Delta Media Group published its three-year read on brokerage AI anxiety on August 26, and the number moved the wrong way. The average worry score among brokerage leaders sat at 6.50 out of 10 in 2024, fell to 5.80 in 2025, then climbed back to 6.38 in 2026. Two more years of daily use did not make anyone calmer. The split underneath is the useful part. Among female brokerage leaders, 71.4 percent named regulatory compliance as a top AI concern, against 39.1 percent of men. Men worried about cost and return instead, 51.6 percent to 28.6 percent.
Those are two different questions about the same software. One group asks whether the subscription pays for itself. The other asks who answers when the output is wrong. If you run a solo shop or a five person team, you own both questions. Only one of them has a regulator attached to it.
The Regulator Already Answered Your Supervision Question
The California Department of Real Estate issued an advisory on March 17, 2026 that most agents scrolled past. The short version: using AI to perform licensed activity can amount to letting an unlicensed assistant perform licensed activity. The DRE also warned that brokers face discipline for failing to supervise how their agents use these tools. That is not a proposed rule waiting on a comment period. It is the existing supervision duty, applied to software you already pay for.
A neighboring industry supplied the case study. Earlier this year a federal judge sanctioned the managing partner of a California law firm after a junior attorney filed a brief containing a citation that did not exist. The partner did not draft the brief. He did not sign it. He stayed counsel of record, and the court fined him and ordered him to complete training on supervision and ethical AI use. Swap broker for managing partner and agent for junior attorney. The structure holds without modification.
Compliance consultant Summer Goralik has a name for the failure mode: reverse compliance. You adopt the practice, something breaks, and the complaint teaches you what the rule was. For a solo agent the dollar exposure is smaller than a brokerage faces and the proportional damage is worse. One fabricated square footage figure, one altered listing photo published without disclosure, and the thing on the table is your license.
Inventory the Tools Before You Write Any Rules
You cannot supervise a stack you have never written down. Block twenty minutes this week and build a four column list. Column one, the tool. Column two, what client data goes into it. Column three, what it produces that reaches a consumer or the MLS. Column four, the licensed human who reads that output before it ships. No software required. A spreadsheet is fine.
Most small teams find six to nine tools once they actually count. ChatGPT Plus at roughly $20 per person per month. Gemini folded into a Google Workspace plan somewhere around $14 to $22 per user. Microsoft 365 Copilot at $30 per user per month. A CRM with generative features baked in, Lofty or Follow Up Boss or BoldTrail, running anywhere from $80 to several hundred a month depending on seats. A listing description generator. A video tool. Then the free browser extension somebody installed in March and nobody has thought about since.
Column two is where the surprises live. Ask each person on the team, plainly, whether they have pasted a seller’s financial situation, a buyer’s name and address, or an inspection report into a general consumer chatbot. The honest answer is usually yes. That is the moment data privacy stops being a survey category. Delta found 65.7 percent of female leaders flagging data privacy against 48.4 percent of men, and the person doing the pasting is not thinking about either number.
The One Page Policy You Can Copy Today
Long policies do not get read. Write one page, drop it in the shared folder, and have every licensed person on the team sign the bottom. Here is a working draft. Change the names, keep the shape.
- Approved tools only. Only tools on the attached list touch client work. Adding a tool requires written approval from [broker or team lead].
- Never paste. Client names, addresses, financials, inspection findings, and signed documents do not go into a general consumer chatbot. Use the approved CRM or a business account with a data agreement in place.
- Human review. Every AI drafted item that reaches a client, a consumer, or the MLS gets read line by line by a licensed person first. That person owns the content as if they typed it.
- Facts come from the record. Square footage, school assignment, tax figures, HOA dues, and permit history come from source documents, never from a model. No exceptions, including for a rush listing.
- Image disclosure. Any listing image that was virtually staged or altered beyond lighting and color correction gets labeled, and the unaltered original stays available. Check your MLS rule, because most already require this.
- No legal or lending output. The model does not write contract interpretations, fair housing language, or qualification statements.
- Log it. When AI drafts something client facing, note the tool in the transaction file. One line is enough.
- Ninety day review. Every quarter, re-read the tool list and cancel what nobody uses.
Item seven is the one everybody skips and the one that pays. If a complaint arrives eighteen months from now, a single dated line in the transaction file is the difference between a documented process and a shrug. Regulators treat those two things very differently.
Vendors Ship Faster Than Your Rules Do
The tooling is not waiting for you to catch up. Inman named Lofty its official AI partner in August, wiring agentic capability into the new Inman 2.0 platform so the trade publication becomes something agents run decisions through rather than something they read. RealAnalytica launched Atlas Agents, an always-on workflow layer that watches listings and transactions across CRM, MLS, tax data, and e-signature with more than 30 integrations. Zillow keeps widening the beta on its AI mode, which now books tours inside the conversation. Compass is pushing its Home Platform across the brands it absorbed in the Anywhere merger.
Inman ran a panel on September 4 built around what the hosts called the cheesecake test: does this tech actually earn the calories. It is the right question and it is incomplete. A tool can pay for itself in hours saved and still create exposure nobody priced. An always-on agent that touches your MLS data and sends client email is exactly the kind of tool that needs a named human in column four of your inventory.
Here is the signal worth watching over the next 30 days. Inman merged its Best of Proptech and Best of AI awards into one program with 25 categories, nominations open August 31 through October 1 and winners announced October 27. One of those categories is responsible deployment. When a trade publication carves out a standalone award for supervision, the market has decided that guardrails are a selling point rather than a tax. Vendors will start marketing their audit logs. Ask for them.
None of this requires you to slow down your adoption. It requires you to write down what you already run, name who reviews the output, and sign the page. That is a Tuesday morning task, not a project. Do it before the tool list grows again.
At Atlas Unchained we build the sites, automations, and operating docs that let small teams move fast without leaving holes behind. If you want a second set of eyes on your stack, start here. And if you want the week’s operator signal in one email every Friday, subscribe to The AU Brief.
