Commercial Real Estate AI Agents: Connect Claude via MCP

By Jude Lee · · Custom

commercial real-estate brokers and brokerage operations teams using workflow automation software

The gap between an AI that talks about CRE and one that knows your deals

Ask a general-purpose assistant about cap rate compression in industrial and you’ll get a competent, generic answer. Ask it “which of my tenant-rep clients have leases expiring in the next nine months in the Dallas submarkets I cover?” and it has nothing — because that answer lives in your CRM, a folder of abstracted leases, and someone’s spreadsheet.

That’s the whole problem. The model isn’t the bottleneck; access is. An AI agent — as opposed to a chatbot — is an assistant that can call tools, take multi-step actions, and come back with a result: query the CRM, pull the matching leases, cross-reference the marketing calendar, draft the outreach list. It can only do that if something hands it those tools.

What a custom MCP server actually is

MCP is an open protocol, introduced by Anthropic in late 2024 and documented publicly at modelcontextprotocol.io, for connecting AI assistants to external data and tools. A custom MCP server is a small piece of software your firm runs that exposes a handful of named capabilities — search_listings, get_deal, list_lease_expirations, find_comps — and returns structured data when the assistant calls them.

It is not a data lake, not a model, and not a rewrite of your systems. It’s a translator plus a bouncer: it speaks your CRM’s API on one side and the protocol on the other, and it enforces who’s allowed to see what. Client and vendor support for MCP varies and moves quickly — check the current documentation for your AI client and your CRM before assuming a given tool speaks it.

The Monday-morning questions that justify the build

The test for whether this is worth doing isn’t technical. It’s whether you can list five questions your team asks every week that currently require a human to open three systems.

A realistic first set for a mid-size brokerage:

  1. Which active listings have had no marketing activity logged in 21 days?
  2. Which of my client’s leases expire in the next three quarters, sorted by square footage?
  3. What did we last discuss with this principal, across email and CRM notes?
  4. Which deals in the pipeline have a closing date in the past and no updated status?
  5. Show me every industrial comp we’ve touched in this submarket in the last 18 months.

Each maps cleanly to one MCP tool. A broker asks in plain language; the agent picks the tool, passes parameters, and reports back with the underlying records cited so the broker can verify. That last part matters more than anything: an agent that returns a list without the record IDs behind it is unauditable, and unauditable output is how confidently wrong answers reach a client.

If the agent can’t show you the records it based its answer on, you don’t have an assistant — you have a rumor generator with API access.

Skills: teaching the agent to do the job your way

The second building block is skills — reusable, packaged instructions that teach an assistant to perform a specific job the same way each time (Anthropic documents its Agent Skills format publicly; other platforms use similar prompt-and-script bundles under different names).

For a brokerage, a skill is where your house style lives. “Quarterly tenant-rep check-in email” as a skill encodes: pull the lease expiry and last-contact date, use the firm’s three-paragraph structure, never quote a rate that isn’t in the CRM, always end with two proposed call times, flag for human review if the client is in an active negotiation. Written once, it produces output that follows the same structure and constraints whether the analyst or the managing director triggers it. Be precise about what that buys you: a skill constrains format, sourcing rules and guardrails — not exact wording. Language models vary run to run, so expect consistent shape and consistent rules, not identical sentences. That’s still the difference between AI as a novelty and AI as an operations standard — the same logic that makes CIM and offering memorandum production repeatable rather than artisanal.

Build it in the order that de-risks it

  1. Write down five questions, with known answers

    Pick five real questions and answer them manually today, recording the specific records behind each answer. Those answers become your test set. If you skip this, you’ll have no way to tell whether the agent is right.

  2. Inventory access, not features

    Can you reach your CRM through a documented API? Your document store? Your email? Data-provider licences frequently restrict redistribution and derivative use — read the terms and confirm with the vendor before piping licensed market data into an AI system.

  3. Ship read-only first

    Four to eight tools, read-only, scoped to the permissions of the person asking. A broker should not be able to ask the agent about a colleague’s confidential assignment. Entitlements belong in the server, not in the prompt.

  4. Evaluate before you expand

    Re-run the five known-answer questions weekly against current data, compare the record IDs the agent returns to the record IDs in your manual answer, and log every miss in one of two buckets: missing-record (it left something out) or fabricated-record (it returned something that doesn’t exist, or asserted a figure no record supports). The second bucket is the dangerous one. Our 10-point scorecard for vetting CRE AI tools works just as well on something you built yourself.

  5. Add skills for the repeatable outputs

    Follow-up emails, listing blurbs, tour recaps, weekly pipeline summaries.

  6. Only then allow write actions — with approval

    Updating a deal stage, logging an activity, creating a task. Every write gets a human confirmation step until you’ve watched it behave for a full quarter.

Choosing between off-the-shelf AI, a custom agent, and nothing at all

There is no single “best AI tool for commercial real estate,” and anyone who names one is selling something. The best tool is the one connected to the data your question depends on.

Start with the cheapest option: no agent. If a question’s data lives entirely in one system, a saved view, a scheduled report, or a Zapier-class rule answers it outright. Of the five questions above, numbers 1 and 4 are pure saved-filter-and-alert work inside most CRMs, and number 2 is too if your abstracted leases already sit in the same system. Only 3 (email plus CRM notes) and 5 (internal comps plus a data provider) genuinely cross boundaries. That’s the real shape of most brokerages: three of five need configuration, not AI.

Off-the-shelf AI features
Already sees the data inside that platform. No engineering, no maintenance, vendor handles security review. Buildout is built around listing marketing production — flyers, offering memoranda, email campaigns off listing records — so “produce and send collateral for this listing” is a question it answers inside its own workflow. Apto is a broker CRM whose native reporting answers “what’s in my pipeline and what activity did I log this month” without any extra layer. CoStar’s research platform is where “what comps and availabilities exist in this submarket” already lives, under a licence you’re paying for anyway. Verify current AI-specific features against each vendor’s own documentation. Weakness: it stops at the platform boundary, and you inherit the vendor’s roadmap.
Custom agent + MCP server
Spans systems your vendors will never connect for you, encodes your firm’s specific logic, and you control retention and permissions. Weakness: real build cost, ongoing maintenance when APIs change, security surface you now own, and it needs an internal owner. Rarely worth it for a team of five.

A useful heuristic — opinion, not measurement: build only when the question crosses two or more systems, gets asked weekly, and the answer has money attached. Everything else, buy or configure. We go deeper in custom vs off-the-shelf CRE software.

What this actually does to the broker’s job

The realistic effect of agentic AI is compression of the middle: the assembly, formatting, chasing and reconciling between origination and closing. You’ll see headlines claiming brokers free up huge percentages of their time; treat any specific figure as unverified unless you can read the methodology behind it.

What doesn’t compress is the part clients pay for. Sourcing an off-market opportunity, reading a landlord’s real position, knowing which principal actually signs, taking accountability when a deal goes sideways — none of that is a retrieval problem. Ops roles change more than production roles: less rekeying, more exception handling and system ownership. And on the “30% rule” people keep asking about: there’s no canonical definition behind that phrase. Our rule of thumb, stated plainly as opinion — automate the mechanical majority of a task, leave the judgment tail to a human, especially anything that quotes a rate, states a legal term, or goes to a client under your name.

Modelling the payback without inventing a number

Don’t accept anyone’s ROI headline, including ours. Model it yourself:

(hours per week on the task) × (loaded hourly cost of the person doing it) × 52 = annual cost of the status quo

Then estimate the share that genuinely disappears — not the share a vendor claims — and subtract build and run costs. Add second-order effects only if you can defend them: recovered analyst hours reallocated to underwriting more deals, fewer missed lease-expiry outreach windows, fewer commission reconciliation errors. Plug your own numbers into our CRE automation payback model rather than trusting a benchmark.

The figures below are planning defaults we use to scope a first build. They are starting points to argue with, not findings.

4–8
tools in a sensible first MCP server
Planning default, not a finding
5
known-answer questions in your evaluation set
Planning default, not a finding
90
days read-only before enabling any write action
Planning default, not a finding

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