Lesson objective

Distinguish rules-based automation, machine learning and generative AI in brokerage tools, and match the check you apply to the way each one fails.

Plain-language explanation

Why this matters now

"AI" now appears on almost every product a brokerage touches: the CRM that scores new leads, the transaction platform that sets task reminders, the listing tool that offers draft remarks, the photo editor that furnishes an empty room, and the chat assistant that writes a follow-up email. The label is the same. The systems underneath are not, and they fail in different ways.

The best recorded picture of how agents use these tools comes from the National Association of REALTORS® Research Group's 2025 REALTORS® Technology Survey, published on September 18, 2025. Its numbers need their sample stated before they mean anything:

  • In July 2025 NAR invited a random sample of 49,233 active REALTORS® to an online survey and received 1,241 usable responses, a response rate of 2.5 percent, with a margin of error of plus or minus 2.78 percent at 95 percent confidence.
  • Respondents were 68 percent sales agents and 14 percent associate brokers, with a median of 11 years' experience.
  • Asked how often they had used AI in their business over the past 12 months, 20 percent said daily, 22 percent weekly, 27 percent a few times a month, and 32 percent had not actively tried AI for their business. Those categories sum to 101 percent through rounding.
  • Asked about AI's impact on their business, 17 percent reported a significantly positive impact, 33 percent moderately positive, 46 percent neutral or no noticeable impact, 2 percent moderately negative and 2 percent significantly negative.

Read those figures only as far as the sample allows. A 2.5 percent response rate means the results describe the REALTORS® who answered, not all licensees, and licensees who are not NAR members were not sampled at all. "Impact" is a self-report, not a measurement. This course holds no study that measures how much time agents save with generative AI, so no productivity figure appears here.

Consumers, meanwhile, still come to people. NAR's 2025 Profile of Home Buyers and Sellers (highlights published November 4, 2025) surveyed buyers of a primary residence who bought between July 2024 and June 2025: 173,250 surveys were mailed and 6,103 returned, an adjusted response rate of 3.5 percent. It reports that 88 percent of buyers purchased through a real estate agent or broker and 91 percent of sellers sold with an agent's assistance. NAR's news article on the report says agents were "the most trusted and frequently used information source—well ahead of online listings." The public highlights contain no questions about AI, so the Profile tells you nothing about how buyers feel about AI.

On its Artificial Intelligence (AI) in Real Estate page, NAR describes consumers as increasingly relying on REALTORS® as the "human in the loop" for AI-assisted tasks like home searches and price estimates. That is a trade association's overview, not a rule. But the point holds on its own: if consumers trust the licensee, the licensee needs to know which kind of system is behind each answer.

A question to sit with

Your CRM ranks today's leads, your transaction platform sends the same reminder on every file, and your listing tool offers a polished paragraph of remarks. All three are sold as "AI". Which of them could be wrong in a way a buyer would notice before you did?

What goes wrong

Treating every "AI" label as one thing. A brokerage writes one instruction, "check AI outputs", and applies it the same way to a fixed reminder rule and to generated listing remarks. The rule is checked every week for no reason and the remarks are checked too lightly every time.

Checking a generated draft the way you check a rule. A rule that worked on one file behaves the same on the next file with the same facts. A generative tool writes new text every time it runs. Yesterday's accurate draft says almost nothing about today's.

Reading a score as a verdict. Lead scores, match scores and "confidence" ratings are the system's own estimate of how strongly its learned patterns point one way. They are not proof, and they are weakest on the lead, the property or the buyer who does not look like the past.

Assuming fluent language means knowledge of the property. A well-written sentence about a "renovated chef's kitchen" can rest on nothing at all. Fluency is a property of the words, not of the facts behind them.

Letting the tool decide who owns the result. No product category moves responsibility. The licensee who would have owned the advertisement, the client email or the market summary without the tool still owns it with the tool, under the supervision of the sponsoring or designated broker.

Treating a survey as a mandate. "Most agents use AI" is not what the NAR survey says, and even an accurate figure about what others do tells you nothing about whether a use is safe for your client.

The practical answer

Sort every AI feature into one of three families before you decide how to check it. Many products combine families, so sort the feature, not the product.

Rules-based automation Machine learning Generative AI
How it works A person writes explicit if-then logic The system learns patterns from past examples, such as past leads or past sales A model trained on very large amounts of text or images produces new text, images or audio
Typical brokerage use A CRM sends a set email to every open-house sign-in; a transaction platform creates a task list when a contract is marked accepted Lead scores; photo room-type tagging; value estimates built from past sales Draft listing remarks, follow-up emails, market summaries; virtual staging images
How it fails Consistently: the rule is wrong, or a file falls outside it Quietly, on unusual cases; drifts when the market changes Plausibly: confident, well-formed content that is false
What a wrong output looks like The same wrong action every time A reasonable-looking number or label A fluent, complete-looking draft or a realistic image
How you check it Test the rule on real cases; review what it did; control changes Sample outputs; look hardest at unusual cases; ask what data it learned from Verify every fact against the listing record, seller documents or the rule itself

Three questions turn the table into a habit.

  1. Which family produced this output? If you cannot tell, ask the vendor or your broker, and write the answer down.
  2. How does that family fail? Match the check to the failure. A rule needs a test. A score needs sampling and a second look at unusual cases. A generated draft needs line-by-line verification against the record.
  3. Who owns the result? That depends on the work, not the tool:
    • the licensee for every advertisement, listing description, property representation and client communication, under the supervision of the sponsoring, designated or responsible broker, as each state's licence law defines that duty (Module 4);
    • the seller for the seller's own property disclosures (Module 10);
    • the housing provider for a rental decision (Module 11).

Each state licenses and supervises in its own way, so check your own commission's rules; Module 4 shows how.

Apply that idea one feature at a time. For each AI feature in your work, write one line: what the system does, what a person does, and who decides.

Feature System does Person does Who decides
Open-house follow-up rule Sends a set email to every sign-in Reviews the template and the sign-in list weekly Team leader, under the broker's policy
Lead score Ranks new leads Calls in order, and reviews low scores before discarding any lead The licensee working the lead
Draft listing remarks Produces a paragraph Checks every claim against the MLS record and seller documents, reviews for fair housing, edits The listing licensee, under broker supervision

Remember this

  • "AI" covers three families (rules-based automation, machine learning and generative AI) that fail in different ways.
  • Match the check to the failure: test rules, sample scores, verify generated content against the record line by line.
  • A score is the system's own estimate, not proof, and it is weakest on unusual cases.
  • A survey figure means nothing without its sponsor, sample and date; the NAR 2025 survey describes 1,241 responding REALTORS®, not all licensees.
  • The tool never moves ownership: the licensee owns the output, under the broker's supervision, and fair housing law reaches every family.

Real estate example

Copperline Properties is a fictional 22-agent brokerage. Before a sales meeting, designated broker Renata Osei asks team leader Marcus Hale to list every AI feature his team touched in one week and say how each is checked. Marcus finds three. The first is a CRM rule, written two years ago, that sends "Thanks for visiting 14 Alder Crescent! It is still available, and we would love to show it again" to everyone who signs in at an open house. The listing went under contract on Tuesday; the rule kept sending the same email through Saturday. It is rules-based automation: it fails the same way every time, so Marcus switches the rule off, rewrites the template without the availability claim, and adds a step to change templates when a listing status changes. The second is the CRM's lead score. A buyer who registered through the brokerage website sat at 14 out of 100 and was never called; she bought through another brokerage. The score came from patterns in past leads, and she did not look like them. Marcus asks the vendor what the score learned from and sets a rule that no lead is closed out on score alone. The third is a generative tool that drafted remarks for a fictional listing, Tri-County MLS number TC-40418. The draft called the kitchen "recently renovated with quartz counters." The fictional MLS record and the seller's property information sheet both say "original 1998 cabinets, laminate counters." The phrase was generated because it is common in listing remarks. The listing agent, Dana Whitlock, deletes it, rewrites the sentence from the record, reviews the draft for fair housing wording, and approves it. Marcus's one-page note lists each feature, its family, its check and its owner: the team leader for the rule, the working licensee for the lead, and the listing licensee for the remarks, all under Renata's supervision.

Responsible-use warning

Do not let a score or a polished draft shorten your review. A lead score or match score is the system's own estimate from patterns in past data, and it is least reliable on the case that does not resemble the past. A fluent generated paragraph can describe a feature the property does not have. Neither is evidence. Before you rely on an AI output, name the family that produced it and apply the matching check: test the rule, sample the scores, or verify every claim in the draft against the MLS record and the seller's documents.

Human-review requirement

The licensee who owns the work decides whether an AI feature was used appropriately and checks its output: every advertisement, listing description, property representation and client communication belongs to the licensee, under the supervision of the sponsoring, designated or responsible broker as their state defines that duty. The broker confirms that each feature in use has been sorted into its family, that the check matches its failure mode, that no output was accepted on a score or fluent wording alone, and that every feature that drafts, targets, answers or decides has a fair housing review step. The seller still owns the seller's disclosures.

Practical exercise

Map the AI features in one brokerage workflow you know, using fictional or anonymised examples only. Do not paste client or transaction data anywhere for this exercise.

  1. Choose one recurring task, such as launching a new listing, working a week of website leads, or managing a file from acceptance to closing.
  2. List every feature in the tools you use for that task that is described as AI, smart, automatic, suggested or scored.
  3. For each feature, decide its family: rules-based automation, machine learning or generative AI. If you cannot tell, write "unknown" and the name of the person you will ask.
  4. For each feature, write one sentence on how that family fails.
  5. For each feature, write the check that matches the failure: a rule test, a sample of scores, or line-by-line verification against the record.
  6. Complete a four-column table: feature, what the system does, what a person does, who decides.
  7. Mark every feature that drafts, targets, answers or decides, and note where a fair housing review would sit.
  8. Name the owner of each output: the licensee, the supervising broker, the seller or the housing provider.
  9. Keep the table in your workbook; you will build on it in the risk matrix later in this Level.

Knowledge check

  1. Why should a CRM follow-up rule and AI-drafted listing remarks not be checked in the same way?

    Answer guidance: The rule is explicit logic that fails consistently, so it is tested and its actions reviewed. A generative tool writes new text each time and fails plausibly, so every claim in each draft must be verified against the MLS record and seller documents before publication.

  2. What does a lead score tell you, and what does it not tell you?

    Answer guidance: It is the system's own estimate of how strongly patterns in past leads point to this lead converting. It does not prove anything about this consumer and is least reliable for leads unlike those it learned from.

  3. What must you say about the NAR 2025 REALTORS® Technology Survey before quoting its AI use figures?

    Answer guidance: That it drew 1,241 usable responses from a random sample of 49,233 active REALTORS®, a 2.5 percent response rate, so it describes responding members rather than all licensees, and that non-members were not sampled.

  4. What does the 2025 Profile of Home Buyers and Sellers tell you about buyers and AI?

    Answer guidance: Nothing. The public highlights contain no AI questions. It reports that 88 percent of buyers purchased through an agent or broker, and NAR says agents were the most trusted information source, but it cannot support any claim about AI.

Sources and further reading