Jiberish

AI Consulting

AI in a real-estate brokerage: what to implement first, and the rules that apply

Where AI helps a brokerage with leads, listings, and transactions, what stays human, the US and Utah rules that apply, and a 90-day plan.

Kyle Del FranciaUpdated 11 min read
In this piece

If you run a brokerage or a team, the first AI work worth implementing is drafting with review: a first reply to a new enquiry that an agent approves, a listing description checked against the listing record, and a transaction checklist that a coordinator verifies. Those tasks repeat, their inputs already exist in your systems, and a wrong draft can be caught before it reaches a client. Pricing opinions, negotiation, and anything that commits the brokerage to a buyer or seller should stay with licensed people.

Your agents are probably experimenting already. In NAR's 2026 REALTORS® Technology Report, released September 22, 2026, 23% of members surveyed said they use AI tools daily and another 25% weekly; among AI users, 75% write listing descriptions with it (NAR announcement). The broker's decision is less whether AI gets used than where review sits, which tools are approved, and which rules each workflow touches.

Sources below were checked on October 31, 2026. The rules are summarized to help you plan, not as legal advice; confirm specifics with your brokerage's counsel, your MLS, and your state regulator.

Where AI pays off in a brokerage

Speed-to-lead triage with a human handoff. An enquiry arrives from a portal, your website, or a text. An AI step can sort it (buyer, seller, renter, vendor, spam), pull out the property, timeline, and requested showing time, attach the original message, and draft a reply for the agent on duty, who starts from a structured record instead of a blank screen. Keep what the system may send by itself narrow: an acknowledgement that a named person will follow up, with your actual hours. The lead-handoff walkthrough covers failure cases such as duplicates and outages.

Listing descriptions checked against the listing data. This is where agents already use AI most, and where invented details slip in: a "remodeled kitchen" inferred from an old photo, or a lot size the input sheet never stated. Make the input sheet the only source and have the reviewer confirm each claim against it. MLS remarks rules apply too. UtahRealEstate.com (URE), for example, keeps contact information, showing instructions, and occupancy status out of Public Remarks (Rule 1.24) and requires remarks to comply with the Fair Housing Act (Rule 20.5), according to its rules reference guide revised February 11, 2026.

Market updates built from MLS reports. AI can turn a statistics export into a readable neighborhood summary, as long as the numbers come from the report, not the model. The same guide requires listing information in statistical reports to be aggregated unless the listing broker consents, with a source disclaimer (Rule 13.1).

Client updates and transaction tracking. Seller updates that summarize showing feedback, and milestone emails drafted from the checklist, save time while the agent decides what goes out. An assistant can also extract deadlines from an executed contract into the checklist; the contract governs, so a coordinator verifies every date before it drives a reminder.

Agent onboarding and training. An assistant that answers from your policy manual, citing the section it used, cuts repeat questions to the broker, and the corrections your reviewers log become training material.

What to leave human

Pricing opinions, offer strategy, and negotiation belong with licensed agents. So do questions about whether an area is safe or right for a particular family, which raise fair-housing risk however they are phrased; final approval of every ad, listing, and photo; and anything involving payment or wiring instructions. We would also keep people responsible for deciding which leads get attention, because a scoring rule built on location or similar proxies can reproduce a protected-class pattern nobody intended.

The rules that apply

Fair housing in copy and targeting. The Fair Housing Act prohibits housing ads that indicate a preference based on race, color, religion, sex, handicap, familial status, or national origin (42 U.S.C. § 3604(c)). HUD's regulation also covers choosing media or locations for ads so that particular segments of the market are denied housing information because of a protected characteristic (24 C.F.R. § 100.75(c)(3)). HUD withdrew its April 2024 guidance on digital-platform advertising in a notice published April 6, 2026, effective September 17, 2025, but the notice says conduct that does not comply with the Act itself remains subject to enforcement. NAR's director of emerging technology warns that persuasive AI copy drifts toward describing an ideal buyer, with phrases like "perfect for young families."

Buyer agreements. Since August 17, 2024, NAR's practice changes have required MLS participants working with buyers to sign a written agreement before touring a home, in person or by live virtual tour; asking an agent about their services does not trigger it (NAR consumer guide). Offers of compensation are no longer allowed on MLS platforms (NAR settlement summary; URE Rule 5.0.1). An assistant can explain the agreement but should not confirm a tour before it is signed, and remarks it drafts must leave compensation out.

MLS data. MLS licenses limit how listing data may be used, and some address AI directly: Metro MLS in Milwaukee prohibits using its data for AI training or machine learning without express written permission, NAR reported in July 2026. URE bars giving anyone else your MLS login (Rule 10.3), which matters if a tool asks for credentials to pull data for an agent, and bars using another member's listing photos without written permission (Rule 1.22). Check your MLS license before anyone pastes an export, especially other brokers' listings, into a general-purpose AI tool.

Altered and virtually staged photos. URE requires listing photos to be a true representation of the property (Rule 1.21), and NAR's 2026 Code of Ethics requires a "true picture" in advertising. California goes further: Business and Professions Code § 10140.8, in effect since January 1, 2026, requires a disclosure on digitally altered images in property advertising and access to the unaltered versions, and CRMLS Rule 11.5.2 requires the original to appear next to the altered image, labeled with a term such as "Virtually Staged." In Utah, misleading advertising is grounds for discipline, and an ad must name the brokerage or, online, link to it within one click (Utah Code § 61-2f-401(12)).

Texts and AI voice calls. The FCC confirmed in February 2024 that AI-generated voices are "artificial" voices under the TCPA, so those calls need prior express consent unless an exemption applies (FCC 24-17); telemarketing robocalls need prior express written consent. A stricter one-to-one consent rule never took effect; the Eleventh Circuit vacated it on January 24, 2025. An FCC order released October 1, 2026 lets callers treat an opt-out from informational robocalls and texts as covering only that category and designate one exclusive opt-out method; opt-outs from telemarketing still revoke consent for all of that caller's telemarketing, honored within ten business days (FCC 26-67). It was published in the Federal Register on October 26, 2026 and takes effect November 25, 2026. Make sure your form consent language covers the channels and technology you actually use.

Utah's AI disclosure law. The disclosure rules in Utah's 2024 Artificial Intelligence Policy Act, rewritten in 2025, now sit in Title 13, Chapter 77. A business using generative AI with a consumer must disclose it when the person clearly asks (§ 13-77-103). People in occupations licensed through the Department of Commerce must disclose prominently, at the start, when the interaction is "high-risk," which includes collecting financial data or giving personalized recommendations or advice someone could reasonably rely on for a significant personal decision, such as financial advice (§ 13-77-101). The Division of Real Estate sits within that department (§ 13-1-2), so licensees are covered. The definition targets systems that converse with limited or no human oversight, so in our reading a chatbot or AI phone answerer is the clear case, and one that asks about pre-approval or suggests an offer price sits close to the high-risk line. The safe harbor is a clear disclosure at the outset and throughout (§ 13-77-104), and we would use it for every AI chat or voice interaction.

A worked example: one buyer enquiry

The enquiry, listing, agent, and draft below are invented for illustration. The draft was written to show common failure patterns; it was not generated by or tested against any model.

Hi, is the 4-bedroom on Juniper Hollow Lane still available? We're pre-approved for $650,000, have two young kids, and want a safe area near good schools. Can we see it Saturday? Would the seller take $600,000?

Suppose an AI step drafts this reply for the agent: "Great news, it's still available! Juniper Hollow is a very safe, family-friendly neighborhood with great schools. I've booked you for Saturday at 10. With $650,000 you have room, and I think the seller would consider $600,000. – Megan"

Draft elementVerdictWhat the reviewer does
CRM summary: property, requested time, questions, original message attachedAcceptableConfirm the fields match the message
"Still available"Needs correctionCheck status in the MLS at send time, or say the agent will confirm
"Very safe, family-friendly, great schools"Needs correctionRemove it; point to the school district's boundary information and let the buyer judge the area
"Booked you for Saturday at 10"Needs correctionOffer times subject to agent confirmation, and explain the written agreement before a tour
"With $650,000 you have room"Needs correctionDo not comment on finances; record the pre-approval in the CRM for the agent
"The seller would consider $600,000"Needs correctionRemove it; offer strategy and the seller's position are for the agent, not the assistant
Signed "– Megan" with no AI disclosureNeeds correctionMegan reviews and sends it herself, or an automated reply identifies itself as an assistant at the start

The acceptable version thanks the buyer, says a named agent will reply about availability and timing, explains the written agreement that comes before a tour, and leaves price for the conversation. Most corrections come from rules rather than tone, so a style check alone will not catch them.

A first 90 days

These are planning judgments; adjust them to your volume and staffing.

Days 1–30: policy and one drafting workflow. Write an AI use policy naming approved tools, the client data that never goes in, and who approves what; NAR offers members a brokerage AI policy template. Find out which tools agents already use, then start listing drafts with a checklist built from your MLS rules. The check: every published description traces to its input sheet, and corrections per draft are recorded so you can see whether they fall.

Days 31–60: lead triage on one channel. AI classifies and drafts for one inbound source; agents send. Add the disclosure line and review your form consent language. The check: median time to first human reply against your pre-pilot baseline, and the share of drafts sent as written, edited, or discarded, with every fair-housing or factual correction logged.

Days 61–90: checklist extraction and training. AI extracts contract dates into the transaction checklist and the coordinator verifies each one. Run a training session built on real corrections from the first two months. The check: no unverified date reaches a client, and agents can explain why each correction was needed. Then expand, adjust, or stop each workflow based on its correction record. We would keep human approval on every outbound message through this period.

How Jiberish Labs helps

For the workflows above, Custom AI Systems & Agents maps the workflow and what stays human, connects the tools in scope where they support it, such as your forms, inbox, and CRM, and builds the AI step with review points, failure handling, and documentation. Focused builds are quoted after discovery; Embedded AI Partner provides three months of ongoing ownership for $37,500 paid upfront. AI Advisory & Training runs hands-on sessions on your agents' own tasks, including how to review outputs and handle client information, quoted after discussing team size and format. An agent portal or transaction dashboard of your own falls under AI-Embedded Apps & Portals, priced by custom proposal. An advisory hour at $200 is a reasonable way to test this plan against your brokerage before you buy anything.

Our published real-estate work is website design. Jared Looser's live Utah website brings listing discovery and buyer and seller journeys together for a client, and Alder & Summit is our self-initiated demonstration for a fictional brokerage, not a client engagement; the website journeys article walks through both. The AI-Ready Websites service is $3,499.99 a month, month-to-month, or $2,899.99 one-time with 30 days of basic support.

Put this into practice

Bring the decision you’re working through.

A free discovery session is a chance to review your goal, the tools or website you have, and the help you need. We’ll identify a sensible next step and discuss scope before any project commitment.

Custom AI Systems & Agents →AI Advisory & Training →AI-Ready Websites →Book a free call