Anonymised, illustrative composite. A brokerage website chatbot answered a visitor's question about walking away from a firm deal — confidently, definitively, and incompletely.
At a glance
A brokerage had deployed a website chatbot, a general-purpose AI tool, to answer visitor questions about listings, showings and the buying process around the clock. It handled routine questions well — hours, availability, neighbourhood basics — and nobody had configured it to refuse or defer on anything more specific.
A visitor already under a firm agreement of purchase and sale, deposit paid, no conditions left outstanding, typed a direct question into the chat window: could they still walk away without losing their deposit. The chatbot answered yes, in general terms, without asking whether the agreement was still conditional or already firm — a distinction that decides the entire answer. The visitor took that answer at face value and relayed it to their agent as though it settled the question.
The deposit at issue was $22,000. The agent, trusting the client's account of what they'd read online, began drafting a mutual release on the assumption the buyer could exit cleanly. The file only got a second look because the brokerage runs a routine errors-and-omissions file review on any in-progress termination before a release goes out for signature — standard practice, not a response to a complaint. That review is what surfaced the chatbot transcript and the gap in what it had actually told the client — exactly the kind of file an errors-and-omissions audit exists to catch before it becomes a claim.
The transcript itself was three exchanges long. The visitor asked the deposit question once, got the unqualified yes, and never asked a follow-up — there was no second question about conditions, financing, or what "firm" meant in their own agreement, because the first answer had already sounded complete. That is the ordinary way a short, confident chat exchange becomes the basis for a five-figure decision: nobody involved treated it as a first pass to be checked, because nothing about the exchange signalled that it needed to be.
RECO Bulletin 1.1 states the competence duty in plain terms: where a registrant — and by extension, a tool the brokerage puts in front of clients on the registrant's behalf — cannot competently answer a question, the rule is to “advise the client… to obtain services from another person,” not to answer anyway, the same principle a confidently wrong AI draft runs into elsewhere in this hub. CREA's own AI guidance closes the loop on accountability: “the adoption of AI does not alleviate the professional responsibilities of REALTORS®. REALTORS® must remain fully accountable for the information, advice and services they provide to clients.” A chatbot's answer, once a client acts on it, is the brokerage's answer.
The E&O review flagged the transcript before any release or mutual agreement was signed. The agent stopped drafting, went back to the client directly, and explained that whether they could walk away without losing the deposit turned on facts the chatbot never asked about — the client was connected to a real estate lawyer the same week to review the actual agreement before anything else moved. No deposit was lost, and no release was signed on the wrong premise. The brokerage reconfigured the chatbot the same month to hard-refuse any question that reads as a specific legal question about a live transaction, routing those to “speak with your agent or a real estate lawyer” instead of attempting an answer.
Had the E&O review not caught the transcript before the release went out, the realistic path was a signed mutual release the buyer had no legal basis for, a seller who could reasonably argue the deal was still firm, and a $22,000 deposit dispute layered on top of a transaction that should have simply closed. Unwinding a signed release on the wrong legal premise is not a quick fix — it is exactly the kind of file an errors-and-omissions claim gets built around, with legal costs on both sides that would have dwarfed the price of routing one chatbot question to a lawyer in the first place.
The tell was the chatbot's own confidence: it gave a flat, unqualified yes to a question whose real answer is “it depends on whether a condition is still open,” with no follow-up question and no hedge. Any AI tool client-facing enough to be asked a transaction-specific legal question needs a rule for recognizing that shape of question and deflecting it — a confident, complete-sounding answer to an incomplete question is the exact failure mode a human reviewer would have caught by asking one clarifying question back. Keeping a human in the loop on anything transaction-specific is the standing fix, and the same lesson — a tool's confidence is not evidence of accuracy — is what an AI-edited listing photo teaches in an AI edit that removed a visible defect.
A 30-minute call is enough to tell you whether AI pays for itself here.
