Every brokerage in America now has an unlicensed workforce it never hired, cannot see, and is legally responsible for. It lives on agents’ phones and laptops. It answers buyer inquiries at 2 a.m., writes listing remarks, produces pricing opinions, books showings, and increasingly negotiates. This summer a federal appeals court told us who owns its mistakes. It is not the software company.

On August 4, 2026, in Amazon.com Services, LLC v. Perplexity AI, Inc., the Ninth Circuit vacated the injunction that had kept Perplexity’s Comet shopping assistant off Amazon. The panel’s reasoning was blunt. However advanced the assistant may be, “it is a tool, not a person for statutory purposes.” The person who sends the agent out is the one who acts. Not the developer. Not the bot.

The case turned on the Computer Fraud and Abuse Act, it was decided on a preliminary record, and Amazon may still seek further review. Brokers should take no comfort in any of that. The logic is the logic every regulator, plaintiff’s lawyer, and state real estate commission will reach for. Software holds no license. It cannot be disciplined. It cannot make a consumer whole. The human who deploys it can. In real estate, the human at the top of that chain is the broker.

The duty that never moved

State license law has always rested on the duty to supervise: the broker answers for the licensed activity conducted under the broker’s license. Agents may be independent contractors for tax purposes. They are not independent of the license. When an real estate agent unleashes an AI bot, the complaint names the brokerage, the E&O claim hits the brokerage policy, and the fair housing complaint cites the name on the yard sign.

For thirty years that duty was manageable because brokers could see the work. Contracts moved through transaction management. Advertising passed through brokerage review. Supervision was imperfect, but it was possible.

AI ended that.

Why supervision just became nearly impossible

Three things changed at once.

The work went dark. A personal chatbot account, a browser agent, an AI bot service bought on the agent’s own credit card: none of it touches a brokerage system. A broker cannot supervise what the broker cannot see, and most brokers today cannot produce a list of the AI tools their agents use.

The volume exploded. A productive agent sends dozens of messages a day. An AI bot sends thousands, without fatigue and without second thoughts. Each one is a potential misrepresentation, a potential fair housing violation, a potential federal telemarketing claim. That last exposure deserves its own section, because it is the one most brokers have not priced.

The tools started acting on their own. Generative AI drafted words to copy and send. Agentic AI does the work now. It replies to leads, schedules showings, sends documents for signature, moves ad budgets, and responds to offers all without supervision. The gap between instruction and action, the gap where broker review used to live, has collapsed to zero.

The TCPA problem nobody priced in

What the TCPA is. The Telephone Consumer Protection Act of 1991 is the federal law that governs telemarketing calls and texts. Three of its features matter to every broker. First, it prohibits calls to cell phones and residential lines that use an “artificial or prerecorded voice” unless the consumer has given prior express consent, and for marketing calls that consent must be in writing. Second, it created the National Do Not Call Registry and requires companies to honor their own internal do-not-call lists. Third, and this is what makes it dangerous, it gives every consumer a private right to sue: $500 per call, trebled to $1,500 per call when the violation is willful or knowing. There is no cap. Multiply those numbers across a class of thousands and a single calling campaign becomes an enterprise-level event.

What the FCC did. In February 2024, the FCC unanimously issued a declaratory ruling, effective immediately, confirming that AI technologies that simulate a human voice, including voice cloning, are “artificial or prerecorded voices” under the TCPA. The ruling did not ban AI calls. It did something more consequential: it erased any argument that a conversational AI voice agent is anything other than a robocall in the eyes of the law. If the consumer hears a synthetic voice, the call needs prior express consent, and a marketing call needs prior express written consent. The FCC also confirmed that AI voice calls must carry the same caller identification and opt-out disclosures required of any prerecorded call.

Do not take comfort in the Supreme Court’s 2025 decision that federal courts are not bound by FCC interpretations of the statute. A voice generated by a machine is artificial under any plain reading of the word. The plaintiffs’ bar does not need the FCC to make this argument. The FCC simply made it easier.

Why this lands on the broker. Real estate already carries TCPA scar tissue, and the pattern is always the same: agents dialed, the brokerage paid. Keller Williams agreed to pay $40 million in 2023 to settle claims that its agents made unsolicited prerecorded calls, including to numbers on the Do Not Call Registry. Anywhere agreed to a $20 million settlement over roughly 700,000 calls allegedly made by Coldwell Banker agents between 2015 and 2020. In the Intero case, a federal judge found the brokerage vicariously liable for calls made by its agents, and the matter settled for about $13 million. Each of those cases involved old-fashioned dialers and recorded messages.

Now consider what AI voice changes. An agent can subscribe to an AI calling assistant for a modest monthly fee, upload a list of expired listings or stale CRM leads, and turn loose a voice that sounds human, handles objections, and books appointments around the clock. Every one of those calls is an artificial-voice call under the FCC ruling. The expired seller never consented. The portal lead may have consented to hear from someone, but consent captured by a lead vendor rarely extends cleanly to a particular agent’s AI voice campaign. And the more human the voice sounds, the worse the problem: a consumer who is never told the caller is artificial has been denied the disclosure the rules require.

Run the arithmetic. One agent’s AI assistant placing 300 calls a day for 90 days generates 27,000 calls. At $500 per call, that is $13.5 million in potential statutory exposure. If a court finds the conduct willful, it is $40.5 million. That is one agent, one quarter, one tool the broker may not know exists. AI texting tools carry their own do-not-call exposure on top of that.

This is where the supervision problem and the contracting problem converge. The broker cannot audit consent records the broker never sees. The vendor agreement must require the vendor to capture and retain consent, scrub against federal, state, and internal do-not-call lists, disclose the artificial voice at the start of every call, honor opt-outs immediately, retain call recordings and logs, and indemnify the brokerage for violations. The agent addendum must prohibit AI voice or automated outbound calling on brokerage business without written brokerage approval and documented consent.

Agentic AI is already practicing real estate

Here is the truth few vendors will say out loud. In most states, advising on price, negotiating terms, representing a party, and holding oneself out to consumers in a representative capacity are licensed activities. Unlicensed assistants are confined to clerical and scheduling work, and a human assistant who told a buyer what the seller would accept would be violating license law, with the broker answering for it.

Agentic tools are doing those exact things today. The lead-response bot that tells a buyer the seller will probably take 3 percent under. The assistant that counters an offer while the agent is at dinner. The valuation generator that delivers a value opinion straight to a seller with no licensee in the loop. Software does not get an exemption because it is software. After Amazon v Perplexity, the theory is clean: the tool did nothing; the person who deployed it did. Either a licensee engaged in licensed activity through a tool, which demands supervision, or an unlicensed actor did, which the broker permitted.

NAR’s own template draws the same line. Its NAR Brokerage AI Policy Kit says agentic tools may not perform licensed activities or create agency obligations for the brokerage, and it prohibits fully autonomous execution of high-impact actions without human review. The trade association has drawn the line. Most brokerages have not gotten the licensee to sign it.

The exposure is not theoretical. Algorithmic steering is a fair housing case waiting for a plaintiff. A hallucinated square footage or invented amenity in a listing is misrepresentation, and NAR has said plainly that the agent and brokerage remain responsible when that content goes out. Client data pasted into consumer AI accounts is a privacy event. An eager agent bot can create an agency relationship nobody intended. And when the dispute arrives, the AI agent’s logs will be evidence: the post-Perplexity litigation has already produced motions to compel an agent’s logs. If the brokerage does not control those logs, the brokerage does not control its defense.

NAR did the drafting. Brokers have to do the signing.

In March 2026, NAR released the Brokerage AI Policy Kit: a 12-section AI Use Policy template, a glossary, an agent and staff FAQ, and a one-page summary. WAV Group called it one of the most important brokerage resources of the year, and it is. It requires human review before AI content reaches a listing, a client, marketing, transaction documents, or a valuation. It makes agents and staff fully accountable for the AI content they use. It calls for audit logs. Section 9 sets the vendor bar: data ownership and portability, no training on brokerage or client data, retention, deletion and breach notification, and compliance with privacy and licensing law. Section 10 puts supervision of agents’ AI use squarely on the managing broker.

Now read what is missing. The kit has no agent acknowledgment and no signature line. NAR describes it as a resource it does not require anyone to adopt. That is appropriate for a trade association. It is not a defense for a brokerage.

A policy sitting in a shared drive is a wish. A policy incorporated by reference into the independent contractor agreement, with the agent’s signature beneath it, is evidence of supervision. NAR understood this architecture years ago: its listing content templates include work-made-for-hire and assignment language written specifically to be attached to the broker’s independent contractor agreement. AI governance needs exactly the same treatment.

Two contracts every brokerage needs now

The agent AI addendum, attached to every independent contractor agreement:

  • A list of approved tools, and a requirement to disclose every AI tool the agent uses on brokerage business, including tools the agent pays for personally.
  • No client or transaction data in unapproved tools or personal consumer accounts.
  • Human review before AI content reaches a consumer, a listing, an advertisement, or a document.
  • No autonomous licensed activity: AI may draft, a licensee decides and sends.
  • Log retention, and broker access to those logs on request.
  • Annual training acknowledgment.
  • Cooperation, indemnification, and termination for breach.

The vendor AI agreement, for every technology partner whose product touches brokerage or client data:

  • Everything in NAR’s Section 9: data ownership and portability, no model training on brokerage or client data, retention, deletion, and breach notification.
  • Logging of every action the product takes on the brokerage’s behalf, with audit access for the broker.
  • A representation that the product does not execute licensed activity autonomously, with human-in-the-loop controls the broker can configure and enforce.
  • Fair housing testing and disclosure of known limitations.
  • Compliance with MLS rules and data license terms.
  • Indemnification backed by insurance.

Some brokers will object that controlling how agents work threatens independent contractor status, which NAR has rightly spent years defending. It does not. License law already requires the broker to supervise outcomes. A contract that governs which tools touch client data and what must be reviewed before it reaches a consumer controls compliance, not hours or methods. That is the same line brokers have walked on advertising review for decades.

The MLS has a role, and an opportunity

The MLS is the other contracting chokepoint in this industry. MLS Grid already requires data consumers to accept an AI addendum before they can request MLS data. That is the model. MLSs should make AI terms a standard part of every vendor data license: disclosure of AI use, no training on listing data outside the license, logging, and a commitment not to automate licensed activity.

This is a relevance question for the MLS, not a compliance chore. The MLS that governs how AI agents touch listing data becomes the trust layer for the industry. The MLS that does not becomes a data feed, and data feeds get disintermediated.

The profitability question

Brokerage margins are already thin. A single fair housing class action or TCPA judgment can erase years of company dollar. The brokerage that can prove supervision, with signed agreements, approved tools, and logs it controls, can adopt AI aggressively and capture the productivity gains. The brokerage that cannot prove supervision has two choices: ban AI and lose its best agents to firms that allow it, or tolerate AI and bet the company on nobody suing. Neither is a strategy.

Strategic Recommendation for the Broker

  1. Inventory within 30 days. Identify every AI tool used on brokerage business, including tools agents pay for themselves. You cannot supervise what you have not counted.
  2. Adopt and sign within 90 days. Customize the NAR AI Use Policy and incorporate it by reference into every independent contractor agreement as a signed addendum. No signature, no AI on brokerage business.
  3. Re-paper every vendor. Hold every AI-enabled vendor to NAR’s Section 9 plus logging, audit access, a no-autonomous-licensed-activity representation, and insured indemnification. Vendors that will not sign are telling you something.
  4. Draw the licensed-activity line in writing. Specify what AI may draft and what only a licensee may decide and send: pricing opinions, negotiation, offers, and representations about property condition.
  5. Own the logs. If the brokerage does not hold the record of what its AI agents did, it cannot defend itself in a complaint, an audit, or a courtroom.
  6. Call counsel and your E&O carrier this quarter. Confirm how your coverage treats AI-generated content and AI-executed actions before a claim answers the question for you.
  7. Press your MLS. Ask your MLS to make AI terms standard in its data licenses so that every vendor in your market plays by the same rules.
  8. Freeze unapproved AI calling today. Prohibit AI voice and automated outbound calling and texting on brokerage business until the tool, its consent records, and its do-not-call controls have been approved in writing. This is the fastest-moving, most expensive exposure on the list.

The Ninth Circuit settled who is accountable when an AI agent acts. The only open question is whether your brokerage can prove it was paying attention.

WAV Group is available to consult with brokers and MLSs draft and adopt policy. The broker who does nothing is at risk. 

Hire WAV Group

  • Please select a service.
  • How can we help you?

 

 

 

Sources

  • Ballard Spahr, “Ninth Circuit Opines on Agentic AI in E-Commerce” (August 2026): https://www.ballardspahr.com/insights/alerts-and-articles/2026/08/ninth-circuit-opines-on-agentic-ai-in-e-commerce
  • Ropes & Gray, “Tool or Intruder? What Amazon v. Perplexity Means for Agentic AI and the CFAA”: https://www.ropesgray.com/en/insights/alerts/2026/08/tool-or-intruder-what-amazon-v-perplexity-means-for-agentic-ai-and-the-cfaa
  • ComplexDiscovery, “Ninth Circuit vacates Amazon injunction; AI agent logs emerge as eDiscovery issue”: https://complexdiscovery.com/ninth-circuit-vacates-amazon-injunction-ai-agent-logs-emerge-as-ediscovery-issue/
  • NAR Brokerage AI Policy Kit (March 2026): https://bayeast.org/wp-content/uploads/NAR-Brokerage-AI-Policy-Kit-March-2026.pdf
  • NAR, “Why Every Brokerage Needs an AI Use Policy”: https://www.nar.realtor/news/broker-news/why-every-brokerage-needs-an-ai-use-policy
  • NAR, Agent Independent Contractor “Work Made for Hire”/Assignment Language: https://nar.realtor/legal/managing-listing-content/agent-independent-contractor-language
  • WAV Group, “Here Is Your Free Brokerage AI Policy Kit”: https://www.wavgroup.com/2026/04/01/here-is-your-free-brokerage-ai-policy-kit/
  • MLS Grid, AI Addendum Processes: https://wiki.mlsgrid.com/
  • Womble Bond Dickinson, “FCC Adopts Declaratory Ruling Confirming AI Voice Calls Are Subject to TCPA” (February 2024): https://www.womblebonddickinson.com/us/insights/blogs/fcc-adopts-declaratory-ruling-confirming-ai-voice-calls-are-subject-tcpa-takes
  • Wilson Sonsini, “FCC Rules AI-Generated Voices Are ‘Artificial’ Under the TCPA”: https://www.wsgr.com/print/v2/content/49018368/FCC-Rules-AI-Generated-Voices-Are-E2809CArtificialE2809D-Under-the-TCPA.pdf
  • Inman, “Keller Williams to pay $40M to settle cold call class action lawsuit”: https://www.inman.com/2023/01/13/keller-williams-to-pay-40m-to-settle-cold-call-class-action-lawsuit/
  • National Mortgage News, “Anywhere Real Estate to settle TCPA suit for $20 million”: https://www.nationalmortgagenews.com/news/anywhere-real-estate-to-settle-tcpa-suit-for-20-million
  • Tycko & Zavareei, “$13 Million Settlement in Real Estate Telemarketing TCPA Class Action”: https://www.tzlegal.com/our-successes/final-approval-granted-in-real-estate-telemarketing-tcpa/