It is nine in the evening and your buyer has read the inspection response addendum three times. They send you a photo of paragraph 4 and ask what happens if the seller does nothing by the deadline. You do not know off the top of your head, their attorney will not answer until tomorrow, and the answer is fifteen seconds away in an app on your phone.
This is the single most common way AI real estate contract review goes wrong, and it is not because the model is unhelpful. It is because it is extremely helpful, instantly, in the confident register of somebody who went to law school, about a form it has never seen.
Two different problems, and only one of them is new
Pasting a contract into a chatbot creates two risks that people tend to blur together. They have different fixes.
- Unauthorized practice of law. Explaining to a client what a contract provision means is legal advice in most states, whether you worked it out yourself or read it off a screen. This rule is decades old and has nothing to do with AI.
- Confidentiality. An executed contract contains names, addresses, financing details and sometimes far more. Putting that into a consumer chat product is a decision about your client’s data, and it is one most agents make without noticing they made it.
The second one has a straightforward answer, covered further down. The first one is where careers get damaged.
What the line actually is
The National Association of REALTORS® puts it plainly. Agents may insert factual information — the parties, the property, the amounts — into blank, attorney-approved form agreements. What they may not do is draft contracts, make modifications beyond factual entries, interpret what a provision means, or answer legal questions. Article 13 of the Code of Ethics requires recommending counsel rather than supplying advice, and state rules sit on top of that with their own definitions and their own penalties, which run to fines and license revocation.
Notice what that list does not say. It does not say “unless you are confident”. It does not say “unless the client asked you directly”. And it does not say “unless you got the answer from somewhere else”, which is precisely the loophole an agent feels like they have found the first time a model produces a clean, plausible paragraph about a clause.
The model does not have your form
This is the part that surprises people, so it is worth being concrete about how wrong the answers get.
In September 2025, Inman published the results of a compliance expert, Summer Goralik, putting a series of California real estate law questions to ChatGPT. The answers were not gibberish. They were the far more dangerous thing: mostly right, fluent, and specifically wrong in ways only somebody who already knew the answer would catch. It gave the wrong deadline for depositing trust funds. It cited the wrong subsection of the Business and Professions Code for an exemption, and described that exemption as covering entities it does not cover. It gave an incomplete answer on dual-role disclosure that left out the twenty-four-hour written disclosure requirement entirely.
Every one of those errors would read as authoritative to an agent who was asking because they did not know. That is the whole failure mode. The same limitation is why a general assistant cannot reach into your MLS and why it should never be trusted with a figure: it produces the shape of a correct answer from patterns, not from the current text of your state association’s form.
Your purchase agreement was revised at some point in the last few years. Paragraph numbering shifted. Your state added a disclosure. The model’s version of all of that is an average of everything it read, weighted towards whatever was most common on the internet, which is usually neither your state nor this year.
Where it is genuinely useful
None of this means close the app. It means moving the work to the side of the line where the output is for you rather than for the client.
| Task | Safe? | Why |
|---|---|---|
| Explaining a clause to your client | No | Legal advice, regardless of source |
| Telling you what a term of art generally means | Yes | Vocabulary for your own understanding, not advice you pass on |
| Turning your confusion into a precise question for the attorney | Yes | This is the highest-value use and almost nobody does it |
| Summarizing a clause you already understand | Careful | Fine as a draft you verify line by line; not fine as a shortcut |
| Calculating a deadline from an effective date | No | Arithmetic, plus business-day rules it does not know |
| Drafting an addendum or amendment | No | Drafting is prohibited in most states, full stop |
| Comparing two offers on price and terms you typed in | Yes | You supplied the facts and you are the audience |
The third row is the one worth building a habit around. Attorneys bill by the question, and a vague one produces a vague answer and another round of email.
The prompt that turns confusion into a question
I am a licensed real estate agent, not an attorney. I am not asking for legal advice and I will not repeat your answer to a client.
Here is the situation in general terms: [describe the fact pattern with no names, no addresses and no account numbers].
Write the three questions I should ask a real estate attorney licensed in [state] so that one email gets me a usable answer. For each question, say in one line why it matters. Do not answer the questions yourself. Do not tell me what the law is.
The instruction not to answer matters. Left to itself the model will helpfully volunteer the law anyway, and you will read it, and it will lodge in your head as something you know.
The prompt for understanding vocabulary
Explain what “liquidated damages” means as a general concept in US residential real estate, in plain English, in under 120 words.
Then list three things that vary by state or by contract form, so I know what I would still need to check. Do not tell me what my contract says. Do not give me a number.
Used this way it is a glossary that answers follow-up questions, which is a real improvement on the version of professional development where you nod and look it up later and never do.
The data question
Separate from all of the above: do not paste an executed contract into a consumer chat account. Not the signature page, not the lender’s pre-approval, not the page with the buyer’s employer and salary on it.
- Strip names, addresses, account numbers and dates of birth before anything goes in. If stripping it makes the question unanswerable, that is a sign the question belongs with the attorney.
- Check whether your account’s settings train on your conversations, and turn that off. It is one toggle and most agents have never opened that screen.
- Ask your broker whether the brokerage has an approved tool. Some do now, with a business agreement behind it, and using the sanctioned one costs you nothing.
- Assume anything you paste could be read by a human reviewer at some point. That is not paranoia, it is how these services are documented.
The same caution applies anywhere client information meets a tool you did not vet, which is the recurring theme in the piece on AI and cold calling.
What to say at nine in the evening
The client is not really asking you to practice law. They are asking you to make the anxiety stop. That can be done without answering the question.
That paragraph is doing something specific and I’m not going to guess at it, because if I’m wrong it costs you money and it’s my license. Here’s what I can do tonight: I’ll send it to [attorney] now with the exact question, and I’ll tell you what the practical options usually look like so you’re not sitting there in the dark. Nothing is due before Thursday, so we have time.
Three things happen there. You have declined without sounding evasive, you have named the deadline, and you have given them something to hold onto until morning. Agents who handle this well are not the ones who know the answer. They are the ones who are unembarrassed about not knowing it.
Common questions
Is it unauthorized practice of law if I just forward what the AI said?
The prohibition is on the advice reaching your client through you. Forwarding it does not create a separate author who takes responsibility. If anything it is worse, because you have now put an unverified statement about the law in writing under your name. The same logic decides what to do with an AI-written buyer letter.
What about the AI review features built into transaction software?
Better on the data question, since there is usually a business agreement behind it. Identical on the legal one. A summary generated inside a compliance platform is still a summary, and reading it aloud to a client is still interpretation. Ask the vendor directly whether their output is intended for client delivery, and notice how carefully they answer.
Can I use it to check my own understanding before I call the attorney?
Yes, and this is the best use of it. Write out what you think the clause does, ask the model where a reasonable person might disagree with you, and take both to the attorney. You will have a sharper call and a shorter bill.
Does any of this change if I tell the client I used AI?
No. Disclosure fixes a trust problem, not a licensing one. When and how to mention AI at all is worked through in the piece on telling a client you used AI, but no phrasing converts legal advice into something you are allowed to give.
My state lets agents fill in standard forms. Does that cover me?
It covers filling in blanks with facts. It does not extend to explaining the effect of what you filled in. That distinction is the entire subject, and where exactly it sits is a question for your state commission rather than for a chatbot, for reasons this article has hammered on about enough by now.
If the broader question is what these tools are actually good for in a working week, the honest inventory is in the overview of how agents use AI, and the short version is that the jobs where being wrong is cheap are the ones worth automating.
Sources
- National Association of REALTORS®, what constitutes the unauthorized practice of law, including the distinction between entering factual information and interpreting provisions.
- NAR Code of Ethics, Article 13 on recommending legal counsel rather than providing legal advice.
- Inman, compliance consultant Summer Goralik’s test of ChatGPT on California real estate law, 23 September 2025.
Checked August 2026. General information, not legal advice. Rules on the unauthorized practice of law are set by each state; confirm with your broker, your state real estate commission and an attorney licensed where you practice.