How commercial real estate brokers use AI

Where does AI save a broker the most time?

On work that is repeated, rule-bound and sourced from records you are allowed to use. That is most of the research and paperwork around a deal, and very little of the deal itself.

How do brokers use AI to prospect owners?

By asking questions of the county roll that the county website cannot answer. Most county assessors publish the owner of record, the tax-bill mailing address, a use classification, building size, an assessed or market value and at least the last sale for every parcel; what else they publish, and in what format, varies by county and state. Loaded into a database an AI assistant can query, that becomes one question: industrial buildings between 20,000 and 60,000 square feet in one county, held more than ten years, owned by an LLC with an out-of-state mailing address.

The second step is the person behind the entity. Each state's corporate registry lists an LLC's registered agent, and many list its managers or officers too, so the model can go from the name on the deed to the names on the filing, and from there to everything else those names own. Some filings name only a registered agent or a law firm, and an LLC formed in another state may show nothing useful in yours, so treat an unmatched name as unknown rather than as a finding. In Florida the registry is Sunbiz; this is the work our county records connector does inside Claude across Hillsborough, Pinellas, Pasco and Polk.

Keep it to sources you are allowed to automate. Licensed platforms usually restrict automated access and AI use in their terms; Can Claude connect to CoStar? covers why, and what to do instead.

How does AI help with BOVs, OMs and comps?

It assembles the evidence and drafts the prose, and the broker supplies the judgment. A broker opinion of value is mostly gathering: the parcel record, the sale history, a screened set of comps, and a layout in the firm's style. mortr, a CRM we built for CRE teams, generates that document in the firm's branding from county records and leaves the conclusion to the broker.

An offering memorandum draft works the same way. Given a rent roll, photos and the broker's notes, a model can write a first pass of the property, location and tenancy sections. Expect to rewrite the parts that sell. For comps, the value is in the screening: removing $100 deeds and multi-parcel portfolio sales, and stating the file date so a thin recent month is not read as a slowdown. A reusable Skill is the right place to write those rules down once.

How can AI keep a CRM, market updates and follow-up current?

By running the same small checks every week without anyone remembering to. Three patterns work well:

• CRM hygiene. Connected to the CRM, a model can list contacts with no activity in 90 days, deals with no next step, and duplicate companies spelled three ways. The broker decides what to merge. See Claude in CRE: Connectors.
• Market updates. A scheduled task can pull the week's recorded sales in a submarket each Monday and draft a short summary. Read it before it goes anywhere.
• Follow-up. After a meeting, the model drafts the follow-up email and the CRM note from your notes. It drafts. You send.

For file-heavy work such as rent rolls and underwriting models, Cowork, which Anthropic began folding into the main Claude app on September 16, 2026, starting with Pro and Max, lets Claude work inside an approved folder on your computer instead of bouncing files through a browser.

Where do personal AI agents like Meta's Muse fit?

On the broker's own marketing, not on firm data. A personal agent works for one person across that person's accounts. Meta launched Muse in the US on September 8, 2026, with a free tier and paid plans; it runs tasks in a cloud browser you can watch. Instinct, a similar assistant reached by text message, is still invite-only as of September 2026.

The useful job is the one a broker already does by hand: turning a weekly submarket summary or a closed-deal announcement, drafted and checked in Claude, into posts for the Facebook and Instagram accounts and neighborhood or business groups the broker already works. Two limits apply. Meta does not document direct posting to Facebook or Instagram as of September 2026, so test what it will do in your own accounts before you rely on it. And a consumer agent is not a place for client files, owner lists or licensed data unless your firm has approved it. Read every post before it goes out, and each group's rules on promotion. Personal AI agents for business covers the category.

What should a broker not hand to AI?

Anything where being wrong costs a client or breaks a rule. Specifically:

• The value conclusion and the pricing recommendation. They carry your license and your name.
• Advice to a client about whether to sell, what to accept, or how to structure a deal.
• Licensed data your subscription does not allow in AI tools. Read the terms before pasting a screen into a chat.
• Confidential client material in a consumer account without the data controls your firm has approved.
• Unreviewed sends. Nothing goes to a client, a counterparty or a mailing list without a person reading it.
• Legal, tax and environmental conclusions, which belong to the professionals who sign for them.

What does a 30-day AI pilot look like for a brokerage?

Pick one workflow, one owner and one measure, and run it on real deals for four weeks.

1. Week 1: choose and map. Pick the workflow that is most repeated, often owner research or BOV assembly. Write down every step as it is done today and how long it takes.
2. Week 2: build it by hand. Run the workflow with AI on three live properties. Save the prompts that worked. Note every correction you made.
3. Week 3: make it repeatable. Turn the prompts and corrections into a Skill or a checklist, connect the sources you are allowed to automate, and have a second person run it.
4. Week 4: measure and decide. Compare time per property and error count against week one. Keep it, change it, or drop it.

A pilot that ends in "drop it" is still a result. A pilot that never names its measure is not.

A starting prompt for week 2:

For the parcel at [address], list the owner of record, mailing address,
use code, building area, year built, just value and every recorded sale.
Mark any sale under $1,000 or any deed that conveyed more than one parcel.
Then list qualified sales of the same use code within [size band] in the
last 24 months, and state the date of the sales file you used.