Most of the argument about AI in real estate is an argument about capability: what the tools can and cannot do this month. That framing is useless in practice — it changes every few weeks and it never tells you what to do on a Tuesday with four listings to write up.
Exposure is the better frame, and it is the one we build against. Some work costs you nothing when it comes back wrong. Some work costs you a client, a complaint, or a licence. Sort by that.
The three questions that sort the work
Before handing a task over, we run it through three questions:
- Would you have edited this anyway? If yes, a machine-written first draft costs you nothing. Listing copy, a nurture sequence, a caption — you were going to rewrite them regardless, so starting from something beats starting from nothing.
- Is the output reversible? A bad draft is deleted in a second. A price that has already anchored a seller's expectations is not so easily undone.
- Will someone treat it as advice? The moment output leaves your hands and a client reads it as your professional opinion, it is your professional opinion. The tool does not carry your licence.
Everything below follows from those three.
What to hand an AI
First-draft listing copy
Writing the fourth description of the week is a chore, and the blank page is most of the cost. Editing is faster than generating, so start with something to react to.
Two conditions. First, feed it real detail — the actual feature sheet, the actual measurements, the actual reason the layout works — because a model given thin input fills the gap with plausible fiction. "Hardwood throughout" and "recently renovated" have a way of appearing from nowhere, and misrepresentation in a listing is your problem, not the tool's. Second, read every line against what you have personally verified.
Then watch the language drift. Models trained mostly on English-language web text, where American property marketing dominates, will happily produce copy that describes the buyer rather than the property — "ideal for professionals," "perfect for a young family." Provincial and territorial human rights legislation does not care that a machine wrote it. Describe the house, never the person you imagine living in it.
Summarising comparables you have already chosen
There is a sharp edge here and it matters which side of it you are on. Selecting comparables is judgement — it is the part of a CMA you are actually paid for. Deciding that the sale two streets over does not count because it backs onto the arterial, or that a January close reflects a market that no longer exists, is expertise.
Once the set is chosen, what remains is clerical. Turning eight sales into a paragraph a seller can follow, noting that six of them closed under list, laying out the spread in price per square foot — that is summarising, and machines summarise well. Keep the selection. Hand over the write-up. That is the boundary our listings and CMA tools are built around.
First-pass email and follow-up sequences
A five-touch sequence is a structure problem before it is a writing problem, and structure is what these tools produce quickly. Getting the shape of a follow-up campaign out of your head and onto a page is worth real time, even if you rewrite every sentence afterwards. That first pass is what our email campaign tools are for.
Then check compliance yourself. Canada's Anti-Spam Legislation (CASL) applies to commercial email no matter who or what drafted it, and it generally expects consent to send, clear identification of the sender, and a working unsubscribe mechanism. What counts as consent, and which messages are treated differently, is not something to work out from a chat window — check your brokerage's guidance or the regulator's own material. A model will cheerfully write you a sequence with none of it, because it has no idea it is operating in Canada.
Repurposing one piece of work across channels
This is where the hours actually go. You wrote the market update; now it needs to be an email, three social posts, a caption, a short video script, and a block on a listing page. None of that is new thinking — it is the same thinking in six shapes. Reformatting is one of the highest-return, lowest-risk uses of AI in this business, precisely because the substance was yours before the tool ever saw it. That is why social content was one of the first places we pointed it.
What to keep human
Pricing judgement
An automated valuation is a starting number derived from what a dataset can see: beds, baths, square footage, close dates. It cannot see that the basement suite is not legal, that the roof is at the end of its life, or that the two strongest sales on the street were pre-list deals that never reflected open-market demand. It also cannot read the seller across the table who has to be walked from an expectation down to a number.
Use model output as a sanity check on your own figure. Never present it as the figure. And think hard before showing a seller an AI-generated price at all — once a number is on the table it anchors everything after it, and you will spend the rest of the appointment negotiating against a figure you did not choose.
Disclosure and compliance wording
Anything with a regulator behind it stays with you: disclosure of known latent defects, representation and agency agreements, the wording of conditions, client identification under FINTRAC, anything touching a trust account. These are not writing tasks that happen to be regulated. They are regulated instruments that happen to be made of words.
The specific hazard is that this is exactly where AI is most convincing and least reliable. It will produce a clause that reads precisely like a clause. It will use the right vocabulary and the right cadence. It will not know that your province revised the form, that your brokerage has its own required wording, or that the phrasing it borrowed comes from a jurisdiction with different obligations entirely. Fluency is not authority. Use your brokerage's forms and your regulator's language. The same caution applies to the client information you feed a tool in the first place: what may leave your systems is a question for your brokerage's policy, not one to decide case by case at your desk.
Anything a client will treat as advice
The line here is not the subject, it is the reception. The same sentence about a condition period is background information in a blog post and professional advice in a text to a client at nine at night. If someone will act on it, own it — read it, verify it, and be ready to defend it, because you are the one who will be asked to.
The failure mode that catches agents
It is not obvious nonsense. Obvious nonsense is easy to spot and delete. The dangerous output is plausible, specific, well-written and wrong.
Ask a general-purpose model for a school catchment, a zoning designation, a strata or condo bylaw on rentals, a land transfer tax rule, or how long a comparable sat on market, and it will often answer with total confidence and no source. Sometimes it will be right. You cannot tell which times from the output alone, because the wrong answers arrive in the same assured voice as the correct ones.
We treat every fact that comes out of a model as unverified until it has been checked against the source you would have used anyway — the board, the municipality, the strata documents, the land title. If verifying takes longer than looking it up would have, do not use the model for that task at all.
How to set this up this week
- Write the not-list first. Decide in advance which work never goes to a machine. Pricing, disclosure, negotiating strategy and anything read as advice is a sensible starting set.
- Give it your real inputs. Vague prompts are what produce invented detail. Paste the actual feature sheet and the actual comparable set. Client notes are a different question — check what your brokerage's policy allows before anything that identifies a client goes into a general-purpose tool.
- Feed it your own writing as the example. Generic output is usually a prompt problem, not a ceiling. Two or three pieces you wrote yourself will move the voice further than any amount of instruction.
- Verify every number, every time. No exceptions, including the ones that look obviously right.
- Time the task before and after. If it is not saving real minutes on a specific job, stop using it for that job.
The agents getting the most out of this are not the ones using it for the most things. They are the ones who decided early and clearly which work is theirs. That decision is what makes everything else safe to automate.