Real Estate & Property: what to automate, what to keep human.

Adoption is not the question in real estate anymore. In the National Association of Realtors 2025 Technology Survey (a random sample of 49,233 active Realtors, fielded July 2025), about 68% of agents already use AI. But only 20% use it daily, only 17% report a significant positive business impact, and 46% see no noticeable difference, with most of it general-purpose ChatGPT (58%) writing listing descriptions and captions.

That gap, everyone uses it, almost nobody closes more because of it, is the whole story. The cosmetic work AI is aimed at is not where deals are won or lost. Deals leak in the minutes before a lead gets a call back, in the follow-up that quietly dies after the second attempt, and in the transaction-coordination drag that turns a producing agent into an unpaid project manager. None of that is what the demo shows you, and all of it is where the commission actually lives.

So the question a serious broker asks is not 'which real-estate AI tool.' It is 'which workflow, fixed, puts more deals through the office, and is AI actually pointed at it.'

The revenue-critical workflows in a brokerage are well known: speed to the inbound lead, persistent follow-up across a long buying cycle, and the coordination work that keeps a deal moving to close. These are repetitive, time-sensitive, and currently lost to whoever is busy. They are exactly what AI is good at accelerating, and exactly what it is rarely deployed against.

Start where a mistake is cheap and the gain is visible: instant lead response and callback, structured follow-up that does not depend on an agent remembering, deal-coordination reminders, document checklists, and draft client updates a license-holder sends. None of these put an unreviewed number or claim in front of a buyer or a lender, and each one is measurable in booked appointments, response time, and deals that did not stall.

Prove the gain on one of these before spreading AI across the office. The point is not more AI in the brokerage. It is one workflow where AI demonstrably moved a number a broker cares about, with an owner and a measure, before the next one.

There is exactly one category where speed is not the goal: any output that carries the license. An AI-assisted valuation that is off, listing or advertising language with steering-adjacent wording, or an AI tenant-screening decision is not a productivity item, it is a fair-housing or valuation exposure with a license attached, and a fast wrong answer there is a regulator, not a rewrite.

This is not a judgment call anymore, it has dates on it. The federal Quality Control Standards for Automated Valuation Models rule (OCC, Federal Reserve, FDIC, NCUA, CFPB, and FHFA, under Dodd-Frank) took effect October 1, 2025 and requires AVMs used in covered credit decisions to meet quality-control standards including testing and compliance with nondiscrimination law. HUD's May 2, 2024 Fair Housing Act guidance states plainly that the Act applies even when AI performs screening or advertising, and the housing provider stays responsible. The AI does not absorb the liability. A named person does, or no one does.

So the leash is narrow and specific, and it exists to protect the deal flow you just sped up, not to slow it down. A licensed person signs off any valuation or CMA before it informs a price or a lender. A human reviews AI-generated listing and ad language against fair-housing rules before it publishes. Screening criteria are human-set, and AI never makes the adverse decision alone. AI drafts and computes everywhere else. It owns nothing that carries the license.