Who Is Liable When an AI Agent Makes the Mistake?

A chatbot that writes a listing description is one thing. An AI agent that schedules the inspection, sends the notice, answers the buyer's question, or pulls a document from a portal is something else. It's taking an action, and actions create records and obligations.

That shift has been the theme of a lot of recent AI news. In June, an AI in Real Estate Summit in New York kept coming back to the same point: the industry is moving from AI that generates content to AI that can act without a person approving each step. In late August, Anthropic made its agentic tools generally available, and Cloudflare launched payment rails built for AI agents. These tools are showing up in the software we use faster than most of us expected.

So here is the question nobody wants to answer until something goes wrong: when the AI makes the mistake, who is responsible?

The short answer is probably you

The vendor's marketing says the tool is smart. The vendor's terms of service say something different. Read them and you will usually find language limiting the vendor's liability, often to the fees you paid, and putting responsibility for reviewing output on the user.

Your license works the same way. A buyer or seller doesn't care whether a person or a piece of software sent the wrong deadline. They hired you. If an AI tool under your name tells a client the wrong thing about a contract term, misses a date, or sends a document to the wrong party, expect the complaint to name you and your brokerage, not the software company.

Courts and regulators haven't settled how this plays out. New York S7263, for example, would make operators of AI chatbots liable for substantive answers that would count as unauthorized professional practice if a person gave them, and a disclaimer that the bot is an AI would not protect the operator. It hasn't passed, but real estate is licensed work in every state, and bills like this tend to spread to other states. If you put an AI assistant on your website or client portal, you're probably the operator.

Where the gaps usually are

• Your contract with the vendor. Look for who is responsible when the tool takes an action on its own, whether there is any indemnification, and what the liability cap is. If the answer is "nothing" or "the cost of a month's subscription," know that going in.
• Your E&O policy. Ask your carrier directly whether losses caused by AI tools are covered, and whether the policy has an exclusion for automated or AI-generated output. Get the answer in writing. Some policies were written before any of this existed.
• Your brokerage's policies. If you work under a broker, find out whether they allow AI tools on client communications and what they require for review. Supervision duties don't go away because the work was automated.
• Your clients' data. If you connect an AI agent to your email or calendar, it can see wire instructions, closing dates, and personal information. Read the data retention terms yourself.

A simple standard to hold yourself to

Treat an AI agent like a new, unlicensed assistant on their first week. You wouldn't let them send contract notices or answer legal questions without checking their work, and you wouldn't blame them when a deadline got missed.

• Start with low-stakes tasks. Scheduling, reminders, and internal notes before anything that goes to a client or a counterparty.
• Keep a person in the approval step for anything that creates an obligation, moves money, or interprets a contract. This is the "human in the loop" step.
• Keep a log. If something goes wrong, you want to be able to show what the tool did, when, and who reviewed it.
• Ask every vendor one question in writing: when your AI takes an action, who is responsible for it? An evasive answer is an answer.
• Never let an AI approve a wire change. A verified callback to a known number still applies, no matter how capable the system asking for the change appears to be.

What to take from this

You don't have to avoid agentic tools. Some of them will save real time on the parts of a transaction that are tedious and repetitive. But the efficiency gain is yours and so is the risk. Before one of these tools goes anywhere near a live transaction, know what it can do, who checks it, and what your policy and your contract say when it gets something wrong.

- Jason