How real estate agents use AI in 2026: 9 jobs, 6 prompts

Agents use AI mostly to write: 75% of Realtors who use it draft listings

Real estate agents use AI mostly for writing, and the National Association of Realtors has the numbers. In its 2026 Technology Report, released September 22, 2026, 23% of members use AI daily and another 25% weekly. Among those who use it, 75% write listing descriptions with it, 56% social posts, 52% emails and follow up, and only 27% use it to review or summarize documents. The share not using it and with no plans to fell to 12%, from 32% a year earlier.

Those rows are six jobs a general chatbot (ChatGPT, Claude, Gemini) does from a good prompt, and each gets one below. Three more jobs appear on every agent AI list and need a dedicated tool instead: instant lead response, CRM follow up and virtual staging. In the same NAR report, 46% of Realtors use a CRM and 41% use AI-generated content tools, NAR's top technologies used. Nine jobs is fewer than the longer lists of AI tools for agents, on purpose: every job here gets either a prompt you can paste or a tool with a published price and the legal rule that applies, and an entry with neither would be a name, not a method. The real estate prompts pack has 20 more, from showing feedback a seller can act on to outreach for an expired listing.

The last chart row is the one this page argues with. Our reasoning is that explaining a 40 page inspection report to a nervous buyer is where a general model saves an agent the most, because a buyer who understands the report at 9pm does not call at 9pm; that is this page's judgment, not a measured result, and three quarters of agents using AI have not tried it. One rule runs through every prompt here: AI writes the words, the agent owns every number.

Lead response, CRM follow up and virtual staging need a dedicated tool, and two carry a legal rule

These three jobs need software wired to your leads, your contact history or your photos, and a chatbot has none of those. A general AI workspace, Whizi included, has no MLS feed, no CRM sync and no staging engine. Published prices as of September 26, 2026, from each vendor's own pricing page:

An AI voice agent that calls a lead back is an artificial voice under federal law. The FCC's declaratory ruling, released February 8, 2024, says callers "must obtain prior express consent from the called party before making a call that utilizes artificial or prerecorded voice simulated or generated through AI technology." Structurely meters voice by the minute, so this is a live question the day you switch calling on: ask the vendor how its agent records the consent before it dials.

Virtual staging carries a disclosure rule in California. Assembly Bill 723, chaptered October 10, 2025 and in force since January 1, 2026, requires an agent who advertises with a digitally altered image to include "a statement disclosing that the image has been altered and a link to a publicly accessible internet website, URL, or QR code that includes, and clearly identifies, the original, unaltered image." The bill counts AI edits and added or removed furniture as alterations, so a staged room is covered; white balance, exposure and cropping are not. Outside California, your MLS rules decide.

Follow Up Boss notes on its pricing page that some of its AI features use data from its calling add-on, so the $69 seat is not the full price of the AI on the Grow plan. The last table row is the scheduling and admin automation that sits beside the AI jobs: ShowingTime names no AI on its pricing page, so it is not counted among the nine, but its app and live appointment specialists handle showing requests for $15 a month per agent.

Listing descriptions, and the Fair Housing line the prompt has to draw

A listing prompt has to forbid describing the buyer, because the Fair Housing Act, signed April 11, 1968, makes it unlawful to publish any advertisement that indicates a preference based on race, color, religion, sex, national origin, familial status or disability. HUD's Office of Fair Housing and Equal Opportunity said on May 2, 2024 that the Act applies to housing advertising "including when artificial intelligence and algorithms are used." NAR's own AI guidance points members back to Article 2 of the Code of Ethics, which prohibits Realtors from exaggerating, concealing or misrepresenting pertinent facts about a property.

Nothing in a plain "write a warm listing" prompt stops a model from writing "perfect for a growing family" or "steps from St. Mary's", phrases that fill the listings it learned from. The prompt has to forbid them in plain terms, and then you still read the draft.

Write a listing description of 120 to 160 words for the property below. Use only the features I give you; if a number is missing, write [NEEDED] instead of guessing. Describe the property and its features only. Never describe who should live there, and never mention families, children, professionals, age, religion, nationality, disability, safety, or any place of worship. Do not use the words quiet, safe, exclusive, or ideal for. Plain sentences, no exclamation marks, no superlatives you cannot see in the feature list. Property: [address, price, beds, baths, square footage from the tax record, lot size, year built, updates with dates, parking, HOA, school district name only]. Highlights I want first: [two or three].

HousingWire's May 15, 2026 checklist for AI written MLS remarks reduces it to one line: describe the property, not the people. The first four rows are the phrases that checklist names. Accessibility features in the last row are property facts, and stating them is encouraged.

The square footage line in the prompt is deliberate. A prompt that says nothing about gaps leaves the model free to fill them, and a filled gap in square footage is a misrepresented fact under Article 2. "[NEEDED]" makes the gap visible so the tax record can close it.

Follow up emails come from CRM notes; inspection reports are the job agents skip

Follow up from CRM notes. At 52% in the NAR report, this is the third most common job, and the one where a draft beats a blank page most plainly. Paste the notes verbatim, typos included, and let the model do the packaging. The AI for email writing guide covers the three part prompt in full; this is the agent version of it.

Here are my CRM notes on a client: [paste notes: name, buyer or seller, what they saw, what they said, price range, the last thing we agreed]. Write a follow up email under 110 words that references one specific thing from the notes, proposes one next step with a day and time, and asks one question. Warm, not salesy. No "I hope this finds you well". Do not mention any property, price, or date that is not in the notes.

Explaining a contract clause or an inspection report. Only 27% of agents using AI do this, and this page ranks it first on payoff, because a buyer who understands the inspection report at 9pm does not call you at 9pm. Upload the PDF to a chatbot that reads files. Chat with a PDF explains the one habit that makes this safe: ask the model to quote the passage before you trust its summary, so you know it is reading your document and not a generic contract from memory.

Attached is [the purchase agreement / the inspection report] for [address]. Explain [clause 14 / the roof and electrical sections] to my client in plain words, under 200 words, at the level of someone who has never bought a home. Quote the exact sentence you are explaining first. Say what it means, what the client can do about it, and by when. Do not give legal advice and do not estimate any repair cost. If anything is unclear in the document, say so rather than guessing.

The last two sentences matter. A repair cost from a model is a number with no contractor behind it, and your client will remember it. The rule from why AI gets things wrong applies with money attached: the model finds the passage, you and the contractor make the judgment.

The CMA narrative, social posts and open house prep all run on facts you paste

The CMA narrative, with one hard rule. AI cannot see the MLS: no comps, no days on market, no pending sales. Any sold price in a CMA it writes without your data came from nowhere, however real it looks. Paste the comps you pulled and make the prompt say they are the only comps that exist.

Below are the comparable sales I pulled from the MLS for [subject address]. These are the only comps. Do not add, adjust, or estimate any other sale, price, or square footage. Write a 250 word narrative for a seller explaining how these comps support a list price range of [low] to [high]: which comp is the closest match and why, which one the seller will point to and why it is not a match, and what the days on market say about pricing. Use the numbers exactly as given. If the comps do not support the range, say that. Comps: [address, sold price, sold date, beds, baths, square feet, lot, condition note, DOM].

The narrative is the part sellers read and the part agents postpone. In RPR's February 2026 survey of 225 NAR members, misinterpretation of market data was a concern for 47% of agents, and this prompt answers it: the model is not allowed to interpret data it was not given.

Social posts. 56% of agents using AI do this already. The weak version is generic ("Just listed! Stunning home in a sought after neighborhood."), and the fix is the listing prompt's: fewer adjectives, one concrete detail per post.

Write three social posts for [platform] about this listing: [paste the listing description you approved]. One post leads with the single most unusual feature, one with the neighborhood fact below, one with the open house date. Under 60 words each, no emoji, no hashtags, no exclamation marks, no words about who the home is for. Neighborhood fact: [one verifiable thing]. Open house: [day, time].

Open house prep. The least written about job, and a real saving on a Saturday morning. The model does not know the property, so this prompt gives it everything and asks for the list a good assistant would make.

I am hosting an open house at [address] on [day] from [start] to [end]. Listing: [paste]. Known issues a visitor might ask about: [list]. Give me: the 10 questions visitors are most likely to ask about this specific property with a two sentence answer for each using only the facts above, a 30 second walk in greeting that describes the property and never the visitor, and a checklist of what to set out. Mark any question I need to verify with the seller with [VERIFY].

What does AI get wrong for real estate agents?

AI gets three things wrong for agents, and all three look confident on the page.

It does not know local zoning. A model can recite what a single family zone usually allows. It cannot tell you what your municipality allows on that lot, whether the accessory unit was permitted, or what changed at the last council meeting. Zoning answers come from the planning department, and any prompt that asks for one should end with "and tell me which office to confirm this with".

It fills gaps you leave open. Square footage, lot size, year built and HOA dues are the fields where a guess does the most harm, which is why the listing prompt carries "[NEEDED]". The tax record is the only source for the number that replaces it.

It cannot see the MLS. No live listings, no sold data, no pending status, no showing instructions. Every fact about the market in an AI draft came from you or from nowhere. The RPR survey put output accuracy at the top of agents' concerns, at 63%, with compliance and legal issues second at 49%. The answer to both is a prompt that names its sources and forbids the rest.

Will AI replace real estate agents?

Nothing in NAR's numbers points that way. In its 2025 survey, fielded in July 2025 with 1,241 responses, only 17% of Realtors said AI had a significant positive impact on their business and 46% reported no noticeable impact. The work AI took on in the 2026 report is drafting: listings, posts, emails.

NAR's deputy chief economist, Jessica Lautz, put the 2026 finding this way: "Real estate agents are making practical choices about technology, and the two payoffs they want most are time and a smoother experience for their clients." The same release says faster routine work leaves agents more time for guidance, negotiation and advice, which is where the job was never a draft.

The decision rule: if the job is words, use a general model with one of the six prompts above. If the job is a number, it comes from the MLS, the tax record or the planning office, and the model is not allowed to touch it.