development, a price change or a relevant rule update. A French buyer considering Paphos should not receive the generic Crete newsletter because both records say "Mediterranean".

What breaks is frequency without relevance. Once the system can produce text cheaply, teams are tempted to send more of it. Every draft should have a reason to exist, use current consent and recognise what the buyer already knows. A human approves claims and sensitive advice.

Measure days between useful touches, replies and reactivated conversations, not email volume. Review a sample as a client would, in sequence. Three individually plausible messages can still form a repetitive and faintly desperate relationship.

Drafts everywhere, sends by humans

Reply suggestions, viewing itineraries, debriefs, seller updates, offer summaries and price-reduction proposals all follow one pattern: the machine drafts with the full record in front of it, while the person edits and sends. For a seller in Tuscany, the draft can gather viewings and objections without pretending that a model has the standing to recommend a new asking price. The one exception is the narrow, disclosed service reply described under after-hours below, which answers only from verified data within limits the agency has approved in advance.

Approvals must be real, not a decorative button everyone taps. Highlight changed figures, unsupported claims and missing source fields. Measure editing time, correction themes and the rare drafts agents reject completely. Those rejects often reveal a broken source field rather than weak prose.

SLAs, KPIs and the anomalies that matter

The system watches response by language and source, quiet deals, overdue seller updates and unbalanced workloads. It does not send another dashboard. It gives the manager a short list of exceptions with evidence and a suggested action.

For a Gold Coast team, that might be an overseas enquiry assigned during the wrong working hours, a listing with activity but no seller report, and one agent carrying the week's urgent viewings. The measure is whether the exception is resolved and whether it recurs, not whether somebody opened the alert.

After hours, without lying

A late enquiry from Stockholm receives a Swedish answer based on the real listing, area file and process. The system can propose slots from the actual calendar and say clearly that the response is automated. In the morning, the agent sees the summary and the questions that require judgement. Our Aria receptionist and sales agent infrastructure follows that pattern.

After-hours systems fail when access to data outruns permission to speak. They should not invent availability, calculate a tax position or promise that an owner will accept an offer. Voice needs particular care: accent, interruptions, addresses and prices expose weak evaluation quickly. Test each supported language with real agency vocabulary.

Measure the share answered in the buyer's language, useful handovers, bookings later confirmed by a human, correction rates and opt-outs. Open a sample conversation every week. A pleasant voice can conceal a poor answer.

Content, newsletters and campaigns that come from the data

The system knows which segment has gone quiet, what is new for it and which verified market change matters. For an Athens Riviera launch, a campaign can start from the developer's current unit table, payment schedule and actual questions from previous enquiries. An editor with taste still decides what leaves the building.

The concrete test is traceability. Every price, availability claim and rule should point back to a current source. If the draft began with a blank prompt rather than company data, it is merely faster content production.

Pipeline and task automation from reservation to notary

The names of the steps change by country, but every transaction has owners, deadlines, documents and dependencies. In Spain that can include reservation, arras, NIE, checks and notary completion. In Dubai, off-plan files need project and payment documentation. The system creates tasks, chases missing items and escalates silence.

The arras deadline on Thursday while the lawyer has gone quiet is the kind of exception that matters. Automation should reveal it early, never imply that the legal work has been done, and preserve who confirmed each milestone.

Reporting that answers questions

"Which source produced completions rather than enquiries?" "Why did viewings in Provence fall?" "Are our Arabic handovers slower than our English ones?" Managers should be able to ask in ordinary language and receive an answer with the query, fields and period used.

For a Miami brokerage, that can separate Canadian referral business from portal volume without exporting another spreadsheet. The safeguard is reproducibility: if finance or the sales director cannot inspect the calculation, the polished chart is only an opinion wearing a tie.

The numbers that tell you it is working

Measure your own baseline before installing anything. Use the same definition, source and reporting period afterwards. A system that shortens response time while lowering qualified viewings has not improved the operation; it has improved one attractive cell in a spreadsheet.

The practical rhythm matters. Give each measure an owner, inspect exceptions weekly and review the set quarterly. Do not pay people to maximise a proxy they can game. If agents are rewarded only for complete briefs, every record will soon be "complete", including the ones full of guesses.

There is still no honest plug-and-play version of this entire layer. A professional agency has several offers, departments, languages, systems and habits. A product made for everybody can automate what everybody shares, which is the shallow part. The part that changes the numbers is specific to how your company sells. That is why the answer to "which CRM should we buy?" is often "the one you can build into". We compare the options in custom CRM versus off-the-shelf CRM; the smart CRM in Dominium is our own answer for international agencies.

Layer 2: your private AI team, the shift that defines 2027

The smart CRM makes existing work faster. The larger change in 2027 is a different category: company-specific AI agents working as back-office colleagues for the human team, on infrastructure the agency controls, technically or contractually. Think less about a chatbot improvising with buyers and more about a disciplined team behind the screens, preparing work that named people review.

For each person in the office, that team works as a set of personal assistants behind the screen. The agent starts the day with a briefing and a research pack for the afternoon's viewings, the listing manager finds the portal updates already prepared, the compliance officer receives an evidence file that is complete, and the director reads the week's numbers with the anomalies explained. Almost any preparation that happens on a screen can be handed over, within the permissions each assistant holds, so that every person spends more of the day on the decisions and conversations only they can handle, and arrives at them better prepared.

Technically, a private AI team runs on an agentic framework: software that lets AI agents plan a task, split it into steps, use tools and pass work to one another and to people, rather than answering one prompt at a time. In practice, it looks like this. You rent a server in a data centre in your own jurisdiction, EU for European agencies, the UAE or the US for those markets, or install one in your office. The law does not always demand local hosting, but keeping the data at home removes most of the transfer questions before they are asked. An orchestration layer receives requests and chooses the appropriate agent. The runtime gives each agent tools and permissions. A knowledge store holds your processes, inventory, area notes and approved templates. Connectors reach the CRM, portal feeds, document archive, calendars and developer stock lists.

There is no need to make one ideological choice about models. Sensitive document work can run on an open-weight model on your server, while difficult reasoning goes to a frontier model through an API under proper terms. A provider may also offer regional processing: Mistral made regional inference endpoints generally available in August 2026, allowing customers to choose Europe or US processing. Connectors can use the Model Context Protocol, the open standard Anthropic handed to the Agentic AI Foundation in December 2025, so the same governed agent can reach several systems without a fresh integration for every pair. Our October 2026 overview of how businesses use AI covers that model market in more detail.

On top of the infrastructure sits the team. A property specialist knows the inventory and locations. A legal researcher prepares material but never advises. A compliance assistant assembles evidence for the responsible person. A CRM manager can propose a change, but write only when permitted. Roles are useful labels for humans; underneath, they are bundles of skills, knowledge and permissions. Each role needs an escalation rule, and none should talk to a client merely because it can.

Because they share that framework, the agents can work through multi-step jobs together. They pass work to one another, ask a named colleague when information is missing and report back in the channels the team already uses, as the Thursday in the illustrated week below shows. You will hear this set-up called managed agents more and more: agents that run in a managed environment, where memory, tools, permissions, logs and updates are handled centrally rather than by each employee. Anthropic launched Claude Managed Agents, a fully managed service for running AI agents, in public beta in April 2026 and added multiagent orchestration in May, and we expect the term to become common in 2027. For an agency, the useful question is who does the managing: where the agents run, who sets their permissions and who answers when one gets something wrong.

The skill library: what a private AI team actually does

A skill is a repeatable job with a defined input, a defined output and a guardrail. A role such as property specialist, legal researcher or compliance assistant is a bundle of those skills plus permission to read or act in particular systems.

For the sales floor. Agents feel these first, because they change what is on the desk before a viewing rather than what is in a report afterwards.

Property and location research briefs. An agent needs to turn an attractive listing into a decision-ready brief rather than another portal printout. For an Italian buyer considering a two-bedroom home on the Athens Riviera, the skill combines the verified listing record, the agency's past transactions, current portal evidence, official planning and hazard sources, area notes and the buyer's stated use. It returns comparable evidence, running costs, transport and amenity notes, planning questions and an explicit list of unknowns, in the buyer's language, with a source on every number and a clear line where a lawyer, surveyor or tax adviser must take over. Measure it by the share of cited facts that survive human review, and by whether agents fix the knowledge store instead of repairing the same error twice.

Buyer readiness. The window is after qualification and before anybody books flights. A British family preparing to buy in Mallorca receives a personalised preparation pack built from its agreed brief, residence position, financing status and the approved country guidance: financing routes for non-residents, purchase costs and taxes from sourced tables, the likely stages, the documents to gather and items such as the Spanish NIE. The skill prepares and never advises; it does not predict a mortgage approval or turn an estimate into a quote. Success means fewer viewing trips with an unresolved funding problem and fewer document chases after reservation.

Off-market sourcing against buyer mandates. Only a signed, current mandate with clear boundaries switches this on. For a Canadian buyer seeking a mixed-use building in Lisbon, the skill searches the sources the agency may lawfully use, such as auction notices, public filings where permitted, bank stock, developer remainders and the agency's own dormant listings, and returns a ranked research list with the source, fit, likely route of approach and missing evidence. It never pretends an address is available or scrapes around access controls, and the agent approves every owner contact. Track the opportunities agents accept and how promptly the work stops when the mandate changes.

The viewing debrief that writes itself into the CRM. A voice note recorded in the car after a Lake Como appointment becomes structured feedback: reactions by property, objections, changed priorities, promises, next actions and a concise entry for the seller report. With approval, it updates the CRM and feeds matching, nurturing and the weekly vendor report. It separates what the buyer said from what the agent inferred, asks when two statements conflict and never adds a protected characteristic to scoring. The measure that matters is whether the next shortlist reflects the objections instead of repeating them.

For listings and sellers. Sellers judge an agency by what happens between the valuation and the first offer. That is where these skills work.

Listing intake and seller onboarding. The moment is a proposed mandate, before photography and certainly before translation. For a villa in Provence, the skill reads the seller's documents, the title and planning material available through the proper process, energy and diagnostic records, the agency checklist and a recorded walk-through. It returns the country-specific missing-document list, verified listing fields, questions for the owner and descriptions in every required language. Descriptions may have style, but every factual claim must come from verified material: a store room never becomes a bedroom, and a distant blue strip never becomes a sea view. Measure incomplete mandates caught before launch and factual corrections after publication.

Comparable market analysis for the listing presentation. Nobody should ask a seller to agree a price, or revise one, without this having run first. For a house on the Gold Coast, it returns a range supported by named comparables, adjustments the agent can explain, competing and withdrawn stock, evidence gaps and a presentation-ready appendix. Asking prices stay visibly separate from achieved prices, weak comparables are labelled, and thin evidence never becomes false precision. The listing agent owns the recommendation. Monitor how many comparables the valuer accepts and how far the agreed strategy later differs from real market feedback.

Weekly market intelligence and content from real data. A schedule drives this one, after the reporting period closes, rather than the moment somebody needs an empty social post. A Puglia agency receives a short market note, seller talking points, content drafts by audience and language and a source pack for every claim, built from its own CRM activity, verified inventory movements, completed deals it may report and official releases. Small samples are labelled, client confidentiality is protected and the agency's enquiries are never presented as the whole market. A human editor chooses what deserves publication, and the useful measure is whether agents reuse the material in real conversations.

For compliance and legal. None of these skills advises or decides anything. Their job is to make the lawyer's and the compliance officer's file arrive complete, dated and sourced.

Developer and off-plan due diligence. This runs before a project is promoted and again when material facts change. It reads the developer file, licence and project registration evidence, draft contracts, payment schedule and title material, and returns a dated checklist, missing documents and questions for the lawyer, measured against protections that differ by country. In Spain, from the building licence, advance payments must be secured for repayment plus legal interest if construction does not start or finish by the agreed date, and held in a separate account, under Ley 38/1999; Dubai requires project registration and a separate project escrow account under Law No. 8 of 2007; Italy requires a qualifying guarantee covering advances and legal interest if the developer enters a defined crisis, plus a ten-year indemnity policy at transfer, under Decreto legislativo 122/2005; Cyprus requires, for contracts signed from 12 December 2023, a search certificate no more than five working days old under Law 132(I)/2023. It flags, never clears, a project. Measure missing items found before marketing and legal-review corrections.

The AML, KYC, PEP, sanctions and UBO assistant. The agency's jurisdiction and risk policy decide when this starts, and it repeats whenever a party, ownership chain or risk fact changes. For a buyer acquiring a Dubai apartment through a holding company registered in a third country, it returns a document gap list, an identity and beneficial-ownership map, screening evidence, a draft risk assessment, suspicious indicators and a complete audit trail. It does not decide that a client is safe, file a suspicious report on its own or reveal one; the responsible person decides, following the country table in Layer 3. Measure ownership chains resolved, false matches cleared with evidence and file completeness at internal audit.

Contract and clause preparation for the lawyer. Between agreed commercial terms and counsel's first draft, there is a gap this fills. For an offer on a Miami property, it turns the approved deal sheet, party and property details, agreed dates, financing and inspection conditions and the lawyer-approved clause library into one structured instruction pack, with a first draft only where authorised and the unresolved questions in plain language. It cannot invent a clause, choose legal strategy or send a signature version, and jurisdiction, document version and lawyer ownership appear on every page. Success is a lawyer who receives one coherent pack instead of terms scattered through email and WhatsApp.

Translation with a legal terminology memory. Any controlled document, compliance request or client explanation that crosses languages goes through this, for example between a French seller and an Emirati buyer. It translates against a jurisdiction-specific glossary and prior lawyer-approved wording, and returns a parallel translation with defined terms aligned, ambiguities flagged and a change record. It never swaps a legal concept for a convenient but different foreign equivalent, output involving rights or obligations goes to qualified review, and the source version stays controlling unless counsel says otherwise. Measure terminology deviations and whether the same approved term appears in the CRM, the document pack and client messages.

For owners, landlords and investors. Management work is repetitive, deadline-driven and unforgiving of a missed licence renewal, which suits a well-supervised skill better than a tired administrator.

Rentals and property management. This runs across onboarding, tenancy and maintenance, with separate permissions for each stage. A London portfolio receives tenant-screening preparation for a human decision, polite arrears drafts, triaged maintenance tickets, contractor briefs and licence-renewal alerts for short lets where relevant. The skill must not make a solely automated tenancy decision, infer sensitive facts or promise a contractor visit the calendar does not support, and emergencies go straight to a person. Measure unresolved maintenance age, missed licence dates and human overrides in screening, reviewed for fairness rather than treated as errors.

Investor and portfolio reporting. An agreed reporting cut-off, after accounting and property data have reconciled, is the only sensible start. An Indian investor with units around Dubai Marina then receives a consistent portfolio pack: property-level movements, cash-flow explanations, upcoming decisions, evidence links and stated currency assumptions. The skill never fills missing rent with an estimate, hides a vacancy inside an aggregate or turns market commentary into tax advice, and a portfolio manager approves the narrative. Measure reconciliation differences and the time managers spend assembling rather than interpreting the pack.

For the back office. Money and performance are where hidden criteria do the most damage, so these skills run only against definitions management has published.

Commission statements, invoice reconciliation and agent performance. The financial close, and any change of status on a deal or invoice, wakes this up. A brokerage in Orlando receives draft commission statements, unmatched-invoice exceptions, referral liabilities and performance views tied to the definitions management has published. The skill cannot alter a split, approve payment or rank people using hidden criteria; disputed items go to finance, and access is restricted because one agent does not need another's pay. Success is measured by exceptions found before payment and fewer arguments caused by different spreadsheets, and agent performance should connect to the client outcomes in Layer 1 rather than raw activity.

A week with the team

What follows is an illustrative week rather than a case study: the people and times are fictional, while the workflows are the ones we build.

Monday, 07:30. Each agent receives a ten-line briefing in their language: today's viewings in driving order, three unanswered weekend leads, a vendor who has gone quiet in the CRM, an approaching contract deadline, and a note that a developer changed a phase delivery date. The agent reads it beside the first coffee instead of opening six systems.

Tuesday, 12:10. Halfway through a Costa del Sol viewing day, an agent hears that the selection is wrong. The German couple keep returning to road noise and garden privacy. From the car, she records a note asking for quieter options slightly inland, within the existing budget, which listing agents can show soon. The property specialist returns a reasoned shortlist and availability requests. The agent chooses what to show and what the clients receive.

Wednesday, 16:45. Stuck on the A-7 behind a caravan from Groningen, another agent asks the CRM manager to move the Janssen viewing, record their need for a two-car garage and retrieve the remaining penthouse plans. The agent confirms the first two actions but pauses on the third: the live price sits above the budget in the client record. That pause is exactly the point of company memory.

Thursday. An investor in Riyadh asks about a plot near Paphos. The property specialist prepares the commercial brief; the legal researcher compares the supplied title material and planning evidence, then flags a discrepancy for local counsel before anybody books a flight. Later, a resale pre-check finds two missing documents before photography is scheduled.

Friday. International vendors receive draft updates in their own language, built from actual views, enquiries, viewings and feedback themes. Responsible agents approve them. The price review identifies listings whose strategy has drifted away from the available comparable evidence, and a new mandate arrives with its potential buyer matches and reasons already prepared.

Saturday, 22:20. A buyer in Dubai sends a detailed Palm Jumeirah enquiry after the sales team has left. The after-hours assistant says it is automated, answers from verified listing and process data, records the financing question it cannot answer and offers real calendar slots. On Sunday morning, the responsible broker finds the whole thread and one clear task rather than a cold notification.

Nothing in that illustrated week removes the agent. It removes the repeated screen work that makes a good agent late, tired and less attentive in the room.

This is the architecture we build as Dominium, our global real estate system. It is why we give the same answer when an agency asks for an AI: the model is now the inexpensive, replaceable part. The lasting work is the data, connectors, permissions, evaluations and documentation of how the agency actually operates. Those assets remain yours whichever model is fashionable next year.

Layer 3: the guardrails, which decide what the machines may touch

The private team becomes useful because it can see real work: names, negotiations, passports, company structures, viewing feedback and contracts. That is also why governance cannot be a policy pasted onto the project after launch. The first design meeting should decide which data can enter which model, who may trigger each skill, what requires approval and how a client can challenge a consequential decision.

Data protection follows the client and the processing

For EU operations, GDPR requires a binding processor contract and documented instructions, confidentiality, security, audit support and control over subprocessors under Article 28. International transfer rules also follow onward transfers, so a vendor's claim about its front-end location is insufficient. An agency needs the processor list, transfer mechanism, retention period and deletion route. Client files do not belong in consumer chatbot accounts. Put approved tools behind company access, minimise what each skill reads, and keep passports out of a location-research prompt that does not need them.

GDPR also gives people a right not to face certain decisions based solely on automated processing, with safeguards including human intervention and the ability to contest where an exception applies, under Article 22. Lead scoring for internal attention is different from refusing a tenancy or determining creditworthiness. The closer an output gets to money, housing access or another person's rights, the more meaningful the review must be. A button labelled approve is not meaningful if staff are trained to click it without seeing the evidence.

The UK has its own version of this framework. All stages of the Data (Use and Access) Act 2025 are in force, and the rules now permit a wider set of significant solely automated decisions where there is a valid reason and safeguards, although special-category data remains more restricted, according to the ICO's June 2026 summary. That is not permission to score carelessly. Accountability, fairness, transparency and a route to challenge still belong in the workflow.

In the UAE, Federal Decree-Law No. 45 of 2021 provides the federal framework for processing controls, individual rights, security and cross-border transfers, as summarised by the UAE Government. DIFC has a distinct regime whose Regulation 10 expressly covers autonomous and semi-autonomous systems, including AI. In the US, privacy is a state patchwork: IAPP counted 19 enacted comprehensive state laws by January 2026, while California's automated decisionmaking requirements for significant decisions apply from 1 January 2027 under the state regulator's approved rules. The practical response is a processing map by operation and client, not one global checkbox.

Anti-money-laundering duties for estate agents, by country

The period from 2026 to 2027 closes gaps on four continents. Australia has brought property professionals into its regime, the EU's directly applicable rulebook begins, and established regimes from Canada to the UAE continue to put identity, ownership, records and suspicious activity inside the property business. A group selling in several countries cannot copy the Spanish checklist into every branch and call that compliance.

The compliance assistant prepares: it requests documents, checks expiry and consistency, maps ownership, records screening evidence and drafts a risk assessment. The human responsible decides whether the evidence is sufficient, whether enhanced checks are needed and whether a report must be made. Suspicion reporting also requires strict access and non-disclosure. Automation should make a defensible file easier to assemble, never spread a sensitive decision across a general team channel.

Identity and fraud become part of the client experience

Every EU Member State must make at least one EU Digital Identity Wallet available by the end of 2026 under Regulation (EU) 2024/1183. Adoption by agencies and verification providers will take work, but a high-assurance credential offers a better direction than passport photographs bouncing around inboxes. In the UK, conveyancers that meet HM Land Registry's Digital Identity Standard can receive a safe harbour from recourse claims, and the standard uses cryptographic checks, biometric matching and evidence of control over the property or transaction, according to Practice Guide 81.

The threat is already material. The FBI's 2025 figures, as reported by HousingWire, record 12,368 real-estate fraud complaints and 275 million USD in losses, and the same report warns about AI-enhanced schemes including deepfakes and voice cloning. A familiar voice on a rushed call cannot authorise a changed bank account. Agencies need an out-of-band verification routine for payment instructions, identity checks for parties and staff, and a clear client promise: we verify everyone, including ourselves. In practice, that means no changed payment instruction accepted from email alone, no company buyer cleared without understanding who ultimately owns it, and no colleague's identity assumed because a face on a video call looks familiar. Done well, the checks feel calmer rather than more bureaucratic: explain each one, request each document once and show the client a secure status page.

Talking to the public has its own rules

EU AI Act Article 50 transparency duties have applied since 2 August 2026, including disclosure when a person interacts with AI and labelling duties for certain synthetic content, according to the European Commission timeline. The AI Omnibus entered into force on 27 July 2026 and moved Annex III high-risk obligations to 2 December 2027, as the Commission confirms. Most agency CRM, listing and follow-up uses are not high-risk merely because they involve property. Employment decisions and systems evaluating creditworthiness can fall into Annex III under the AI Act. Our EU AI Act guide walks through the duties by role.

Channels add separate constraints. In the US, the FCC has confirmed that AI-generated human voices count as artificial voices under the TCPA and are subject to its restrictions, according to the FCC ruling. In the EU, automated direct-marketing calls without human intervention require prior consent under Article 13 of the ePrivacy Directive. Meta's terms effective 15 January 2026 prevent AI providers from making general-purpose AI the primary function of a WhatsApp Business Solution except where law requires access, including EEA and Brazil numbers; businesses may still use an AI provider for ancillary automation under the WhatsApp Business Solution Terms. That supports qualification and service workflows, not an unbounded general assistant wearing the agency logo.

Scoring and advertising also need a fairness review. HUD says the Fair Housing Act applies when algorithms are used in tenant screening and housing advertising, including practices with an unjustified discriminatory effect, in its AI housing guidance. NAR separately warns brokers about fair-housing, provenance, brand and advertising exposure when agents operate these systems in its agentic AI risk guidance. Remove protected characteristics and proxies where they do not belong, test outcomes, document scoring reasons and provide human review. GDPR Article 22 remains relevant when an automated decision has legal or similarly significant effects.

The permission model is the operating system

Keep a register for every AI agent and skill: owner, purpose, input data, allowed tools, allowed outputs, write permissions, review point, retention and kill switch. Start with read access. Add approved writes only after evaluations show the output is dependable in every language used. Money instructions, legal commitments, suspicious-activity decisions, employment outcomes, tenancy decisions and anything affecting a person's rights require named human sign-off.

Logs should show the evidence read, tool calls made, version used, output produced and person who approved it. Narrow credentials by task rather than giving an orchestration layer the managing director's access. Test ordinary cases, malicious instructions inside documents, stale inventory, mixed languages and unavailable systems. A kill switch should stop the action without erasing the evidence needed to understand what happened.

Finally, separate legal duties from house rules. Disclosure, data rights, AML obligations and calling restrictions come from the applicable law. Our preference for drafts before sends, source links on factual claims and a human on money or rights is deliberately stricter. A team should know which rule is mandatory and which reflects the agency's risk appetite, because both matter and only one changes when the law changes.

AI on the agency website: what buyers should see, and the slop to avoid

The three layers work mostly behind the screen. Clients meet them in front of it: on the website, in the listing, in the chat window and in the video a friend forwards to them. The purpose of AI there is the same as behind it, a better experience for the client and more output from the team. In a business built on relationships and honesty, that means using AI to guide a buyer towards a better purchase, never to mislead them into a quicker one.

Parts of the market are doing the opposite. Merriam-Webster chose slop, meaning low-quality digital content produced in quantity by AI, as its 2025 word of the year, and property listings have their share: skies that are always blue, kitchens that were never fitted, a tired flat rendered as a showroom. A study by the real estate software vendor Coraly found signs of digital alteration in nearly 11% of roughly 40,000 primary listing photos on Zillow, Redfin, Realtor.com and Homes.com in the first quarter of 2026, according to Real Estate News, and frustrated US buyers have started calling it being "housefished". The same report notes that California has required agents to disclose digitally altered images, with a link to the original, since January 2026, that Wisconsin will require disclosure of materially altered marketing images from 2027, and that NAR tells agents to use AI to "show possibilities" in a home, not to "rewrite reality".

Each of those images spends the one asset AI cannot yet create for an agency, its reputation, and the buyer discovers the difference at the front door. The same models can also do things for a buyer that no brochure could: read the whole inventory, answer in the buyer's language at midnight, show what a room could become and explain a neighbourhood with its sources. The difference lies in how they are used, and three rules keep a website on the right side. Whatever is generated says so, on the image rather than in the small print. Whatever is presented as fact links to where it came from. Every promise has a named person behind it.

The first question gets an answer, not a form

Most agency websites still answer a question with a form and a promise to call back. The assistant that replaces the form should know which page the visitor is on, which listings they have opened and which language they are reading in, and answer from the same verified inventory, area notes and approved answers as the smart CRM in Layer 1. It follows the rules of the after-hours replies described there: it says that it is an AI, offers a named person, books into a real calendar, writes the conversation into the client record so nobody asks the same questions twice, and hands over when it does not know.

Search changes in the same way. A buyer can type "three bedrooms, walking distance to an international school, under twenty minutes from Málaga airport, a garden for the dog", as many already do inside ChatGPT, and the agency's own site turns that sentence into filters over its own inventory. Distances should be calculated from the map rather than guessed by the model, and each result should say why it matched, so the buyer corrects the search instead of scrolling past it.

Discovery that learns what the buyer means

Social media learned long ago to show people more of what holds their attention. A property website can do the same for a better reason, because a buyer actually wants help with the decision. If someone asks for a modern villa but keeps returning to stone farmhouses with timber ceilings, the site can start showing similar homes and say why: "because you saved two fincas with exposed beams". The stated brief stays visible, the buyer can correct or switch off what the site has learned, and nothing that matches the brief disappears because an algorithm guessed otherwise. Learn property preferences, never protected characteristics or their proxies; the fairness review in Layer 3 applies to recommendations as much as to scoring.

The same signals help the team. A third visit to one listing or a downloaded floor plan suggests when a follow-up may be welcome, and the conversation tells the agent what to say. A click should never reduce the service a buyer receives, tracking of this kind needs the visitor's consent where the law requires it, and showing buyers why a suggestion appeared lets them judge or correct it.

Listings that show the home as it is, and what it could become

This is where most of the slop lives, and where some of the most useful tools live too. From the same listing photos, a model can now produce a short branded video in the buyer's language, furnish an empty room, restyle a dated kitchen or build a walkable 3D impression of the home. Each can help a buyer understand a property, and each can mislead. Our line follows the NAR phrase: show possibilities, do not rewrite reality.

In practice, the photograph is the evidence and the render is an idea. Keep the original next to every staged or restyled version, label the altered image on the image itself, and never remove what the buyer will find at the door: the crack, the pylon, the motorway, the neighbour's wall. A video made from photos may move the camera, but it must not add a room, a view or a finish. A 3D tour rebuilt from a handful of photos has to invent whatever the camera never saw, so use it for orientation, label it and keep a captured scan as the evidence. Trailers cut from real footage around what a particular buyer cares about, the kitchen for the cook and the garden for the dog, are a better use of the technology than footage of rooms that do not exist. In the EU, Article 50 of the AI Act requires disclosure when AI-generated or manipulated images resemble existing places or objects and would falsely appear authentic, which can include a render of a real home that passes for a photograph; putting the label on the image itself is our stricter house rule.

We think the feature buyers of older homes will value most is a renovation perspective. Let the buyer ask for the kitchen opened up or the terrace covered, and show the illustration next to what the agency actually knows: a cost range from builders it has worked with, whether the community of owners must approve the change, whether the facade is protected or the plot falls under coastal rules, and what the energy certificate says. The render starts the conversation, and the facts make it a better purchase. A buyer who asks to see the view from the main bedroom at sunset should receive the real photograph or scan from that window, with the sun's path for that address and season, which can be calculated, rather than a generated sunset nobody can check.

Facts about the place, with their sources and their uncertainty

Valuations and area data complete the modern listing page. An automated valuation gives a seller or a buyer a first range before anyone visits, and a good one shows the comparables it used and how confident it is. Present it as a range with its evidence, and leave the advice on asking price with the agent, as the drafting rule in Layer 1 already requires.

Neighbourhood forecasts are more tempting and riskier. Heat maps of gentrification, rental yield or climate exposure can help an investor, but a forecast drawn as a map is still a forecast, and a contested one damages trust quickly. In November 2025, Zillow began removing First Street's climate risk data from its listings after the California Regional MLS questioned its accuracy, while Redfin and Realtor.com kept showing it, according to HousingWire. Prefer official sources where they exist, such as national flood maps, strategic noise maps and the municipality's tourist licence rules; show the date of each dataset; and label every forecast as a forecast, with its range. A yield estimate for a Spanish apartment that ignores whether it can hold a tourist licence is worse than no estimate at all.

The rest of the journey, from the client's side

The website is only the first stage. After the enquiry, the client should experience one continuous conversation rather than a series of departments: an itinerary in their language that reflects what they said mattered, a debrief that remembers their objection, an offer summary that explains the process in the country where they are buying, and a secure status page from reservation to signing, with each document requested once. After completion, the same record serves the rental management, the maintenance and, years later, the resale. Clients rarely see the three layers, but they notice when every person they speak to already knows their story. The team notices too: fewer "is it still available?" calls, briefs that arrive structured and media produced from the listing record rather than commissioned for each portal. Hold the website to the same two tests as the layers behind it, a better experience for the client and more output from the team, and measure both with the numbers in Layer 1.

Our real estate AI predictions for 2027, with the evidence behind each

Predictions are cheap. Each of these begins with something that has already happened, then states what we think an agency should do next.

Buyers start inside the assistant, and most agencies are invisible there

ChatGPT had 900 million weekly users by April 2026, according to OpenAI, and Google's AI Mode had passed 1 billion monthly users by May 2026, according to Google. Property supply is moving into those interfaces too, as the idealista, Zillow and Rightmove launches described at the start of this article show.

Attention inside an answer engine behaves differently from a list of links. Pew found that users clicked a traditional result in 8% of visits containing an AI summary, against 15% without one, while a cited source received a click in only 1% of visits, in its study of US Google use. Those figures are broad search behaviour, not property-specific evidence, but the operational lesson is useful: being understood and cited can matter before the website visit.

A FlyDragon benchmark, reported by HousingWire and explicitly a vendor study, tracked 8.2 million AI conversations and found 8.4% of US agents appearing in high-intent generated answers, with the top 1% taking 47% of citations, according to HousingWire. Agencies should publish consistent business facts, answer transaction questions with primary sources, maintain structured listing and organisation data, and make expertise attributable to real people. The aim is to become the clearest reliable source about the places, property types and buying process the agency genuinely knows, rather than to stuff prose for a bot.

Personal AI agents move into glasses and pockets, and start acting for buyers

The next shift concerns who does the searching. Meta's personal AI agent Muse reached number one among free apps on Apple's US App Store in September 2026, a week after launch and ahead of ChatGPT, and Meta says it can research, fill out forms, buy online and book reservations while asking for approval before sensitive actions. At its Connect conference in September 2026, Meta announced VR glasses of about 100 grams, with the processor and battery in a puck that clips to a pocket and its AI agent built into the operating system, on sale from spring 2027 at 1,299.99 USD. The glasses double as a computer with several virtual screens, and the same event introduced Muse Charm, a keychain device for the agent, while Muse is also coming to Meta's smart glasses.

In our view, this is the start of an iPhone moment. The first iPhone was expensive and limited too, and what mattered was the habit it started rather than the device. The agent matters more than the hardware. A buyer who can tell one assistant to find three-bedroom homes near an international school, check the community fees, compare flights and book two viewings will increasingly do exactly that, through whichever device is closest. Picture, as an illustration, a Norwegian couple in 2028: their agent has read the portals and agency sites on the Costa del Sol, set aside the listings whose facts it could not verify and requested two viewings, and in the evening they walk through the shortlisted homes in their glasses from a sofa in Oslo before flying out to see one in person.

A dependable shortlist needs listing facts the agent can read and check, so the practical question is what an agency should expose to a buyer's assistant. Start with verified listing facts, current availability, viewing-request slots, document names and an accountable contact, each with the date it was last checked. Do not expose owner data, access codes, negotiation notes or a write route into the CRM. The connection standards such exchanges would use are gaining traction quickly: Anthropic reported more than 10,000 active public MCP servers and more than 97 million monthly Python and TypeScript SDK downloads when it donated the protocol to the Agentic AI Foundation in December 2025, and the related A2A project said it had backing from more than 150 organisations when it joined the foundation in August 2026. For the glasses, start capturing new instructions with a measured 3D scan and floor plan, beginning with the homes most likely to attract buyers from abroad, and keep the scan as the evidence behind any render, as the website section recommends. The office will not need a headset in 2027, but its listings should be ready for the buyers who own one.

The phone comes back, in the buyer's language

For years, agencies pushed people into forms because multilingual phone coverage did not scale. Voice systems now translate, connect to telephony and call tools. OpenAI's May 2026 models added real-time translation from more than 70 source languages into 13 target languages and SIP calling, while Zillow reported a 95% call-success figure against 69% previously. Those are provider and customer figures, not a promise for another agency.

We expect the phone to regain importance for international enquiries, but only where voice is connected to live inventory, calendars and the client record. It should answer from approved sources, hand over with a transcript and follow the disclosure and consent rules in Layer 3, like every other channel. Fluency makes mistakes sound more convincing: a voice that confidently invents Palm Jumeirah availability is worse than voicemail.

Agents that can use a screen bridge the systems nobody replaced

Developer stock sheets, MLS exports, notary and land-registry portals, property portals without suitable APIs and old back-office screens exist in every market. Computer-use agents can move information through those interfaces where an API project would never be justified. The useful 2027 version is modest: retrieve a permitted document, compare a unit sheet, prepare a CRM update, then stop at the approval point.

These systems need their own accounts and narrow permissions. Every click should be logged, the source file retained and the result reconciled. Anything involving a bank account, signed document, binding submission or deletion stays behind human review. Screen use is also brittle when layouts change, so the success measure is completed, checked work, not the number of animated cursors visible in a demo.

Intelligence gets cheap, while judgement becomes dearer

Using data through 2025 and six performance thresholds, Epoch AI found that the lowest price for a fixed level of language-model performance fell between 9 and 900 times per year, with a midpoint example of 40 times per year for reaching GPT-4 performance on a science benchmark; it also warns that the fastest declines may not continue, in its inference price analysis. The direction still changes agency economics. Reading every permitted call transcript, matching structured records and checking document sets can become routine.

The bottleneck moves to data quality, process ownership and judgement. Somebody must decide which source is authoritative when the developer sheet disagrees with the CRM, how long an old viewing objection remains relevant, and when a compliance flag needs escalation. Owners should budget attention for those decisions rather than obsessing over token prices. Cheap intelligence applied to a confused process produces confusion at volume.

Attackers use AI agents too, and larger agencies answer with security agents

Cheap intelligence is cheap for attackers too. Data protection, verification and anti-money-laundering are the parts of an article like this that owners most often skim, and we understand why, but security deserves a slower read, especially in larger agencies and groups. Layer 3 covers the fraud that reaches the client file: cloned voices, deepfaked video calls and fake identities. The database behind the file is a target in its own right. It holds passports, bank details, company structures and negotiation positions for large transactions, which makes the agency and its clients worth attacking.

The evidence has arrived quickly. Anthropic reported in November 2025 that a group it assessed as Chinese state-sponsored had manipulated its Claude Code tool into attempting intrusions at roughly thirty organisations, succeeding in a small number, with AI performing 80 to 90% of the campaign and making thousands of requests at its peak, often several per second. In April 2026, Anthropic said its Claude Mythos Preview model had found thousands of high-severity vulnerabilities, including some in every major operating system and web browser, and a month later the partners testing it had found more than ten thousand high- or critical-severity flaws in their own software. In September 2026, OpenAI said GPT-6 Astra was its first model to reach the Critical level of cybersecurity capability, able, with the right tools and access, to find unknown flaws and exploit them across many well-protected systems without a person guiding each step. In practice, an attacker no longer needs a team working in shifts. Agents can probe logins, forgotten subdomains, old plugins and exposed databases day and night, and the reports above suggest each model generation makes them more capable.

Our prediction is that enterprises and larger agencies will start deploying security agents of their own in 2027: specialised models that live inside the server, the CRM, the email and user accounts, and the domain and network settings (DNS and CDN), each with narrow, logged permissions like any other agent. They watch around the clock, close routine gaps and escalate larger fixes to a person, re-test the defences each time a stronger model is released, and contain rogue agents, whether an intruder's or the agency's own. OWASP's top ten risks for agentic applications, published in December 2025, already include rogue agents, and the kill switch in the Layer 3 permission register is the manual version of that containment. Providers are building for this market: OpenAI released GPT-5.4-Cyber, a variant tuned for defensive security work, to verified defenders in April 2026, and Fortune reported on 24 September 2026, citing people familiar with the plans, that OpenAI was preparing to preview a GPT-6 Cyber model, with a product to help customers deploy it.

A three-person office does not need a security agent of its own in 2027. It needs two-factor sign-in on every account, no shared passwords, narrow permissions for its AI agents and suppliers who patch quickly. A group with several offices, a large client database and high-value transactions should ask its technology partner two direct questions now: which defences run continuously, and who is alerted at three in the morning? Verifying people properly is half of the answer. Defending with the same technology the attackers use is the other half.

Governance becomes part of the product clients buy

The statute is the smaller part of this story, and Layer 3 covers the legal dates. Clients, banks, developers and relocation partners will ask broader questions: where their data goes, who can act, whether output is logged, how each language is tested and who can stop the system. The permission register in Layer 3 turns those answers into evidence. We think agencies that can show controlled automation will gain trust precisely because the market is filling with tools that cannot explain themselves.

Further out, and honestly speculative

Humanoid robots will not show villas in 2027. Models will keep becoming more capable, and much of their internal reasoning will remain opaque. AI also extends beyond language models: pattern recognition over the agency's own data, forecasting, image and document understanding, and agents that take bounded actions.

Some of the ideas now circulating in property technology will arrive sooner than most agencies expect, and others will not arrive in the form promised. Here is our view of four. The same test applies to every other idea: does it rest on verified facts, and is a named person accountable for what it says?

Simulations of daily life in a home. The promise is that a buyer sees how light moves through the house on a wet Tuesday in November and on a July afternoon, and hears the street at school time and at midnight. The ingredients exist: the sun's path can be calculated for any address and date, weather records are public, many European cities publish official noise maps, and a scan or an architect's model provides the geometry. We expect credible versions first in new developments, where the 3D model already exists, and in resale homes once scans become routine. The honest version separates what is calculated, such as hours of direct sun in December, from what is imagined, such as how a grey afternoon feels, and says which is which.

A home that shows its own future. The promise is a timeline slider that ages the roof, the deck and the heating to 2033 and flags when each will need work. Image models can already describe visible wear in photographs, but they cannot see inside a roof, so the useful version is a maintenance forecast with ranges, built from the survey, the energy certificate and the invoices, rather than an animation of decay. In the EU, the energy part of that forecast gains an official source: the revised buildings directive required member states to introduce a scheme for renovation passports, roadmaps for the staged energy renovation of a building, by 29 May 2026, under Directive (EU) 2024/1275. Agencies that can read a passport, a certificate and a survey and explain them in the buyer's language will gain an advantage long before anyone needs the animation.

Closings without paperwork. The promise is that by 2028 an AI coordinates the notary or title company, runs the fraud checks, pulls the zoning data and drafts a deed the buyer signs in minutes, perhaps on a blockchain. The most concrete step so far came from a registry itself: the Dubai Land Department launched a pilot in May 2025 in which investors buy tokenised shares in ready properties. It concerns fractional investment rather than the sale of a family home, but it points to the pattern we expect: registries adopting new technology themselves rather than being bypassed by it. AI agents will assemble, check and chase the transaction file, and registries will accept more digital steps. Notaries in Spain, conveyancers in England and title and escrow companies in the US still carry legal responsibility for key parts of identity, title and money, and in the EU the single AML rulebook makes estate agents, notaries and other obliged entities responsible for customer due diligence from 10 July 2027, whatever tools they use. We do not expect autonomous closings in those markets this decade. The weeks a transaction takes are mostly mortgages, searches, surveys and people making up their minds, and no model removes those.

Machines negotiating with machines. The promise is that a buyer's AI agent talks to the seller's, compares verified finances with the seller's undisclosed minimum and settles price, conditions and dates in milliseconds. The connection standards described in the personal agents prediction above could carry such exchanges one day, but the minimum is another matter. A seller's bottom line is confidential, and an agency whose system reveals it to a buyer's machine has failed its own client. What we expect within a few years is more modest and more useful: structured offers and counter-offers, with verified proof of funds and conditions and dates in a standard form, exchanged between systems, while people approve anything binding and decide the price they will have to live with.

Our honest bet for 2028 is therefore operational rather than cinematic. Each idea above runs on verified facts, client records and permissions, so the agencies that spend 2027 cleaning those will be able to adopt them, while the agencies collecting disconnected subscriptions will face the same integration problem with better demos. It is not a thrilling forecast, but it is the one most likely to appear in the accounts.

If you run the business: what becomes a commodity, and what remains yours

Everything above is operational. This part is for the owner, the managing director and the board, because the decisions that matter most in 2027 have little to do with which model to use.

Intelligence and digital work are becoming a commodity

The price evidence in the predictions above points one way. A unit of reading, drafting, comparing, translating or reconciling costs less every year, and in our experience each generation of models handles more of it with less correction. Research briefs, comparable analysis, the first draft of a contract pack, the reconciliation of a commission statement, an answer at 23:00 in Dutch: in the agencies we know, these were functions a large firm staffed and a small one skipped. The three-person agency in the Algarve from our introduction can now, with a well-built system, produce research, reporting and follow-up with a consistency it could never have staffed. That levels the field in output. It does not level it in trust.

What AI replaces, in our experience, is the work whose value was knowing something others did not, or moving information between systems by hand: the person who knew the price list, the coordinator who retyped viewing feedback, the assistant who assembled the investor pack from four spreadsheets. In our view it replaces those functions whether the agency prepares or not, and it only replaces the people and the firms that were unprepared, because everyone else moves the same hours towards judgement, relationships and accountability, which is where the fees were always earned.

The moat is your processes, your data, your brand and your reputation

If intelligence can be rented by anyone, it cannot be the advantage. Four things remain yours. Your processes: the written way your firm takes an instruction, qualifies a buyer, runs a viewing and closes a file, including the exceptions, which is exactly what a skill needs before it can run. Your data: years of transactions, viewing feedback, seller relationships and local facts nobody else recorded, which is what turns a rented assistant into your property specialist. Your brand: the reason an assistant names your firm when a buyer asks who to call in Sotogrande. Your reputation: the reason the lawyer answers your email first.

AI improves each of those; it cannot yet create them. That is why we do not believe an agency built from nothing with AI and no data of its own will work: the system has nothing of yours to reason over, and a competitor can rent the same intelligence tomorrow. It is also why an established agency letting AI improve what it already has is the logical next step rather than a leap. The large firm's advantage in output is shrinking; its advantage in data, brand and the ability to run governance is not, provided it uses them. The small firm can now compete on output; it cannot skip the data and process work to get there.

Own the durable parts, rent the rest

A simple decision rule follows. Own the client and property data, clean, lawful and with its history; the workflow definitions and their exceptions; the evaluation set, meaning the questions in every language with approved answers that a new model or supplier must pass; the permission register from Layer 3; the client relationship; and the brand. Rent the models, the telephony, the transcription and the generic tools, and keep every rented part replaceable. A useful test before any contract: if this supplier disappeared on a Friday, what would we still hold on Monday? If the answer is a login and some prompts nobody can export, do not sign yet, whatever the demo looked like. The vendor questions in the "What to build first" section below are the practical version of this rule.

Expect your CRM vendor to ship its own version of several features in this article, if it has not already. That is fine, and often worth using. Treat the features as rented. The asset is the data, the definitions and the evaluations that tell you whether a feature works for your buyers, because those move with you when the vendor changes.

Who owns the client record when agents are independent

In many markets, especially the US and in franchise networks, agents are independent contractors or franchisees. The client record is contested from the first day, and a private AI team makes the question sharper, because the team works for whoever holds the data. Three decisions belong on paper before deployment, not after the first departure. What the brokerage owns, what the agent owns and what the client can take anywhere, remembering that in Europe the client keeps their rights over their own personal data whoever holds the record, brokerage or agent. What leaves with a departing agent and what stays, in language a court and a colleague can both read. And what the agent gets from the system that they cannot get alone: the morning briefing, the research brief, the drafted follow-up, the compliance file that arrives complete. In our experience, agents stay for that third list. A brokerage that only takes data and gives nothing back will teach its best agents to keep a private CRM on their phone, and the AI team will then be reasoning over the leftovers.

The numbers a board should ask for every quarter

The operational measures in Layer 1 belong to the sales manager. A board needs fewer numbers and different ones. We suggest six, each with a named owner and the same definition every quarter.

Do not ask for the token bill or a count of automations. Both tend to rise whether or not the business is improving.

A governance rhythm that fits in a normal month

Governance here is a rhythm rather than a committee. Weekly, the workflow owner clears the failure queue and reads a sample of outputs in each language. Monthly, corrections, complaints and near misses are reviewed, and a channel or permission is switched off if the evidence says so. Quarterly, the evaluation set runs against the current models before any permission is widened, and the board sees the six numbers. Annually, the policy, the processor agreements and the legal register in Layer 3 are checked against what changed in the law, which in 2027 will be a great deal. Someone must own the rhythm. In a large firm that is a named executive; in a smaller one a fractional AI officer can carry it without inventing a new department. An incident routine completes it: who can switch a workflow off at 22:00, who tells the affected clients, and who decides when it comes back.

When the machine decides better than you do

The uncomfortable part comes last. The pattern throughout this article is that the system prepares and a person decides. That will not stay true for every decision. For narrow, repeated commercial decisions where the outcome can be measured, such as the order in which an agent calls the morning's enquiries, which listings enter a shortlist before the agent reviews it, or when a follow-up draft is put in front of the agent, a well-evaluated system will sometimes decide better than a tired manager on a Friday afternoon, and it will do so consistently. Once your own evaluation shows that for a specific decision, continuing to take it by hand is a habit, not a control. Sending, pricing and anything that touches money, rights or an external commitment keep the human approval described in Layers 1 and 3; what moves is the preparation and ordering underneath them.

The firms that delegate those decisions with evidence, keep the accountable human on the ones with legal, financial or reputational weight, and review the boundary every quarter will be faster and cheaper than the firms that keep every decision manual out of pride. We expect the second group to be outcompeted quickly, and to be surprised by it, because nothing in their reports will have told them that a competitor changed how decisions were made. The guardrails in Layer 3 exist to make that boundary explicit, evidenced and reviewable, so that moving it is a decision rather than an accident.

The five levels of AI in a real estate agency

Agencies often compare tools when they should first locate themselves. The levels below describe operating maturity, not prestige. A small brokerage with clean records and measured automation may be further ahead than a national brand whose agents each bought a different assistant.

Two patterns stand out in the agencies we meet. The first is the long stay at Level 1, where the technology has arrived but the context has not: a transcription app, a portal autoresponder and a newsletter tool each work, and the agent has quietly become the integration between them. What moves a firm up is one source of truth and one process with an owner, an exception path and an outcome. The second is the jump from a promising demonstration straight to autonomy. Level 2 is where most agencies make their first serious return, and Level 3 pays off when it starts with read access and drafts and earns its write permissions later.

Level 4 describes a governed business that can explain what happened, rather than a fully autonomous one. Buyers' assistants read accurate inventory through controlled interfaces, staff can inspect logs, leaders know which languages have passed evaluation, and money, legal rights and sensitive decisions keep human sign-off. Few firms need to announce a level; they need to know the next honest move.

What to build first, by the size of your agency

The order matters more than the model. We use five steps because they expose weak foundations before an expensive agent starts making the weakness faster.

1. Name the commercial problem and its measure. Choose something visible in the client journey: slow first response, incomplete briefs, weak viewing feedback, missed seller updates or transaction documents arriving late. Record the baseline before changing the process. "Use AI" is not a business objective.

2. Establish the source of truth. Decide where a contact, property, conversation and deal officially live. Clean the fields needed for the chosen workflow and assign ownership. If a developer stock sheet says available while the CRM says reserved, the first project is data responsibility, not a matching agent.

3. Write the normal process and the exceptions. Follow one live file from enquiry to completion. Include the uncomfortable branches: joint buyers using different channels, an offer in another currency, a lawyer who does not reply, a seller changing instructions and a viewing cancelled after hours. This is the material an implementation team actually needs.

4. Automate one complete loop. Intake through human handover is a loop. A viewing note through updated brief and seller feedback is another. Connect the systems, give somebody the failure queue, and run the workflow with approval before granting write access. A focused 90-day transformation can create this operating discipline without pretending the whole agency changes at once.

5. Review the evidence, then widen permissions. Inspect output in every language, compare the agreed measure with its baseline and read the mistakes. Add the next channel or write permission only when the owner can explain the current one. The boring review meeting is where an AI project becomes an operating capability.

A small agency does not need a miniature enterprise programme. It needs one reliable record and fewer dropped handovers. A growing team usually feels routing pain first: leads cross languages, offices and specialisms, while sellers expect consistent reporting. An enterprise or developer operation must solve architecture and governance early because one permission can reach far more people and data.

The geography matters as much as headcount. A Lisbon team selling only local resales faces a different integration map from a similar-sized brokerage handling US buyers, Algarve developments and UK sellers. A Dubai developer desk may prioritise unit availability and payment schedules. A US brokerage may begin with MLS data and state privacy requirements. Use size to set ambition, then use the actual transaction to set scope.

Before signing anything, ask the vendor or implementation partner these questions and record the answers:

• Where is client data processed and stored, which subprocessors touch it, and can we inspect the processor agreement?
• What happens when the model is wrong, a source is stale or a connector fails outside office hours?
• Who owns the prompts, workflows, evaluation sets and integrations if the relationship ends?
• Which actions can the system take without a person, and where can we inspect the log?
• How is output evaluated in each language and market, and who approves a new version?
• How does the system tell clients when they are dealing with a machine?
• How does cost change as records, messages, calls and model use grow?

If nobody can answer the first three in plain language, the quoted price is not yet the important number. Ask us the same questions. The ROI calculator helps turn saved time and changed conversion into assumptions that can be challenged before anyone buys software.

Why this starts now, and where to begin

Every major technology has forced businesses to adapt. Industrialisation changed how things were made, the internet changed how they were found, and social media changed how people talked about them. AI changes how the work itself gets done, and in our view it will reach further and move much faster than any of them. We wrote this in October 2026, and some of the model names in it will sound old by the time you read it. That is fine, because the model was never the durable part.

Human contact is the gold standard in property today, and in a purchase this large it deserves to be. Behaviour changes all the same. Once personal agents do most of the searching and booking, as we expect they will, and the first walk through a home happens in a pair of glasses, the agencies those agents cannot read or trust will be practically invisible, however good their people are. The agencies they can read will meet the buyer better prepared, with more of their time left for the part that still needs a person.

It starts now, with foundations rather than gadgets. The five steps above are how we would begin, and each leaves something that outlasts today's tools: an AI-enabled, scalable basis that later components can join, a better client experience and more output from the team with every measured workflow, processes, standards and insights written down instead of living in a few heads, and clean data from every enquiry, viewing and transaction, which is what the next model, agent or device will work from. Do not wait for the technology to become simple. The firms best prepared for the next wave will not necessarily have the largest software budget, but they will know where their facts came from, who may change them, what happens when a connector fails and which actions always wait for a person, and they will have used the time to give their people more hours with clients rather than giving clients fewer people.

We build exactly these systems for agencies selling across borders. Everything this article recommends building can run on Dominium, our customised smart real estate CRM with its own Property Manager application: the listings, buyer records, follow-up and transaction control of Layer 1 are in the platform, and we build and configure the private AI team of Layer 2 and the permissions, logs and compliance records of Layer 3 around each agency's markets and way of working. The websites we build add the client side described in this article, connected to the same client record from the first visit. Together they form a basis designed to grow: each new capability, from a buyer's agent requesting viewing slots to whatever device follows the glasses, joins when it is dependable and measured, instead of arriving as one more subscription to switch off. Our real estate industry page shows how we approach that operating model across markets. If you want to know where your agency stands and which workflow should come first, that is what a first conversation with us is for.

Frequently asked questions

Will AI replace real estate agents in 2027?

AI will replace parts of the workload, not the accountable human relationship. It can read an enquiry, structure a brief, propose matches, draft updates and prepare research. The agent still tests motivation, notices the objection nobody typed, checks the property, negotiates and carries responsibility for advice and promises. This matters even more when a buyer is purchasing abroad and dealing with unfamiliar law, language and payment steps. Agencies should measure AI by the qualified time it returns to good people and the improvement clients experience. A plan built mainly around removing agents usually automates trust out of the transaction.

Is it legal to put client details into ChatGPT or another consumer chatbot?

Do not put client files into a consumer chatbot by default. In the EU, GDPR requires suitable processor terms and controls, including the Article 28 processor contract and lawful handling of international transfers described in the Regulation; the UK, the UAE and the US states have their own frameworks, summarised in Layer 3. A business account or private deployment can solve some procurement risks, but not lawful basis, minimisation, retention, access or human review. Map the data and jurisdiction first, then approve the tool and its use. The EU AI Act guide covers separate transparency duties.

What is the difference between an automation, an AI agent and a private AI team?

An automation follows a defined rule, such as creating a CRM contact when a form arrives. An AI agent can interpret unstructured input and choose approved tools to complete a goal, such as reading a viewing debrief and proposing a revised shortlist. A private AI team combines specialist agents, company knowledge and separate permissions on infrastructure chosen by the agency, and its agents, sometimes called managed agents, pass multi-step work to one another. A property specialist may read inventory, while a compliance assistant may prepare a file but never approve it. Most agencies need all three. Fixed automation handles predictable movement, agents handle bounded interpretation, and people retain judgement, approvals and responsibility.

Can an AI agent run our anti-money-laundering checks?

It can prepare most of the file, but it should not own the decision. A compliance assistant can request documents, extract identity details, map beneficial ownership, screen approved sanctions and PEP sources, draft a risk assessment and preserve the evidence; the responsible person evaluates the file, resolves possible matches, decides whether escalation or reporting is required and signs off. Whether your agency must run these checks at all depends on the country. Estate agents are covered in the EU, where the single AML rulebook applies directly from 10 July 2027, and in the UK, the UAE, Australia and Canada, while the US federal residential reporting rule was vacated by a court and is under appeal. The table in Layer 3 links each regime; take local advice for each transaction and date.

We are a small agency with three agents. Where should we start?

Start with one CRM as the reliable home for contacts, properties, conversations and deals. Write down how an enquiry becomes a completion, including duplicate contacts, messages outside office hours and missing documents. Then automate intake, conversation capture and the after-viewing debrief, because those loops improve the record used by everything later. Pick one baseline measure, usually time to a useful reply or complete buyer briefs, and review failures weekly. Do not begin with a multi-agent platform or several disconnected subscriptions. A small team gains more from one dependable workflow than from a broad demonstration nobody owns.

Do we need our own server to run AI agents?

No. You need controlled infrastructure appropriate to the data and jurisdiction, which may be a server in a data centre in your own jurisdiction, EU for European agencies, the UAE or the US for those markets, or a contracted cloud service with suitable processing terms. An on-premises machine is one option, not a badge of seriousness. Decide from data sensitivity, integrations, support, resilience and transfer rules. Keep client work out of consumer accounts, separate each agent's permissions, log actions and require approval for money, rights and external communication. The provider, model and hosting choice can change later if the knowledge, connectors and permission model remain yours.

How do we show up when buyers ask ChatGPT or Google's AI which agency to use?

Publish useful, verifiable answers that assistants can understand and cite. Keep the agency's name, locations, services and contact details consistent. Give important topics one clear page, use structured data where appropriate, identify authorship, state update dates and link claims to primary sources. Make property inventory readable through accurate feeds or controlled interfaces, not copied brochure prose. Reviews across credible platforms and specific local expertise also matter. Avoid creating hundreds of thin pages. A real estate SEO vendor study reported by HousingWire found that only 8.4% of US agents appeared in high-intent AI answers, so machine-readable authority is already a commercial distribution issue, not a future branding exercise.

Will buyers use personal AI agents and VR glasses to find homes?

We expect them to start in 2027. General-purpose personal agents that research, fill in forms and book reservations are already popular, with Meta's Muse reaching number one among free apps on Apple's US App Store in September 2026, and Meta's VR glasses with a built-in AI agent go on sale in spring 2027. In our view, agents will take over much of the searching, comparing and booking first, remote viewings in glasses will follow where good scans exist, and the decision, the negotiation and the signature will stay with people for a long time. To prepare, make inventory, availability and documents readable and verifiable by machines, and offer a named person as soon as a buyer's agent asks for one.