Ten of Fifty People Were Uploading MLS Data Into Claude. Write the Job Description Your AI Doesn't Have.

Ten of fifty people in one brokerage office were quietly uploading MLS data into Claude. Their broker did not know. Their MLS could not have known. And the only governance instrument anyone had was a signed piece of paper sitting in a filing cabinet.

Art Carter, CEO of California Regional MLS, told that story on stage in Dallas this week, and it is the most useful sentence spoken in real estate technology in some time.

"They are uploading MLS data into Claude with no governance whatsoever."

He was not describing a security incident. He was describing a Tuesday. Which is the point.

The Week Permission Became the Product

Something structural happened between August 10 and August 12, and almost nobody framed it correctly. Four separate companies shipped the same thing under four different names.

Model Context Protocol is the open standard letting an AI assistant reach into outside systems. It is plumbing, and plumbing is boring, and boring is exactly how infrastructure wins. Holden Page, HousingWire's SVP of technology, put it plainly on stage: the protocol is "basically just the way that we describe our data to AI and helps you query it without having to go through a data scientist." He then had Claude produce a Minneapolis-St. Paul market analysis with a dashboard attached, live, without anyone writing SQL.

Casey Williams, general manager of global mortgage at nCino, described the design constraint that matters more than the demo:

"We're building for a world where lending teams state intent and the system acts, replacing clicks with commands and dashboards with answers. Admins can connect the AI agent they already use directly into the nCino Mortgage Suite, and it only ever acts within their existing permissions, with every action logged."

Read that last clause twice. Only ever acts within their existing permissions. Every serious vendor arrived at the same conclusion in the same 72 hours: the interesting part of an AI employee is not its intelligence. It is its scope of authority.

Which raises an awkward question. Your AI is presumably already working. Where is its job description?

The Document Nobody Has Written

Every human you have ever hired came with an offer letter, a role, a reporting line, a spending limit, and an implicit list of things they must escalate rather than decide. Most AI deployments in residential services have none of these. They have a prompt.

A prompt is not a job description. A prompt is a vibe with punctuation.

Tim Dain, CEO of NorthStar MLS, has been working with Carter for roughly sixteen months on making policy machine-readable rather than paper-readable, and he offered the cleanest diagnostic test in the industry right now:

"First understand what data is being accessed. What use case is that data producing? Should it be governed and controlled? And how do you deploy it safely? Those are your first answers. And if you can't answer those questions, then you don't have an AI strategy yet."

He also said the sentence that should be printed on the wall of every brokerage, title shop, and property management office in the country: "Governance makes innovation safe enough to scale." Not slows. Not constrains. Enables.

What follows is the artifact that does not exist yet. Call it a Scope of Authority, or SOA. It is six sections, it fits on two pages, and you should write one for every Oppy or AI employee you deploy before you deploy the next one.

How To Write a Scope of Authority in Six Sections

1. Jurisdiction: the systems and fields, not the app

Do not write "access to the CRM." Write which objects, which fields, read or write. nCino's model is instructive: higher-impact actions such as archiving a loan officer or changing a branch structure can be configured to require human confirmation, while routine pipeline lookups do not. Dain's team is pushing entitlements down to the field and record level for exactly this reason, because role-based access built for humans in 2015 does not describe an agent acting on behalf of a human in 2026.

Prompt it directly. Give your agent a standing instruction in this shape:

You may read contact records, listing data, and transaction milestones. You may write notes, tasks, and appointment records. You may not modify commission fields, closing dates, or any field marked confidential. If a task requires a field outside this list, stop and ask.

2. Consent basis: the precondition, not the afterthought

Every outbound touch needs a documented basis before it happens, not a defense assembled afterward. The reason is sitting on a Florida docket right now.

On August 11, a federal magistrate in the Middle District of Florida ordered TCPA defendants to appear in person at the Ocala courthouse on August 24 at 10:00 a.m., with a laptop, and produce unfiltered CRM search results in real time for a full four-year period. The court's stated reason, per TCPAWorld's write-up, was that a produced CRM printout looked suspiciously identical to an earlier one:

"Given the inexplicable similarities between the recent CRM search printout and the Original Search printout, and Defendants' failure to offer a satisfactory explanation, the Court will conduct an evidentiary hearing regarding the completeness of Defendants' dialer-record production."

Courts are no longer accepting your summary of your records. They are asking to drive. Separately, an Ohio federal court ordered $83.7 million paid within ten days in a state attorney general robocall action, with a permanent ban on outbound marketing attached. The enforcement environment is not softening. Write the consent rule into the agent, not into a policy PDF.

3. Escalation triggers: the list of things it must not decide

This is where most deployments are thin, and where regulators are getting specific. Materials prepared for the NAIC's Big Data and Artificial Intelligence Working Group, which met August 13 in Columbus, include an AM Best presentation stating that agentic AI requires stronger controls involving permissions, activity logs, human review, and the ability to stop or reverse actions.

Permissions. Logs. Review. Reversal. That is a specification, delivered by an insurance ratings agency, and it applies just as neatly to a title company as to a carrier.

Brooke Anderson-Tompkins, founder and CEO of Bridge AIvisory, gave the geometry of it at the same Dallas summit: she pictures concentric circles narrowing to a core, and "the deeper that you go, the higher the guardrails need to go to be put into place along with it." Agentic sits at the core. Guardrails scale inward, not outward.

Your escalation list should name specific human beings, not departments. Anything touching price advice, legal interpretation, loan terms, or a licensed opinion routes to a person with a license and a name.

The legislative backdrop is moving in the same direction, fast. Hinshaw's compliance team counted nearly 100 chatbot-specific bills introduced across 34 states in 2026 alone. Colorado's Department of Law published draft rules on automated decision-making technology and conversational AI services on August 11, with a public hearing set for October 26. Connecticut's transparency act begins phasing in October 1. None of this requires you to predict which version survives. It requires you to know what your agent does, so you can answer when asked.

4. Licensure boundary: the gray area nobody has mapped

Carter raised the issue most likely to become 2027's expensive surprise:

"AI is increasingly inserting itself into licensed activity, probably on the mortgage, the origination side and on the real estate side. And, you know, that's just the gray area that most states have no clue how to deal with."

California's Department of Real Estate has indicated brokers are responsible for their use of AI. That is a clean answer to the liability question and no answer at all to the practical one. So draw the line yourself, in writing, before someone draws it for you. Your agent qualifies, schedules, follows up, prepares, summarizes, and reminds. It does not advise, opine, quote terms, or interpret contracts. The gap between those two lists is your license.

5. Cost ceiling: the constraint nobody budgeted for

Here is the least glamorous and most immediately useful number of the week. Pritesh Damani, CTO of The Real Brokerage, explained why Leo 2.0 is rolling out selectively rather than all at once, and it has nothing to do with readiness:

"If I threw in all the agents on day one, we would probably spend a [over] $100,000 a day."

Compute is a real operating expense with a real invoice. Kevin Foley, director of AI Labs at Optimal Blue, offered the single takeaway he wanted lenders to leave with:

"If you don't have a dashboard today where you can go and see how much I spent on my tokens in the last 30 days or seven days or whatever, that's maybe [the] No. 1 thing you can take away from this."

Put a monthly ceiling and an alert threshold in the SOA. An AI employee with an unbounded budget is not an employee. It is a subscription with ambition.

6. Audit surface: what you can produce in ten minutes

Every action logged with content, timestamp, and the authority it acted under. nCino logs the action, the timestamp, and the outcome. That is the standard, and it exists because someone will eventually ask, and the question will arrive with a case number.

Dain's warning about what happens to an industry that skips this step deserves its own line:

"We're in a position that if we don't do this, we're going to be buying back the intelligence created from our own data."

What Happens When You Actually Do This

The objection to all of this is that governance is a tax on speed. The week's evidence says the opposite.

David Keener, CEO of Premiere Group at The Real Brokerage, uploaded roughly 3,000 dead leads into a properly scoped agentic system and described the result with an accountant's precision: 1,992 conversations in thirty days, 444 of them live simultaneously, and an inside sales function that used to cost him more than $100,000 a year now gone as a line item. His summary of the change is the best available quote on adoption psychology:

"It's like living in the 1800s when they first turned on electricity."

Damani's forecast is more aggressive than most and worth holding onto: "I do think we're six months away from like 80% of the processes being completely automated." He pairs it with the important caveat that trust and relationship work remain human, because "when you're making a million-dollar purchase, most people don't know what they're doing."

In mortgage, Pavan Agarwal, founder and CEO of Celligence and CEO of Sun West Mortgage, described the end state without hedging in National Mortgage Professional's deep dive: "We originate hundreds of retail loans every month through our distributed retail channel. We don't have any loan processors. AI does literally everything." His colleague Angel Sumlin identified why the economics work: "What AI is giving you for the first time is pretty much a variable cost."

And in the back office, the SetWave and Inside Real Estate launch targets a problem every transaction coordinator will recognize instantly. A meaningful share of transaction files arrive with at least one error or omission, and the review that catches it often finishes days or weeks after closing. Joe Skousen, CEO of Inside Real Estate, described the destination as "a brokerage where an agent can open a transaction, prep the paperwork, and get it signed just by asking." That only works if the agent knows precisely what it is allowed to sign. Scope again.

The Vibe-Coding Warning You Should Take Seriously

One counterweight, from someone who spent this week at Ai4 in Las Vegas among 12,000 attendees. Kevin Hawkins, writing in WAV Group's REAL AI newsletter, described a CVS Health executive who built 400,000 simulated consumer agents as a "flight simulator" for testing programs before they touch a real patient. Fascinating, ambitious, and Hawkins attached exactly the right caution:

"This is not something an independent agent should build in Claude Code over a weekend after watching three YouTube videos."

He is right, and the reason connects back to Carter's ten out of fifty. The failure mode is not that people build bad agents. It is that people build ungoverned ones, quickly, invisibly, from inside your organization, using your data, without telling you. The fix is not prohibition. Carter said so himself: "I don't have to control the sandbox, but I do need to make sure that I do provide those opportunities for our brokers and agents to successfully do what it is that they're going to be doing in this new world."

Give people a sanctioned, scoped, logged way to do the thing they are already doing. That is the whole strategy.

Your Next Ninety Minutes

Pick your highest-volume repetitive workflow. Inbound call triage at a dental office. Renewal outreach at an insurance agency. Milestone updates at a title company. Dead-lead reactivation at a brokerage. Prior authorization chasing at a medical practice.

Write the six sections. Jurisdiction, consent basis, escalation triggers, licensure boundary, cost ceiling, audit surface. Two pages, plain language, one named human per escalation path.

Then hand it to your agent as its standing instruction rather than storing it in a policy folder. In Oppy, that document is the configuration: an AI employee with a defined role across sixty-plus connected business tools, working inside stated limits, logging what it did and why. The SOA is not paperwork wrapped around the product. It is the product specification.

Tyler Morton, broker-owner of REMAX Victory + Affiliates and founder of TroyOS, learned the adoption lesson the expensive way when his first internal build went unused: "It didn't fail because of AI at all. It actually failed because it wasn't something that anyone wanted. I was solving a problem for me and not for them." His advice for starting is the correct advice for this exercise too. Solve the easy problems first. Find the hour a week.

Everyone else is currently arguing about how smart the models are. The firms that will look prescient in eighteen months are the ones writing down what their models are allowed to touch. That document takes an afternoon, and right now, almost nobody has one.

Ten out of fifty. In one office. On an ordinary Tuesday.

References

1. MLS leaders say AI requires a new approach to governing real estate data, HousingWire, August 12, 2026
2. HousingWire expands AI strategy to housing data and news, HousingWire, August 11, 2026
3. nCino launches Mortgage MCP for lenders' AI agents, CFOtech US, August 10, 2026
4. Inside Real Estate Advances AI Strategy With Launch of ComplianceAI and Direct AI Assistant Integration With BoldTrail BackOffice, RISMedia, August 11, 2026
5. RealPage Introduces the Lumina AI Suite, Business Wire, August 10, 2026
6. Court Sets In-Person Evidentiary Hearing to Examine TCPA Defendant's CRM Records, TCPAWorld, August 12, 2026
7. Aaron Michael Jones Ordered to Pay $83,716,500.00 Within 10 Days, TCPAWorld, August 12, 2026
8. NAIC Summer National Meeting to Highlight Innovation and Resilience, NAIC, August 11, 2026
9. Real estate leaders weigh buying vs building AI tools, HousingWire, August 12, 2026
10. The next phase of The Real Brokerage's AI is replacing the work, not the agent, HousingWire, August 12, 2026
11. Optimal Blue's Kevin Foley discusses AI Labs, Virtual Economist rollout, HousingWire, August 12, 2026
12. NMP Deep Dive: The AI Workforce, National Mortgage Professional, August 11, 2026
13. REAL AI: What CVS could teach real estate, WAV Group, August 10, 2026
14. The Agentic Leadership Playbook: The COO as Enterprise Orchestrator, BCG, August 12, 2026
15. Colorado Publishes Draft ADMT Regulations, Consumer Finance Monitor, August 12, 2026
16. 2026 AI Compliance: Upcoming Laws Every Organization Needs to Know, Hinshaw & Culbertson, August 5, 2026
17. Ricky Carruth on changing the infrastructure of real estate, HousingWire Daily podcast, August 12, 2026