What Data Can Your CRE AI Agent Use? 4 Sources Compared

By Jude Lee · · Custom

Commercial real estate brokers reviewing deal documents and a pipeline dashboard on a laptop in a conference room

The model is the easy part. The permissions are the project.

Ask ten brokerage operations leads why their AI pilot stalled and you’ll get versions of the same answer: they picked an assistant before they decided what it was allowed to see. The assistant is interchangeable. The data layer is not.

There’s an architectural pattern worth borrowing here, and it shows up in brokerages of every stripe once the first pilot stalls. Instead of standardizing the whole firm on one chat product, you expose your systems through the Model Context Protocol — MCP, an open standard for giving an AI governed access to your data and tools, originally published by Anthropic — and let people bring whichever assistant they prefer. Adoption has spread beyond Anthropic’s own Claude apps to other assistants and developer tools, but which client supports what moves quickly enough that you should check each vendor’s current documentation rather than trust any dated claim, including this one. The durable asset isn’t the chat window. It’s the governed data layer underneath it.

So the first question isn’t which assistant to buy. It’s what you can safely point one at. Four categories, roughly in order of how much control you have:

  1. First-party data — your CRM, email, deal files, closed comps. You set the rules.
  2. Public records and open datasets — permissive to use, but inconsistent and often lagging.
  3. Licensed market data — the deepest coverage, the tightest contract terms.
  4. The open web — broad and free-looking, brittle and frequently off-limits in practice.

Source 1: Licensed market data

CoStar, Crexi, Reonomy, Placer.ai, Moody’s/REIS-style products — this is the data your brokers already live in, and it’s the data you’re least free to shove into an AI pipeline.

Subscription agreements for commercial data products typically govern redistribution, derivative works, seat sharing, and automated extraction. Terms vary by vendor, by product line, and by how your contract was negotiated, so nothing generic you read online (including here) is a substitute for reading yours. The practical rule of thumb: the fact that a licensed human can view a record in a browser does not mean an agent may harvest it. If you’re planning anything beyond a person reading a screen, get your subscription agreement in front of counsel before you build.

The clean path, where it exists, is the vendor’s own API or MCP integration under a data-access agreement — check the official product documentation rather than assuming. Some platforms are shipping their own AI features precisely so you don’t need to export anything; whether those features do the job you need is a separate evaluation.

Source 2: Public records and open government data

County assessor, recorder and clerk systems; municipal permit and zoning portals; and federal statistical APIs. The U.S. Census Bureau publishes machine-readable data including the American Community Survey through its public developer APIs, and the Bureau of Labor Statistics offers a public API covering series such as employment and wages — both are genuinely usable inputs for a submarket snapshot or a trade-area narrative.

The catch is unevenness. Recording lags vary by county. Owner names arrive as inconsistent entity strings. Some portals offer a clean bulk download; others are a session-based search form that explicitly prohibits automated access.

An agent can handle the parsing and normalization well; it cannot fix the fact that three adjacent counties publish three different things. Where this shines: an agent that assembles the repeatable, defensible parts of a market section — population and household income from ACS, employment by sector from BLS, recent transfers from the recorder — and flags anything it couldn’t verify. A human still writes the judgment.

Source 3: Your own first-party data

Your CRM, your email, your closed-deal files, your rent rolls, your tour notes, your historical comps. This is the data nobody else has, the data you’re not licensing from anyone, and — in our opinion — the correct place to start most brokerage AI builds.

It’s also the data that’s in the worst shape. Deal stages nobody updated after a tour. Four spellings of the same tenant. Lease abstracts living as PDFs in a folder named “Final_v3.” That matters for sequencing: a firm whose CRM is thin, abandoned, or split across three half-migrated systems is almost always better served by a data cleanup plus whatever reporting or AI feature its existing platform already ships, and revisiting a custom build later. Pointing an agent at bad records produces confident, wrong answers faster.

Where the records are reasonably clean, this is exactly the job a custom MCP server does well: you define a small set of tools — search_comps, get_listing, get_deal_timeline, find_expiring_leases — over your own systems, with read-only scopes, per-user permissions and a log of every call. The assistant then answers portfolio questions without anyone exporting anything. We walk through the build in connecting Claude to your CRE systems via MCP.

Source 4: The open web and portal listings

Scraping broker websites and listing portals looks like the fastest route to coverage. It is also the most brittle and the most likely to collide with terms of service, robots.txt directives, and rate limiting. Page structures change without warning; an agent that silently returns stale or half-parsed results is worse than no agent.

A narrower version is defensible: an assistant with browsing that reads one page a human explicitly points it at, the way an analyst would. Mass harvesting is a different activity with a different risk profile — treat it as a business decision with legal input, not a weekend automation.

Licensed depth vs. data you control

Licensed market data
Broad coverage you could never build yourself. Standardized fields. Already trusted by your brokers and clients. But: contract terms constrain what an agent may ingest, API access is a commercial negotiation, and you’re renting the asset. Best used through the vendor’s sanctioned interface.
Your first-party data
Unique, defensible, and yours — closed comps, deal history, relationship context no subscription contains. But: inconsistent, incomplete, and spread across CRM, inbox and file shares. Requires cleanup and a permissions model before an agent is useful. Compounds in value every year you keep it clean.

How to decide what to wire up first

  1. Inventory the systems, not the ideas

    List every place deal data lives: CRM, email, shared drive, data room, accounting, marketing platform. Note the owner and whether an API exists.

  2. Classify rights in a simple traffic-light matrix

    Green = you own it. Yellow = licensed, check the agreement. Red = someone else’s, under NDA, or scraped. One row per source, one named owner per row. Have counsel review the yellow rows.

  3. Pick one question worth answering

    Not “AI for the brokerage.” Something like: “Which of our managed-property leases expire in the next 18 months, and which tenants haven’t been contacted in 90 days?” That question tells you exactly which two systems to expose.

  4. Expose it read-only, scoped, and logged

    Start with retrieval before actions. Read-only tools, per-user permissions mirroring your existing access control, full call logging. Add write actions (CRM updates, draft emails) only after the reads are trusted.

  5. Measure it with your own numbers

    This step produces no benchmark figure — it’s a worked example you fill in. Before: time how many hours a week your team actually spends assembling this answer. Then apply the formula loaded hourly rate × hours × 50 weeks as an assumption-driven estimate; every input is yours, including the 50-week figure, which you should adjust for your own calendar. After: re-measure the same task, and separately track what the recovered hours got reallocated to — more tours, more calls, faster OM turnaround. Recovered hours alone understate the impact if the time goes back into revenue work, and overstate it if the time just evaporates.

What agents actually do once the data is connected

With a clean data layer, the useful applications are unglamorous and specific: pipeline stages updating themselves from email threads, lease and rent-roll terms abstracted from PDFs into structured fields, a draft offering memorandum assembled from the data room, follow-up emails written the same way every time.

None of this is magic. Extraction from scanned, hand-amended leases still needs a human check. Anything going to a client needs a named reviewer. And plenty of tasks — a recurring report with fixed inputs — are better served by a deterministic script or an existing SaaS feature than by an agent that might phrase it differently on Tuesday.

Where AI compresses brokerage labor — and where it doesn’t

Our opinion, stated as opinion: AI compresses the document-handling and analysis labor in brokerage faster than it touches the relationship and judgment layer. The firms that gain are the ones whose proprietary data — their comps, their deal history, their tenant relationships — is structured enough for an agent to use. The ones that don’t will keep paying for subscriptions and keep doing the assembly work by hand.

Your subscription data is rented. Your deal history is owned. Build the agent over the thing you own first.

And if the honest answer after the inventory is “our CRM is too messy to expose, and the off-the-shelf feature in our listing platform already covers most of this” — that’s a legitimate outcome, and worth testing against the actual question rather than assuming either way. Fix the data, use the vendor feature, and revisit the custom build when the question you need answered stops fitting in someone else’s product. We lay out that trade-off in vendor vs custom MCP servers.

Related reading: CoStar vs Crexi vs Reonomy · shadow AI governance in brokerages · a task-by-task map of AI in CRE · rent roll and lease abstraction automation

Not sure where to start?

Get a free automation audit: we map your deal pipeline, marketing, and back-office workflows and show you what's worth automating — before you spend a dollar.

Get a free automation audit

Related guides