What an AI assistant can and can't do for a real estate agent

The pitch for AI in real estate has been the same for three years and it is not quite right in either direction. The optimistic version says the software will handle your business. The dismissive version says it is autocomplete with a sales team. Neither survives contact with a working week.

Here is a more useful frame: AI is good at tasks with a known shape and a forgiving failure mode, and bad at tasks that require judgement, carry legal weight, or depend on knowing things it was never told. Almost everything in this business sorts cleanly on those two axes once you look at it that way.

What it genuinely does well

Answering the phone when you can't

This is the clearest win in the category, and the reason is timing rather than intelligence. A lead who calls and reaches voicemail is a lead who calls the next agent. The value is not that the assistant handles the call better than you would — it does not. The value is that it handles the call at all, at 8pm on a Saturday, while you are at a showing.

What it reliably does: picks up, finds out what the person is calling about, captures the details, answers straightforward questions about a listing, and books a time with you. What it does not do is close anybody. Treat it as the thing that stops a lead going cold, not as a salesperson.

First drafts of anything written

Listing descriptions, follow-up emails, social posts, newsletter sections, a script for a video. AI produces a competent first draft in seconds and a finished piece never. The productive pattern is to let it do the blank-page work and then rewrite it in your own voice, which is faster than writing from nothing and produces something that sounds like a person.

Two cautions, and the second one is serious. First, unedited AI copy reads like unedited AI copy, and buyers have learned to spot it. Second, listing copy is legally regulated text — the Fair Housing post covers what that means in practice, and it means a human reads every word before it is published.

Summarising and extracting

Long email threads, a stack of notes, a document you need the key dates out of. Pulling structure out of unstructured text is the single most reliable thing this technology does. It is unglamorous and it saves real time.

Remembering to follow up

The highest-value thing most agents could do is follow up more consistently, and it fails for a boring reason: nobody can hold two hundred relationships and their next touch in their head. Software that surfaces who is due and drafts the opening line converts an intention into an action. The model is not doing anything clever here. It is doing something you were going to do and didn't.

Being asked questions about your own data

"Which of my buyers have been looking for six months?" "What closed in this neighbourhood this year?" Natural language over your own records is genuinely useful, with one rule attached: verify anything you are about to act on or repeat to a client. Which brings us to the failures.

What it cannot do

Anything requiring judgement about a person

Whether this buyer is serious. Whether this seller is ready to hear the number. Whether the silence after your last message means think-about-it or gone. That is the job, and none of it is in the data.

Pricing

AI can assemble comparables, pull history and produce a range. It cannot walk the property. It does not know that the photos were taken before the roof failed, or that the comparable two streets over backs onto something that costs it real money. A number produced without having seen the house is an input to your opinion, not a substitute for it.

Anything with legal weight

Contract terms, disclosure obligations, advice about what a clause means, what to do about a defect. This is not squeamishness. It is that the error mode is silent — a wrong answer arrives in the same confident tone as a right one — and the consequence lands on your licence.

Negotiating

It has no read on the other side, no relationship with the other agent, no sense of what is actually being traded. It can draft what you decide to say. Deciding is the part that is yours.

Knowing anything it was not told

The model does not know your market unless your market is in the data it was given. It does not know your brokerage's policy, your MLS's rules, or that this seller's sister is the actual decision-maker. Most disappointing AI output is not a failure of intelligence — it is a system answering a question with less context than a new assistant would have had on day one.

The three failure modes to know before you buy

Confident wrongness. The single most important thing to understand. When these systems are wrong they are not hesitant. There is no tone shift, no hedge, nothing that flags the answer as invented. The defence is procedural rather than technical: anything you will repeat to a client or act on, you verify. Build that habit before you need it.

Silent drift. An assistant that worked in March can be worse in September — a configuration changed, an integration broke, a model was updated. Nothing announces it. The practical answer is to actually listen to some of your own calls and read some of your own drafts, monthly. Sampling your own output is unfashionable and it is how you find this.

Doing the wrong thing efficiently. Automation multiplies whatever it is pointed at. Aimed at a poor follow-up sequence, it sends a poor sequence to everybody, faster. Get the process right manually first. Automating a bad process is how a small problem becomes a reputation.

How to think about the decision

The question worth asking is not "is this AI any good." It is "what specifically stops happening in my business if this works, and what does it cost me if it is wrong?"

Missed calls are a good candidate: the thing that stops happening is leads going cold, and the cost of a mediocre answered call is lower than the cost of no answer. First drafts are a good candidate, because a human reviews the output anyway. Pricing advice is a poor candidate, because the cost of being wrong is high and lands on you.

Run that test on any tool you are shown, including ours. It is a better filter than any demo, and it will tell you quickly whether a product is solving a real problem or performing one.