I write this as the person who decides which AI goes into our product and which does not, so I have a conflict of interest and I would rather declare it at the top. That is also why this post includes what our own AI does not do: it is the only way the part it does do means anything.
The underlying problem is that "AI" stopped describing a technology and became a marketing label. In a demo, the same word can mean a language model drafting an email, a fifteen-year-old business rule with a new name, or something that does not exist yet. Telling them apart is most of the work of evaluating.
What genuinely works today
- Drafting. A first message, a follow-up, a property description. AI writes fast and in any language; the person edits and sends. It saves real minutes every day.
- Summarizing. A long file, a chain of twenty emails, three weeks of notes. Turning a lot of text into five lines is among the things it does best.
- Translating properly, not word by word. In Florida this is not a nicety: it is being able to work the same file in two languages without writing it twice.
- Prioritizing over data that already exists. Who to call today, which deal has sat longest, which lead is cooling. Here the AI does not calculate: it orders what the database already knows, and explains why.
Four uses, told as they happen
- Monday, 7:40. Instead of opening the CRM and staring at four hundred leads, the agent opens a list of seven with the reason beside each: "no reply since the 12th, viewed the same property three times." The seven come from rules, not magic; what the model adds is the ordering and the sentence that explains why.
- Tuesday, after a showing. The agent dictates three sentences into the phone — "liked the kitchen, worried about the condo dues, wants to see two more in the area" — and the note lands written and tidy in the file. Here the AI only transcribes and cleans up; the substance comes from the person who was there.
- Wednesday, an inherited file. Forty emails from an agent who left. A five-line summary saves forty minutes and avoids the awkward call asking the client something they already answered twice.
- Thursday, one client who writes in English and another in Spanish. The same follow-up goes out in both languages without writing it twice, and without the second one reading like a phrasebook.
What gets promised and does not yet deliver
Predicting which neighbor will list in the next six months. Propensity models exist and some carry signal, but the precision implied in demos is not in the results: what gets sold is anecdotal hits, not rates. If a vendor tells you their AI predicts sellers, ask for the hit rate across the entire list they handed you, not three success stories.
Predicting the market. No model knows what rates or prices will do next year. When a product implies otherwise, it is selling the illusion of certainty about the one thing this industry already knows is uncertain.
Replacing the agent in the hard conversation. AI does not negotiate a price reduction with a seller emotionally tied to their house, and it does not hold a buyer who has lost three offers. That part of the job — the part that justifies the commission — is still human, and will be for a good while.
The risks that deserve to be taken seriously
The first is compliance, and it allows no nuance: fair housing. An AI that drafts listings or filters clients can reproduce biases from the text it was trained on, and describing a neighborhood by who lives in it rather than by what it has is exactly what the law prohibits. Every generated text meant for publication gets read first with that question in mind, and the legal responsibility sits with the brokerage, never with the vendor.
The second is that these models invent things with total composure. A closing date, a statutory deadline, a file number: if the fact matters, it gets verified at the source before it is sent. We apply that rule even to the posts on this blog, which is why they carry links to statutes and reports instead of loose figures.
The third is your clients’ data. Before pasting a file into any tool, the question is where that text gets processed, whether it is used to train models, and who answers if it leaks. If the vendor does not have it in writing, you already have your answer.
Five questions to evaluate any AI
- What exactly does this feature do when the AI is unavailable? If the screen goes blank, the AI was not an aid: it was the product.
- What is computed from data and what is generated by the model? A temperature score, a count or a date should come from the database. If a model invents them, they will not match twice in a row.
- Does it send anything without a person approving it? An email that goes out on its own to a client is a risk with your name on it.
- What does using it cost and what happens if I go over? Metered usage, a clear cap and a warning before you hit it. "Unlimited" is always paid by somebody.
- Does it touch the money? Commissions, escrow and reconciliation are arithmetic, not narrative. AI has no business there.
What ours does, and what it does not
Our AI does three things: it builds the plan for the day by choosing among candidates the platform computes without AI — overdue tasks, cooling leads, stalled deals — and explains why it picked each one; it leaves the WhatsApp message and the email already drafted in the client’s language, for a person to review and send; and it turns objective behavior signals into a one-sentence read on the client.
And it has four rules we would rather state here than have you discover them: it never sends anything on its own — the last click always belongs to a person; its usage is metered call by call and capped by plan, so there are no surprise invoices; it never touches the money — no AI over escrow or reconciliation; and whatever can be calculated is calculated without AI. Temperature scores, counts and dates come out of the database; the AI only puts the narrative on top. If it ever runs out of quota or provider, the screen keeps working with less text and the same numbers.
That last rule is the most boring one and the one that defines the product most. A brokerage cannot afford its figures depending on whether a model had a good day. The numbers have to add up every time; the AI is there to save you the writing, not to tell you what you billed.