Ask two surveys how many agents use AI and you get two very different answers. The National Association of REALTORS put AI-generated content adoption at 46% in its 2025 Technology Survey. Five months later, a Realtors Property Resource poll of 225 members came back with 82%.
Both are probably accurate. They asked different questions, of different people, at different moments, and in a field moving this quickly five months is a long time. But the gap says something more useful than either number on its own: agents are adopting AI faster than the industry can measure it, and a lot of them are still working out what the thing is actually good for.
What follows is an inventory rather than a sales pitch. Twelve jobs AI handles well enough to be worth your time, four it still gets badly wrong, and the compliance line that runs underneath all of it.
What the numbers actually say
The RPR survey, published in February 2026, found that 82% of the agents it polled use AI in their business and 92% either use it or plan to. More telling than the headline figure: 68% use it daily or several times a week. This is not experimentation any more. It is routine.
The value they report is almost entirely about time. Seventy-one percent named time savings as the main benefit, 68% said they save at least an hour a week, and 34% save four hours or more. Nobody in that survey was claiming AI closed a deal for them.
NAR’s 2025 Technology Survey, released the previous September, gives you the tool breakdown: ChatGPT dominates at 58%, followed by Google’s Gemini at 20% and Microsoft Copilot at 15%. Most agents are using the free consumer chatbot, not a specialised real estate product.
And then there is the number nobody quotes in the marketing emails. In that same NAR survey, 46% of REALTORS said AI had made no noticeable impact on their business. Adoption and benefit are not the same thing. Most of the agents getting real value out of these tools are the ones who worked out which jobs to hand over.
Twelve jobs AI is genuinely good at
The pattern across all of these is the same. AI is useful where the work is words, where you already know what good looks like, and where a wrong answer costs you two minutes rather than a commission.
Writing you were going to do anyway
1. First drafts of listing descriptions. This is the single most common use, and for good reason: the blank page is the expensive part. Feed it the specs, the features and the three things that make the house worth seeing, and you get something to react to in fifteen seconds. You will rewrite half of it. That is fine. Rewriting beats starting.
2. Reshaping one listing for five channels. The MLS remarks, the Instagram caption, the email to your buyer list, the flyer, the text message. Same property, five very different registers and five different length limits. This is exactly the sort of mechanical transformation that language models do faster than any human.
3. The emails you rewrite every week. The price reduction conversation, the offer rejection, the inspection findings, the seller who wants to list 12% over market. You already know what you want to say. Getting it into a professional paragraph without three drafts is worth the subscription on its own.
4. Turning voice notes into usable text. Talk through the showing on the drive home, paste the transcript, ask for a client summary. Agents who work this way tend to be the ones reporting the four-hours-a-week savings, because the alternative was doing it at 10pm or not at all.
The admin nobody pays you for
5. Summarising long documents. A 60-page inspection report, an HOA document set, a municipal disclosure packet. Ask for the five things a buyer would care about and you get a reading aid in seconds. Treat it as exactly that: a map of where to look, never a substitute for reading the section that matters.
6. Cleaning up your CRM. Inconsistent name formats, duplicate entries, notes written in shorthand three years ago. Tedious, low-stakes, pattern-based work. Ideal.
7. Structuring a listing presentation. Not the numbers, which we will come back to, but the running order and the talking points. Give it the property, the seller’s situation and your commission structure, and ask for the objections you should be ready for.
Marketing you never get around to
8. A month of social posts in one sitting. Not because the AI understands your market, but because batching does. Sit down once, produce thirty captions, edit the ten worth posting, schedule them, forget about it. The failure mode of agent social media is never quality. It is stopping.
9. Neighbourhood guides. Structure, section headings, the questions a relocating buyer asks. What it cannot supply is the part that makes the guide worth reading: which coffee shop actually has parking, which school district boundary just moved, which street floods. Bring that yourself and the guide is genuinely useful. Skip it and you have published the same page as everyone else.
10. Newsletter drafts. Three market notes, a listing, a local event. Thirty minutes a month instead of an intention you never act on.
Preparation and practice
11. Rehearsing hard conversations. Ask it to play a seller convinced their home is worth 15% more than the comps support, and to push back the way that seller would. It is a surprisingly good sparring partner, precisely because it never gets tired of you and never lets you off the hook out of politeness.
12. Explaining things you half-know. Assumable mortgages, 1031 exchanges, a title issue you have hit once in six years. Get the plain-English version, then verify it with someone qualified before you repeat it to a client. That second half of the sentence is not optional.
Four things it still gets badly wrong
The failures are not random. They cluster around a single limitation: a language model predicts what a plausible answer looks like. It does not check whether the answer is true.
1. Anything involving arithmetic. CMAs, net sheets, price per square foot, proration, commission splits. Research cited by real estate technologists puts GPT-4 at roughly 71% error on financial calculations without a calculator tool attached, and still failing around 14% of the time with one. Worse, a general chatbot has no access to your MLS. It is not calculating a CMA. It is imitating the shape of one, using numbers it invented. Have your MLS or your CMA software do the maths, then let AI write the narrative around figures you already trust.
2. Facts about a specific property. Ask for a listing description and mention a two-car garage, and you may find a paragraph praising the workshop space and the new roof. Neither exists. It filled the gap because listings usually have those things. Every factual claim in AI-drafted copy has to be checked against the property, every time.
3. Legal and contractual questions. Contingency language, disclosure obligations, what a specific clause commits your client to. It will answer confidently, it will sound authoritative, and it may be describing the law of a state you are not in. Your broker and your attorney exist for this.
4. Predicting your market. Where prices go next quarter, whether now is the moment to sell. It has no live data and no forecasting ability. What it produces is a confident-sounding average of everything written about real estate before its training cutoff, which is a different thing entirely from an answer.
The fair housing problem hiding in AI copy
This deserves its own section because it is the risk agents underestimate most, and because 28% of the agents in the RPR survey named it as a concern while carrying on regardless.
Language models learn from decades of published listing copy, including copy written before anyone was paying attention. What comes back can carry phrasing that describes the ideal occupant rather than the actual property, which is the textbook definition of steering.
The phrases that cause trouble sound harmless. Perfect for young professionals. Ideal for families. Walking distance to church. Quiet neighbourhood. Each one gestures at age, familial status, religion or ability, and each one turns up regularly in AI-generated drafts because it turns up regularly in the training data.
The rule that solves most of it is short enough to keep on a sticky note: describe the property, not the buyer. Not “perfect for joggers” but “backs onto a jogging trail”. Not “great for a growing family” but “four bedrooms and a fenced yard”. The facts do the selling and the risk disappears.
Before anything AI-drafted goes into the MLS, it is worth running the same short check every time:
- Strip any sentence that describes who should live there.
- Remove coded terms such as “safe”, “exclusive” or “desirable”.
- Verify every factual claim against the property record.
- Label virtually staged or AI-enhanced images clearly.
- Read the final version yourself before it publishes.
That last point is not a formality. Under Article 2 of the NAR Code of Ethics you are responsible for what you publish. “The software wrote it” has never been a defence and is not going to become one.
If you are starting from zero this week
You do not need a stack. Most of the agents in these surveys are using the free version of ChatGPT and nothing else, and the ones seeing real time savings got there by narrowing rather than expanding.
Pick one recurring task that costs you time and produces words. Listing descriptions is the obvious candidate. Do it with AI for two weeks, keep the prompts that work in a note, and time yourself honestly. If you are not saving at least an hour a week on that one task, the problem is the prompt, not the tool, and it is worth fixing before you add anything else.
Then add a second task. Not five.
Common questions
Do I have to tell clients I used AI?
For drafting an email or a listing description, generally no, in the same way you do not disclose using a spell checker. It changes when AI alters what a buyer sees, such as virtually staged or AI-enhanced photography, where labelling is expected and in a growing number of places required. Rules vary by state and by MLS, so check yours rather than assuming.
Is the paid version worth it?
If you use it most days, yes. The paid tiers of the major assistants handle longer documents, keep context across a conversation and produce noticeably better writing. If you use it twice a month, the free tier is fine and you should spend the money elsewhere.
Should I use a real estate specific AI tool instead of ChatGPT?
Not at first. NAR’s survey found 58% of agents using ChatGPT against a long tail of specialist products, and that is a reasonable order of operations. Learn what the general tool can do, find the specific job it handles badly, then buy the specialist product that fixes that job. Buying first tends to mean paying for features you never open.
Will AI replace real estate agents?
Nothing in the current evidence points that way. What it replaces is the unpaid hour spent staring at a blank listing description at nine in the evening. The parts of the job that clients actually pay for, being present in a negotiation, reading a room, knowing which contractor answers the phone in August, are not text prediction problems.
Sources
- Realtors Property Resource, AI adoption survey of 225 NAR members, published 12 February 2026.
- National Association of REALTORS, 2025 Technology Survey, published 18 September 2025.
- HousingWire, fair housing checklist for AI-written MLS remarks.
- Neuhaus Realty Group, analysis of AI accuracy in real estate calculations and CMAs.
Figures were checked against the published surveys in August 2026. Survey data reflects the period each study was conducted.