How AI is changing home buying and selling, and where it actually breaks

In short: AI has not replaced the agent. It moved the moment the agent is needed, and it raised what the agent is expected to already know. The transaction did not get automated. The front of it moved into an interface brokerages do not control, and the back of it still breaks in the same place it always did.

Key takeaways

• Buyers now arrive pre briefed. Bank of America's 2026 Homebuyer Insights Report found one in five buyers and owners have used AI in their search. Among Gen Z it is 32%.
• AI is precise exactly where the answer was already known and vague where the money is. Zillow's own accuracy page puts the Zestimate median error at 1.74% for homes already listed and 7.20% for homes that are not. The second number is the one a seller is actually asking about.
• The public stories about AI replacing agents are real, and they are also stories about people who hired a lawyer, paid for a listing service, and got at least one thing legally wrong along the way.
• The operational failure in a brokerage is not the robot. It is the 9pm call nobody answered. In 17 days of one Toronto brokerage's inbound line we logged 167 calls, and 27% of the combined volume landed outside business hours.
• Under TRESA, accountability does not move. CREA's own AI guidance names transparency, accuracy, and accountability, and says a REALTOR® stays fully accountable for advice given. AI does not create a defence.

A call at 9:01pm on a Wednesday

Someone submitted an offer on a listing and wanted the listing agent to call them back. The call came in at 9:01pm. Nobody was in the office. It was answered anyway, the caller's name and number were captured, the intent was logged, and it was in front of the right person the next morning.

We know the exact time because we built the system that answered it. That call is one of 167 we logged over 17 days on a Toronto brokerage's inbound line. We are keeping the client anonymous, but the numbers are real and they are the least interesting part of what they tell you.

That caller had almost certainly already asked an AI what the property was worth, what a reasonable offer looked like, and what happens next. Then they picked up the phone anyway. And they picked it up at 9pm.

That is the actual shape of the change. Nobody got replaced. The moment they were needed moved.

What actually changed

Three things, and only three, are doing real work right now.

Buyers arrive already briefed. Bank of America's 2026 Homebuyer Insights Report found that one in five prospective buyers and current homeowners have used AI tools or chatbots in their process. Among millennials it is 28%, among Gen Z 32%. The most common uses are not glamorous: estimating affordability and carrying costs (57%), general education about the process (55%), and researching neighbourhoods and values (52%). A separate Cotality survey in 2026 found 55% of buyers use generative AI at least monthly, and three quarters assume it is already shaping parts of the transaction.

None of that is a buyer replacing an agent. It is a buyer showing up with the first four questions already answered. The consequence is not fewer agents. It is that the first ten minutes of every conversation are now worthless, and agents who still run those ten minutes sound slow.

The price question got answered before anyone was in the room. This is the one worth being precise about, because the precision is fake in a specific and dangerous way. Zillow reports a median error rate of 1.74% for Zestimates on homes that are actively listed. On off market homes, the number is 7.20%. Within five percent of the eventual sale price, the on market Zestimate lands 85.01% of the time. The off market one lands 37.73% of the time.

Read those two lines together. The model is extremely accurate when a human has already told it the answer by setting a list price, and it is a coin flip when nobody has. Automated valuation is a confidence machine with an accuracy gradient nobody discloses in the interface. We build automated valuation ourselves, so this is not a criticism from outside. It is what the residual looks like when you have seen it.

Discovery moved into an interface the brokerage does not own. When we ran "how is AI changing home buying and selling" through Google's AI Overview from a Canadian IP this month, we got a competent, complete, generic answer citing nobody in particular. Same for "will AI replace real estate agents Ontario." The answer was correct, mentioned TRESA by name, and did not need a single brokerage website to produce it. That is now the top of every funnel in this business, and almost nobody in it is optimizing for being the source that answer is built from.

What did not change, and legally cannot

The two stories everyone in the industry has read this year are Stuart Thompson, a New York Times tech reporter who sold his Hudson Valley home using Gemini and netted about $90,000 more than expected including $36,000 in avoided fees, and Robert Levine, who used ChatGPT to sell in Cooper City, Florida for close to a million and saved $28,000 in commission. Both are real. Both are being passed around as proof the agent is finished.

Read them again and notice what is actually in them. Thompson hired a lawyer for the paperwork. He paid a flat fee listing service to reach the MLS. And Gemini told him to advertise 0% buyer agent commission, which is not permitted. The model was fluent, confident, and wrong, on a compliance point, in a document that goes on the public record. In Levine's case, a broker on the other side of the deal described watching him run positions through ChatGPT and come back with things that simply do not apply in Florida.

The failure mode is not that the model was dumb. It was fluent, and it was fluent about a rule.

In Ontario the boundary is not soft. Trading in real estate is governed by TRESA, and CREA's published AI guidance is short and unambiguous on the point: transparency, accuracy, accountability, and REALTORS® "must remain fully accountable for the information, advice and services they provide to clients." Articles 13 and 15 of the REALTOR® Code cover advertising content and claims, which is where AI written listing copy and AI enhanced photography land. RECO has already told a brokerage that enhanced images need to be disclosed as such in the listing. Discipline runs to $50,000 for a salesperson and $100,000 for a brokerage.

So the honest version of "AI cannot replace the agent" is narrower and more useful than the version everyone repeats. It is not that AI lacks warmth. It is that AI cannot hold a licence, and in this province someone has to.

AI is most convincing precisely at the boundary of what it is allowed to say. That is not a bug you can prompt your way out of. It is a reason to put a named human at that boundary on purpose.

Where it actually breaks

Here is the part we can write and most people commenting on this cannot, because it requires having instrumented a real brokerage's phone line rather than an opinion about one.

Over 17 days on that Toronto brokerage's inbound line, 167 calls. A callback number was captured on 120 of them, which is 72%, close to three in four. Across the combined property management and brokerage volume in the same period, 27% of calls came in after hours or on a weekend. Average call length was about a minute and a half. Zero calls were missed, because the thing answering does not go home.

Now look at what those calls actually were. An agent standing at a property on a Saturday morning whose lockbox code will not work. A buyer's agent at 7:09pm on a Wednesday asking whether a backyard is fully fenced, because their client has small kids. A showing booked at 3:40pm on a Saturday, captured in 38 seconds. An offer notification at 9:01pm.

Not one of those is a job for a large language model's reasoning. Every one of them is a job for something that picks up, understands, captures a number, and routes. That is a plumbing problem, and it is where brokerages are losing deals right now. They are not losing them to a chatbot that writes better listing copy. They are losing them to voicemail.

This is the asymmetry nobody is pricing correctly. AI raised the informational floor at the top of the funnel, so buyers arrive smarter and less patient. It did nothing at all about the operational floor at the bottom, where a lead is still won or lost by whether a human being got the message in time. The gap between those two floors got wider this year, not narrower. Most of the AI spending in this industry is going into the top, which was already fine.

If you want the mechanics of how that plumbing gets built without ripping out the CRM, we wrote that up separately in [real estate AI integrations](/blog/real-estate-ai-integrations).

What we will not build

We get asked for the other version of this. A system whose stated purpose is to run the same volume with fewer licensed people on the payroll. We decline those, and we say why out loud because saying it costs us deals and that is the point of saying it.

AI should give people back their time, not take their jobs. In this specific market that is not a slogan, it is a design constraint with a testable consequence. The systems we ship put the routine intake, the after hours capture, the summarisation, and the routing on the machine, and put the agent back in front of the parts that require judgment and a licence. If a build only works because the headcount came down, we are not the firm for it.

There is also a practical argument. Every one of those 167 calls ended with a human doing something. A system designed to remove the human at the end of that chain does not have anywhere to route to.

What to do this week

If you run a brokerage, five questions. They take an afternoon.

1. Pull your call log for the last 30 days and count the calls that came in outside business hours. Then count how many of those got a callback within 24 hours. That is your real number, and it is almost never the one you would have guessed.
2. Ask your last five buyers what they asked an AI before they called you. Not whether. What. The answers tell you which ten minutes of your process are now dead weight.
3. Ask an AI assistant a question a client in your market would ask, and see who it cites. If it is not you, your competitors are not outranking you. Nobody is being cited, and that seat is open.
4. Find every place AI touches something that goes on the record. Listing copy, enhanced photos, comparative market analyses, anything advertised. Write down who reviews it before it publishes, by name. Under TRESA that person exists whether you have named them or not.
5. Get a written policy for the cases the model gets wrong. Not a plan to reduce them. A policy for what happens when one occurs, who sees it, and how fast. We treat this as a precondition for taking a build, for the reasons in [why most AI pilots die after the demo](/blog/why-most-ai-pilots-die-after-the-demo).

The change in this industry is real and it is not the one on the headlines. The buyer got faster and better informed. The agent got more valuable at a narrower set of moments, and less valuable at the moments AI now covers for free. The brokerages that come out of this ahead will be the ones that noticed the moments moved, and answered the phone at 9pm.

If you want to see what that looks like running in production, the [voice agents page](/voice-agents) has the anonymized call logs the numbers in this piece came from. And if you are trying to work out what any of it is worth before you spend, we wrote down [why we will not give you an ROI percentage](/blog/real-roi-of-ai-implementation).