AI does the assembly work in a comparative market analysis well — the table, the summary, the narrative, the deck. It does not make an adjustment defensible, and it does not turn a CMA into an appraisal.
Key takeaways
AI can do the assembly work in a comparative market analysis — pulling comparable sales into a table, summarising a neighbourhood, drafting the commentary and formatting the deck. It cannot make the adjustments defensible, and it cannot turn a CMA into an appraisal. The opinion of value at the end is still yours, and every Canadian real estate regulator that has published on AI says accountability does not move.
That is the whole answer. What follows is where the line sits in practice, and how to use the tool without producing a document you would not want read back to you.
A comparative market analysis is a registrant’s opinion of likely selling price, prepared to help a seller price a listing or a buyer frame an offer. A formal appraisal is a separate professional product. The Appraisal Institute of Canada grants two designations — the Accredited Appraiser Canadian Institute (AACI) and the Canadian Residential Appraiser (CRA) — and describes designated members as providing independent valuations whose opinions of market value rest on comprehensive research and analysis, as set out on its page on what appraisers do. The AIC is self-regulating and holds its members to the Canadian Uniform Standards of Professional Appraisal Practice, per its professional standards page.
This matters when an AI tool hands you a confident-looking valuation. If the output reads like an appraisal, gets used like an appraisal and is relied on by a lender or a court, you have stepped outside what a CMA is for. BCFSA makes the general version of the point in its Artificial Intelligence Guideline: AI lacks professional licensure, the platforms offer no dependable warranties, and a licensee who uses AI to give advice in areas where their expertise is limited or where they lack proper licensure may be subject to regulatory sanctions.
Four parts of CMA preparation are mechanical, and all four automate well.
Assembly. Turning a set of sold records into a clean comparison table — beds, baths, square footage, lot, days on market, list-to-sold ratio — is data reshaping. A model does it faster than you and does not mistype.
Summarising. Reading twelve listing remarks and reporting what the finished basements and the recent kitchens have in common is exactly what language models are good at.
Drafting. The narrative page that explains the range in plain language, and the alternative version pitched for a first-time seller rather than an investor.
Formatting. Turning your numbers into the same deck every time, with the same headings, so nothing gets left out under time pressure.
What does not automate is the judgment in the middle: which comparables are genuinely comparable, what each difference is worth, and which facts about the subject property change the answer.
RECO’s Bulletin 7.3 on material facts is the most useful Ontario text to read before you let a model near a valuation. It lists the kinds of facts that are often material — knob-and-tube wiring, lead or galvanized plumbing, the age and ownership status of major systems, a history of flooding or fire, renovations and whether permits were obtained, property tax and special assessments, zoning by-laws, prior illicit use, rights-of-way and restrictions, and nearby facilities such as quarries, industrial sites, airports or rail lines.
The bulletin then makes the point that decides whether an AI-assisted CMA is safe: when representing a seller, the reasonable steps required to determine material facts “is not met by simply accepting the seller’s verbal representations” — some research, verification and supporting documentation will be necessary.
An AI tool reading listing data cannot do that verification. It sees “renovated 2021” in the remarks and treats it as a fact. It does not know whether the permit was pulled, whether the rental water heater transfers, or whether the comparable two streets over backs onto a rail corridor. Those are the differences that move an adjustment by five figures, and they are the ones you check yourself.
A three-column habit that keeps a CMA honest
Column one — taken from the record. Facts that came straight from the listing or registry data. Fast, and fine for a model to assemble.
Column two — verified by me. Anything you confirmed against a permit search, a survey, a status certificate, a tax bill or your own eyes.
Column three — assumed. Everything the model inferred or you estimated. If a number in column three is doing heavy lifting in the range, either verify it or widen the range and say why.
CREA’s artificial intelligence guidance defines an AI hallucination as output that sounds accurate or convincing but is false, misleading, incomplete or entirely fabricated, and recommends three practices: require human review and verification before AI-generated content is published, shared with clients or relied on in decision-making; exercise caution on high-risk tasks or those requiring specialised expertise; and provide training and internal policies. It also states directly that a REALTOR® who creates inaccurate, deceptive or misleading information as a result of using AI risks being held responsible.
In a CMA the hallucination risk is specific and easy to miss: a model asked to find comparables for an address it has no data for will produce plausible ones. Street names in the right neighbourhood, sensible prices, dates that fit. The defence is not cleverness in the prompt — it is only ever giving the model the comparable set you retrieved yourself, and treating anything it adds as a fabrication until proven otherwise.
A CMA delivered to your client is a service document. The moment its contents appear in a listing presentation posted online, a “what your neighbour sold for” postcard or a social carousel, Ontario’s advertising rules apply. RECO’s Bulletin 5.1 requires that all statements be factually correct, accurate and verifiable, that comparative claims be truthful and supported by verifiable facts, and that statements about business volume or trading activity explain how the volume was measured. It also prohibits advertising anything that could reasonably be used to determine the contents of an agreement, such as price, or to identify a specific property or a party to a transaction, without consent.
So the CMA you show a seller and the CMA excerpt you post are governed differently, and an AI tool that repurposes one into the other will not notice.
A seller in Hamilton wants a price for a 1,150 sq ft semi with a 2022 kitchen and a finished basement of uncertain permit status. Six sales in the last 90 days are arguably comparable.
The workable sequence is: retrieve the six yourself from your board’s system; paste the fields into the tool and ask it to build the comparison table and flag which of the six differ most from the subject; do the permit search on the basement and the two comparables that also claim finished space; make the adjustments yourself, writing one sentence of reasoning per adjustment; then ask the tool to turn your table and your reasoning into the client-facing narrative in two versions, one cautious and one confident, so you can choose the register.
Elapsed time saved is real — the table and the write-up are most of the evening. What you did not do is ask the model for the price. The moment you do, you have an opinion of value you cannot explain, and explaining it is the job.
You can ask, and you will get a number. It will not be grounded in your board’s sold data, which is not public, and you will not be able to show where it came from. BCFSA’s guideline notes that AI tools are designed to guess answers to questions they do not have complete information on. A number you cannot source is not usable in front of a client.
Treat it as off-limits unless you have checked. BCFSA advises acquiring clients’ informed consent before using their information in an AI tool, and says that unless you are certain the tool does not store or use your input you should avoid entering confidential or personal information at all. RECO’s Bulletin 2.5 on confidentiality is the Ontario counterpart: confidential client information must not be disclosed to a third party without the client’s written consent, and the duty continues after the relationship ends.
CREA’s guidance frames transparency as a core value and says consumers should clearly understand when AI is materially influencing information or representations. BCFSA goes further and asks licensees to communicate to clients when AI tools are being used and explain their limitations. A line in the document saying the analysis was assembled with AI assistance and reviewed by you costs nothing and answers the question before it is asked.
A 30-minute call is enough to tell you whether AI pays for itself here.
