How Real Estate Firms Should Use AI for Sales and Lead Generation

High adoption and low measured impact

The National Association of Realtors surveyed its members on technology in 2025 and found that 68 percent had used AI tools in some form: 20 percent daily, 22 percent weekly, 27 percent a few times a month. That is the fastest adoption of any tool category I can remember in the industry. The next number is the one worth sitting with. Asked what the AI had done for their business, 46 percent said no noticeable difference. Only 17 percent reported a significant positive impact.

The pattern behind those two numbers is visible in the same survey. The most common AI use was generating content, mostly listing descriptions, at 46 percent of respondents. The share using a chatbot for lead capture or client communication was 7 percent. The share using a CRM with AI-driven insights was 21 percent. Agents adopted the use that was easiest to start and hardest to measure, and skipped the two uses where the arithmetic is straightforward.

McKinsey Global Institute estimated in 2023 that generative AI could add US$110 to 180 billion of value to commercial real estate, on the condition that the industry changed how it worked to capture it. The NAR numbers suggest that condition is mostly unmet. I see the same shape in other sectors. The first use of a new tool is the one that feels like the tool, and for a language model that is writing. Writing a listing faster saves an agent twenty minutes a week. It does not bring a single extra buyer to the door, because every competing agent has the same model and the same twenty minutes. The uses that move revenue are less glamorous. They are about who gets called, how fast, and in what order.

The first hour after an enquiry is where the money is

The most useful study on lead response is fifteen years old and was run across financial services, cars, software and education, with no property firms in the sample. James Oldroyd, Kristina McElheran and David Elkington audited 2,241 US companies by sending each a web enquiry and timing the reply. Thirty-seven percent responded within an hour. Twenty-four percent took more than a day. Twenty-three percent never responded. Among the companies that did respond within a month, the average was 42 hours.

Their second study, across 1.25 million leads at 42 companies, measured what the delay cost. Firms that attempted contact within an hour were nearly seven times as likely to have a meaningful conversation with a decision maker as firms that tried even an hour later, and more than sixty times as likely as firms that waited a day. The reasons the authors found for the delay will be familiar to anyone who has run a sales floor: leads pulled from the CRM once a day instead of continuously, agents focused on their own prospecting over inbound, and distribution rules built around fairness to the team at the expense of speed to the buyer.

Property buyers behave the same way as the buyers in that study, with one difference that makes it worse. A buyer who enquires on a listing at 9pm on a Sunday has usually enquired on four other listings in the same sitting. The agent who replies first gets the viewing. The agent who replies on Monday afternoon gets a polite “already viewed, thanks.”

This is the job a well-built AI first-responder does. It acknowledges the enquiry within a minute, on the channel the buyer used, and asks the three or four questions an experienced agent would ask before booking a viewing: timeline, financing status, must-haves, whether they have a property to sell first. It offers viewing slots from the agent’s real calendar. It then hands the conversation to a named agent with a summary, and the agent’s target is a personal call inside the hour. In Singapore that conversation happens on WhatsApp more often than by email, and the tooling now supports that channel natively.

Two rules keep this safe. The AI introduces itself as an assistant and never claims to be the agent. And it does not quote prices, availability, or anything about the property that is not in the listing record. When it does not know, it says a person will come back with the answer. The first rule is a matter of trust and, in several jurisdictions, disclosure law. The second is what stops a helpful-sounding model from inventing a balcony.

Free tool

Agent Risk Assessment Matrix

Place a first-response agent on the autonomy and consequence grid and get the controls it needs before it talks to a buyer, including what it may and may not say unsupervised.

→

Lead scoring should run on behaviour

The second measurable use is deciding who to call first. Most agencies score leads on what the buyer declared: budget band, preferred district, bedroom count, a self-reported timeline. Those fields are filled in once, often optimistically, and never updated.

Behaviour is a better signal, and the portals already collect it. Zillow’s own analysis of two years of US sales found that listings saved by five or more users a day typically went under offer within a week, and listings saved by fewer than one a day sat for over a month. That is a listing-side statistic published by a portal with a product to sell, so I grade it accordingly below, but the direction matches what every experienced agent knows from the floor: the buyer who saves a listing, comes back to it three times, shares it with someone and asks about the maintenance fee is closer to a transaction than the buyer who filled in a form with a large budget and never returned.

A scoring model that reads those signals, from the portal, the agency website and the CRM together, does three things an agent cannot do by hand. It ranks the day’s call list. It flags the lead that went quiet after high engagement, which is usually a buyer who found something elsewhere and is still worth a call. And it notices the seller-side signal that agents miss most often: a past client who has started browsing again, which in a market where sellers hold a home for a median of eleven years before selling is the earliest sign of a listing appointment.

The model does not need to be sophisticated. A weighted score on five or six behavioural events, recalibrated monthly against which leads converted, beats most of what agencies buy off the shelf. The discipline that matters is feeding it real outcomes. If nobody records which leads became viewings and which viewings became offers, the score is a guess dressed as a number.

Listing content carries the most regulatory exposure

Drafting is where most agents started and it does save time. Property descriptions, social captions, the twelve variants of an email for different buyer segments, a translation of the listing into Mandarin or Bahasa: all of that is faster with a model and none of it needs to be worse.

The risk is that the model writes what a good listing usually says, and this property may not have it. The typical failures are a draft that describes a balcony the unit does not have, a “recently renovated” kitchen that was last touched in 2009, and an MRT station that is a 25-minute walk described as “steps away.” None of that is malice. The model is pattern-matching to thousands of listings, and the pattern includes features this unit lacks.

The regulator does not care which tool wrote it. In Singapore, the Council for Estate Agencies’ code of ethics requires marketing materials to accurately reflect the property, and the CEA has said publicly that agents must provide clear disclaimers when AI is used to alter or enhance images or videos, with enforcement action for those who do not. It is reviewing its regulations to give further guidance on AI in advertising. Every ad still needs the salesperson’s registration number and the agency name. The same logic runs through the advertising codes in most markets I work in: the licensed person is responsible for the claim, regardless of who or what drafted it.

Image enhancement deserves its own line because it is where the technology is most tempting. Virtual staging, sky replacement, decluttering and twilight conversion all sell a property harder. Used with a disclosure they are fine. Used without one they are a misrepresentation with a timestamp, and the buyer who turns up to a viewing and finds a different room is the complaint that reaches the regulator.

The control is simple. One person who has physically seen the property signs off every description and every enhanced image before it goes live, and the sign-off is recorded. That takes five minutes a listing and removes most of the exposure.

Targeting and outreach need a fairness and consent check

The third place AI shows up in real estate marketing is deciding who sees the ad and who receives the follow-up. This is where the legal exposure is largest and least understood.

In May 2024 the US Department of Housing and Urban Development issued guidance making explicit that the Fair Housing Act applies to housing advertising delivered through algorithmic targeting, whether or not the exclusion was intentional. The examples HUD gave are exactly the optimisations an ad platform performs by default: an audience model trained on who responded before ends up excluding families with children, particular language groups, or residents of particular neighbourhoods, because the historical data was shaped that way. That guidance now sits in HUD’s archive after a change of administration, but the statute it interpreted has not changed, and Meta settled a Justice Department lawsuit on the same point in 2022.

Singapore does not have a fair housing statute in that form, but the same mechanism appears through two other doors. The Personal Data Protection Act governs what data an agency can use to profile a buyer and requires consent for marketing messages, and the Do Not Call Registry applies to every telemarketing message sent to a Singapore number, including messages an AI drafts and schedules. An outreach agent that sends a “just checking in” WhatsApp to a number on the registry has committed the breach, and the agency’s name is on it.

The practical checks are these. No protected characteristic, and no proxy for one, goes into the targeting model; postcode and language are the two proxies that slip through most often. Every outreach list passes through the DNC check before an AI sends anything. And the AI does not send a first message to anyone who has not enquired, because cold outreach is where consent is thinnest and complaints are most likely.

How far each sales job can be handed to AI

The pattern across the four uses above is the same one I apply to agentic AI in any sector: autonomy should rise with reversibility. A draft can be undone. A message to a buyer cannot. The table below is how I would set the boundaries for an agency starting now.

Sales job	What the AI does	Where the human stays	Autonomy
First response to an enquiry	Acknowledges within a minute, qualifies on four questions, offers real calendar slots, hands over with a summary	Personal call inside the hour; any question about price, availability or the property beyond the listing record	Acts, within a script
Lead scoring and call order	Ranks the day’s list on behavioural signals, flags re-engaged past clients and gone-quiet high-intent leads	Decides who to call and what to say; records outcomes so the score keeps learning	Recommends
Listing copy and translations	Drafts descriptions, captions, segment variants, translations from the listing record	A person who has seen the property signs off every draft; sign-off is recorded	Drafts only
Image enhancement and staging	Virtual staging, decluttering, lighting	Disclosure on every enhanced image; original photo retained	Drafts only
Follow-up and nurture sequences	Drafts personalised follow-ups after a viewing, schedules the cadence	Approves each sequence; DNC and consent check runs before any send	Drafts, sends on approval
Ad audience targeting	Proposes segments from engagement data	Removes protected characteristics and their proxies; reviews audience composition monthly	Recommends

”Acts, within a script” means the AI may complete the interaction unsupervised but only along a bounded conversation it cannot leave. “Recommends” means a person makes the call every time. “Drafts only” means nothing reaches a buyer without a named approval.

Free tool

AI Use Case Prioritisation Matrix

Score the six jobs in the table above, or your own list, on value and readiness and get a sequence for the next two quarters instead of trying all six at once.

→

What to measure in the first ninety days

An agency that does the first-response and lead-scoring work properly will see the result in four numbers, and none of them are “hours saved.”

Median time to first response, measured from the enquiry timestamp to the first human-or-AI reply, with the target under five minutes and the human follow-up under an hour. Contact rate, the share of enquiries that became a two-way conversation, which is the number the Harvard study was measuring. Viewing-booked rate per hundred enquiries. And offers per hundred viewings, which is the check that the AI is qualifying properly and not filling the calendar with tyre-kickers.

Alongside those four, one number that must not move: complaints about listing accuracy or unsolicited contact. If that rises, the programme is buying speed with the agency’s licence, and it should stop until the sign-off and consent checks are fixed.

Where to start this quarter

Pull the last ninety days of enquiries out of the CRM and the portals and measure one thing: median time from enquiry to first human reply. Few agencies have ever looked. When they do, the number is usually measured in hours, sometimes days, with a long tail that was never answered. That single figure is the business case for everything above.

Then put a first-responder on one channel for one team, with the script bounded to the listing record and the hand-off target set at one hour. Run it for a month against the enquiries the other teams still handle by hand, and compare contact rate and viewings booked. If the lift is there, expand to the next channel. Leave the content drafting running in the background; it is useful, and it was never the point.

---

If your agency or sales team is deciding where AI goes first, the ninety-day measurement above is the conversation I usually start with. My consulting work covers scoping that first deployment and building the sign-off and consent controls around it.