Can Your AI Agent Use CoStar or MLS Data? 4 Options

By Jude Lee · · Comparison

Two CRE brokers reviewing listing data on monitors alongside a printed data subscription agreement

The shift that’s forcing this question

Two storylines are running in parallel across the industry as of 2026. MLSs, exchanges and listing platforms are issuing guidance on AI bots — what counts as permitted automated access, who may hold credentials, and what the bylaws actually say when a vendor points a script at a listing system. At the same time, capital is flowing toward CRE lending and brokerage startups whose entire pitch is agent workflows rather than dashboards.

Different corners of the industry, same underlying shift: agents are no longer summarizing documents you hand them. They’re reaching into systems. And in CRE, the most valuable systems — CoStar, Crexi, Reonomy, your regional MLS or exchange — contain data you rent, not data you own.

Separate what you own from what you rent

Before you evaluate any tool, split your data into two buckets. The rules are completely different.

Data you own
Your CRM records, email threads, signed LOIs and leases, closed-deal comps, rent rolls your clients gave you, tour feedback, your own listing photos and OMs. You can put this in front of an AI agent subject to client confidentiality and NDAs — not a vendor’s terms of service.
Data you license
CoStar/Crexi/Reonomy records, MLS feeds, third-party photos, aggregated comps, tenant databases. Your access is governed by a subscription agreement that typically limits redistribution, bulk extraction, credential sharing and automated access. Read yours.

The practical consequence: the highest-value, lowest-risk agent projects in most brokerages run on the first bucket. An agent that can answer “which of our 140 tracked tenants have leases expiring inside 18 months, and when did we last touch the decision-maker?” doesn’t need a single licensed record. That’s the case for building a custom MCP server over your own CRM and listing data before you go anywhere near a vendor feed.

Our position, plainly: the fastest way for a brokerage to lose a data subscription is to let a bot log in with a broker’s password.

Four ways to put market data in front of an agent

1. Human copy/paste into a general assistant

You pull the comp set, export the PDF or paste the table into Claude or ChatGPT, and ask for analysis. A person is in the loop for every retrieval, which is usually what subscription terms contemplate (one licensed human, normal use).

Where it breaks: it doesn’t scale, nothing is logged, and two brokers will paste two different comp sets and get two different answers. Also confirm whether your assistant’s plan trains on your inputs — most business and enterprise tiers let you turn that off; check the vendor’s current documentation, since defaults change.

2. The platform’s own AI features

CoStar, Crexi, Buildout and the major CRE CRMs have all shipped AI features. These are in-bounds by construction — the vendor is using its own data under its own terms.

The limit is scope. Vendor AI answers questions inside the vendor’s walls. It can’t join their market data to your pipeline, your commission splits or last Tuesday’s tour notes. For a single-system job, that’s fine and often the cheapest correct answer.

3. A sanctioned feed into your own environment

Some subscriptions include — or can be upgraded to include — an API, data feed or permitted export. You land that data in your own database, then expose it to an AI assistant through an MCP server alongside your CRM. MCP (the Model Context Protocol, an open standard for giving an assistant governed access to tools and data) is the plumbing; the entitlement is the hard part.

This is the only path that gives you governed, auditable, cross-system answers — and it is also the slowest. Plan in weeks, not days: a procurement conversation, a redlined amendment, then integration and field mapping work on your side. The usual blockers, in our experience of how these contracts are structured, are predictable. Feed entitlement is often priced per seat or per market, so a firm-wide agent costs far more than the single login you budgeted against. Field-level restrictions can strip exactly the fields you wanted — ownership contacts, historical comps, photography. And redistribution clauses may limit who inside the firm can see agent output derived from the feed, which quietly turns your “everyone can ask the assistant” plan back into a licensed-seat problem. Ask your rep three questions in writing: is automated access permitted, what is the field-level scope, and may an AI system process the data.

4. A credentialed bot that logs in as you

A headless browser or scraping agent uses a broker’s login to pull screens on demand. It’s the fastest to stand up and requires no vendor conversation — which is exactly why it shows up in shadow-AI audits.

It is also the path that puts your subscription, and potentially your firm, at risk: credential sharing with an automated process is a common contract violation, MLS participation rules frequently address it directly, and major CRE data vendors have publicly pursued litigation over unauthorized use of their listing content and photographs. If someone in your firm is already doing this, treat it as a governance problem, not an IT curiosity — see shadow AI governance options.

So which AI tool is “best” for commercial real estate?

There isn’t one, and anyone selling you a single answer is selling. The useful question is which tool is best for a specific job and a specific data bucket. Generalist assistants are strongest on drafting and reasoning over documents you provide. Platform AI is strongest inside its own data. Custom agents win when the answer requires joining systems nobody sells together — your CRM, your email, your deal files. We mapped job-by-job fits in how AI helps commercial real estate.

Know the failure modes you’ll actually hit on market data. Assistants misread rent rolls with confidence — collapsing a “current rent” column into “base rent,” or treating a recovery line as income — and the summary reads fluently either way. They will surface comps without a vintage date, so an 18-month-old sale lands in your analysis as if it closed last quarter. And unit-of-measure errors are endemic: NNN quoted against gross, RSF against USF, annual against monthly, with no flag that a conversion happened. Any number that goes into an LOI, an offering memorandum or an IC memo needs a human to re-source it from the original document.

Rules of thumb that don’t survive contact with an agent

Two phrases come up constantly in search, and both deserve a plain answer.

The 2% rule — monthly rent should be at least 2% of purchase price — is a residential small-investor screening heuristic. It has essentially no application to institutional or mid-market commercial underwriting, where cap rates, debt service coverage, lease term and credit drive value. If an AI assistant volunteers it on a CRE deal, that’s a signal it’s pattern-matching on consumer content rather than your underwriting standards.

The 30% rule for AI has no authoritative definition. You’ll see it used to mean “let AI do about 30% of the work” or “expect about 30% of output to need correction.” Neither is a measured industry standard. Replace it with something you can actually run: sample a fixed share of agent output every week, log the error rate you observe in your own files, and set review intensity from that.

A short data-license review before you automate

  1. Inventory the systems

    List every data subscription the firm pays for and who holds credentials. Include MLS or exchange memberships.

  2. Pull the actual terms

    Find the clauses on automated access, bulk extraction, redistribution, credential sharing and third-party processing.

  3. Ask the vendor in writing

    ”May an AI system process data retrieved under this subscription, and is API access available?” Keep the answer on file.

  4. Build on owned data first

    Stand up the agent over your CRM, email and deal files, where you control the rules.

  5. Add licensed data only through a sanctioned path

    Feed, API or permitted export — never a bot with a human’s password.

Modeling what it’s worth, honestly

Don’t buy a dollar headline from anyone, including us. Model it with your own inputs:

(hours per week a broker or analyst spends pulling, reformatting and reconciling data) × (loaded hourly cost) × 52 — then subtract build, feed and subscription cost, and separately ask what the recovered hours get reallocated to. Hours moved from retrieval into client contact are worth more than hours merely saved. Measure the input for two weeks before you believe any of it.

6 hrs/wk × $85 ≈ $26k/yr
Worked example only — substitute your own measured hours and loaded rate, then net out build and feed cost
Illustrative calculation, not a benchmark or observed result

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