MCP-First AI Rollout for CRE Brokerages: 4 Options

By Jude Lee · · Comparison

Brokers and an operations manager reviewing deal pipeline data on laptops in a commercial real estate office

The “MCP-first” question residential brokerages raised first

A phrase has been circulating in brokerage trade press and vendor marketing: “MCP-first, AI of choice.” The idea is that instead of forcing every agent onto one AI assistant, a brokerage exposes its own systems through the Model Context Protocol and lets people bring whichever assistant they prefer. eXp Realty has been named in that conversation.

We are not going to attribute a specific quote or announcement date to any company here, because we can’t verify the wording from the primary source. If you want to cite a particular brokerage’s AI strategy in a board deck, pull it from that company’s own newsroom or investor filings rather than from secondhand coverage. Treat the phrase as a useful framing, not as evidence.

The framing is worth understanding on its own terms anyway — and residential brokerage economics are not CRE economics, so copying the conclusion would be a mistake regardless.

MCP is an open protocol, originally published by Anthropic and documented at modelcontextprotocol.io, that standardizes how an AI assistant connects to external tools and data. Instead of each AI vendor building a bespoke integration with your CRM, MCP defines a common socket. Your systems expose a server; any MCP-capable client can connect to it, under permissions you set.

The real baseline: copy-paste, not agents

Before comparing rollouts, be honest about where most firms actually are. In practice, AI use in CRE usually looks like a broker with a consumer chatbot subscription pasting a lease PDF in and asking for an abstract. That covers real work — lease and rent-roll abstraction, first-draft marketing copy, market narrative for a BOV, tenant follow-up emails, summarizing a long PSA.

What copy-paste does not do is take action. It can’t update the pipeline, pull the current rent roll, check who already received the deal blast, or write a note back into the CRM. That gap is what agents and MCP close — and it’s where the rollout decision starts to matter.

Four ways to roll AI out to your brokers

Embedded AI inside your CRE SaaS
Fastest to turn on. The AI already sits where the data lives, so there’s no integration work and no new vendor review. The limit: it only sees that vendor’s data, only does what that vendor shipped, and you can’t change the prompt behind it. Your CRM’s AI won’t read the data room; your data-room AI won’t touch the pipeline.
MCP-first, AI-of-choice layer
Slower to stand up and requires someone technical. In exchange, one assistant can see the CRM, email, and document store in a single conversation, and you can swap the model later without rebuilding. The limit: you now own permissions, logging, and uptime for a piece of infrastructure.

Option 1: Do nothing formal (shadow AI)

Brokers use personal ChatGPT, Claude, or Gemini accounts on their own. Cost to the firm is zero and adoption is already happening whether you approve it or not.

The real exposure isn’t “AI is risky” in the abstract — it’s that confidential offering material, NDA-restricted financials, and owner contact data leave your governed environment with no log of what went where. If you can’t move on architecture this quarter, two interim guardrails are cheap and immediate: (1) publish a one-page do-not-paste list — NDA-restricted financials, unredacted rent rolls, owner PII, anything under a confidentiality agreement — and have producers acknowledge it in writing; (2) move the firm onto business- or team-tier AI accounts with admin controls and a training-opt-out, and verify that opt-out in the vendor’s current terms rather than its marketing page. Neither gives you audit logs at the document level, but both shrink the exposure meaningfully. Governance options for shadow AI goes deeper.

Fits: very small firms — as a rule of thumb, not a benchmark, under roughly ten brokers — where norms are still enforceable in a single conversation, and only as a deliberate interim step.

Option 2: Turn on the AI features in your existing CRE stack

Several major CRE platforms have shipped AI features — check each vendor’s current documentation for what’s actually available on your plan, because this category moves fast and blanket capability claims age badly.

Strength: zero integration work, data stays inside a system you’ve already vetted, and the AI is tuned to a known schema. Weakness: siloed by design, and you inherit the vendor’s roadmap. If three vendors each ship an assistant, your brokers now have three assistants and still no single place to ask “what’s the status of every deal over $5M in Phase II?”

Fits: most firms, as the default layer. Buy before you build.

Option 3: MCP-first using vendor-published servers

Some platforms publish their own MCP servers; mainstream business tools such as Google Drive, Slack and GitHub have them. You connect a governed AI assistant to the servers that already exist and get cross-system reasoning without writing code.

Weakness as of early 2026: CRE-specific coverage is thin. The server you most want — over your CRM and listing data — may simply not exist yet. Check your vendors’ developer documentation directly before planning around it.

Fits: firms whose critical systems are mainstream business software rather than niche CRE tools.

Option 4: A custom MCP server over your own data, plus skills

You build a server exposing a defined set of read and write operations against your CRM, listing database, and document store, then package repeatable jobs as skills — reusable instruction sets that make the assistant run a lease abstract or a listing activity report the same way every time. The mechanics are covered in connecting Claude to your systems via MCP.

Weakness: this is software. It needs a budget, an owner, permission design, and a maintenance plan. A brokerage that can’t name who owns it in 18 months shouldn’t build it.

Fits: larger firms — we’d use 25+ producers as an editorial rule of thumb, not a measured threshold, on the reasoning that a custom server carries a roughly fixed annual maintenance cost no matter how many people use it, so it amortizes better the more producers share it. A proprietary data advantage (deal history, owner relationships, comps) or a workflow no vendor serves pushes the same direction.

No single AI tool is best — pick by workflow

Anyone naming one best tool without asking what you do all day is selling. The honest decision rule:

  1. Name the one workflow that hurts most

    Not “AI strategy.” One workflow — lease abstraction, deal blasts, pipeline hygiene. Count how many hours it eats weekly across the team.

  2. Check if your existing vendor already does it

    If your CRM or document platform ships that feature on your current plan, test it for two weeks before considering anything custom.

  3. Pick a governed assistant for everything else

    Choose a business-tier plan with admin controls and a no-training-on-your-data commitment — verify it in the vendor’s current terms. Our ChatGPT vs Claude vs Copilot comparison for brokers covers the tradeoffs.

  4. Only then ask whether MCP earns its keep

    If the painful workflow spans two or more systems, or requires writing data back, that’s the signal for a connection layer. If it doesn’t, don’t build one.

Modeling the payback without inventing numbers

We have no benchmark data on CRE AI savings, and you should distrust anyone who claims one. Build your own estimate instead.

hrs/week × weeks × loaded $/hr
Recovered-time formula (fill in your own figures)
Worked example
build + annual maintenance + licenses
Total cost side of the equation
Worked example
1 workflow, 2 brokers, 60 days
Pilot scope — editorial rule of thumb
Editorial recommendation

A worked example — every input is an assumption you should replace. Say lease abstraction currently takes two analysts four hours a week each, and you estimate a governed assistant with human review cuts that in half. That’s 2 analysts × 2 hours × 46 working weeks ≈ 184 hours a year. Multiply by your own fully loaded hourly cost (salary + benefits + overhead, divided by productive hours) to get the recovered-time side. On the cost side, add the quoted build, a realistic annual maintenance figure (plan on a meaningful fraction of build cost each year, not zero), and per-seat licenses. If recovered time doesn’t clear total cost in year one with room to spare, you have a pilot, not a rollout.

The part most firms miss: recovered hours only count if they get reallocated to revenue work. An analyst who saves six hours on abstraction and fills them with more abstraction has produced nothing. Decide in advance what the time is for — more tour days, more owner calls, faster OM turnaround — and the pilot scope above (one workflow, two brokers, 60 days) is our rule of thumb for that: small enough to cancel, long enough to get past the novelty period.

Treat the “30% rule” as folklore, not a benchmark

There is no governing body that defines a “30% rule for AI,” and no authoritative study we can point you to that establishes one. It circulates informally — usually meaning something like assume AI handles roughly the repetitive third of a job and a human handles the rest.

As a planning heuristic it’s not useless: it sets the expectation that output is a first draft, not a deliverable. As a number, it’s folklore. Don’t put it in a board deck as a finding.

The question isn’t whether AI replaces brokers. It’s whether a broker with a governed, connected assistant out-executes a broker without one on the same listing.

AI raises the floor; it doesn’t replace brokers

The structure of the business is the reason, not optimism. CRE transactions turn on off-market relationships, physical inspection, negotiating leverage, and fiduciary judgment with real liability attached. An agent can draft the LOI; it cannot read the room when the seller’s partner goes quiet.

What changes is the floor. Work that was defensible because it was tedious — abstracting a 90-page lease, assembling a tour book, chasing CRM hygiene — stops being a differentiator. Agents versus human virtual assistants is the more useful framing for most ops teams: which tasks move to software, which move to a person, and which stay with the producer.

A defensible sequence

Don’t announce an AI strategy. Pick one workflow, put a governed assistant on it with named humans reviewing output, measure the before-and-after yourself, and only then decide whether MCP earns a line in next year’s budget. An “AI of choice” architecture is genuinely attractive — it keeps you from being locked to a model vendor — but it’s a destination, not a starting point. The firms that get there will be the ones that first proved a single workflow worked.

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