JLL GPT vs Your Own CRE AI Agent: What to Copy

By Jude Lee · · Custom

Commercial real estate brokers reviewing a deal pipeline dashboard on a monitor in a glass conference room

JLL publicly announced JLL GPT in 2023, describing it as a large language model built for commercial real estate, and later positioned AI inside its JLL Falcon technology platform. The specifics have moved since — check JLL’s own newsroom and product pages for the current state, because enterprise AI product lines get renamed and re-scoped constantly (this article is written as of 2026).

But almost nobody searching “AI JLL GPT” runs JLL. The real question underneath is: the big firms have an AI trained on CRE — am I falling behind, and can I get something like that without a nine-figure tech budget?

Honest answer: partly yes, and the part you can get is the part that matters.

The model was never the moat

A general-purpose assistant like Claude, ChatGPT or Gemini already knows what a cap rate is, how a TI allowance works, and how an offering memorandum is structured. That knowledge is commodity. What a general assistant doesn’t know is that 4400 Westheimer went hard on Tuesday, that your client hates ground-floor retail, or that the tenant in Suite 210 has a 2027 expiration with one five-year option at fair market rent.

Enterprise CRE AI closes that gap by connecting the model to proprietary data. That’s it. That’s the moat — a research pipeline, a property database, a CRM, and a permissions layer that keeps one team’s confidential deal out of another’s answers.

The expensive part of enterprise CRE AI isn’t the model. It’s the wiring, the permissions, and the discipline to keep the underlying data clean.

That is also, in our opinion, where most smaller-firm AI projects stall: ambition arrives long before anyone has agreed who owns the CRM’s data quality.

Three routes to “our own CRE AI”

Vendor AI features (Buildout, Apto, CoStar, Microsoft Copilot)

What it is: AI bolted onto software you already pay for — listing description generation, comp summaries, email drafting, search in natural language.

Wins when: the workflow is standard across every brokerage (marketing copy, comp lookup, meeting notes), the data already lives in that vendor’s system, and you want zero engineering.

Breaks when: the answer requires joining data across three systems the vendor doesn’t own, or your process is genuinely non-standard.

Custom agent over your own systems (assistant + MCP)

What it is: Claude (or another assistant) connected via MCP — the Model Context Protocol, an open standard for giving an AI governed, permissioned access to your tools and data — to your CRM, listing database, document store and email.

Wins when: value comes from combining systems, from taking multi-step actions, or from encoding your house rules.

Breaks when: your underlying data is stale or inconsistent, nobody owns the integration after launch, or the workflow was already solved by a seat-based subscription you’re paying for anyway.

Two failure modes deserve naming, because “breaks when your data is a mess” is too vague to act on. First: the assistant reads a CRM record nobody has touched since March, reports the deal as active in fluent, confident prose, and gives you no signal that the field is stale — so build “last updated” timestamps into every tool response and show them in the answer. Second: a read tool scoped to one user’s permissions returns 40 of your 60 active listings, and the assistant summarizes those 40 as if they were the whole portfolio. Partial results look exactly like complete results unless your tools return row counts and the agent is instructed to state them.

There is also an exit cost the build column has to carry. Someone wrote that MCP server. When they leave, does anyone else understand the auth flow, the field mappings, and why one endpoint has a weird retry? If you switch assistant vendors, the MCP server itself is portable — that’s the point of an open protocol — but your prompts, skills and evaluation habits are less so. Before you build, name the maintainer, budget for their time annually, and keep the server’s code and docs in your own repository, not a contractor’s.

The third route — building your own fine-tuned model, the literal JLL GPT approach — is almost never right for a firm under a few hundred brokers. You’d be paying to recreate general CRE knowledge the frontier models already have.

What a custom MCP server actually gets you

Worked example. A 14-broker industrial shop wants to answer, on Monday morning: “Which of our active listings have had no client-facing activity in 10 days, and which of those have an inquiry sitting unanswered?”

No off-the-shelf tool answers that, because it spans the CRM (activity log), the listing platform (status), and the email system (unanswered inbound). With a custom MCP server exposing three read tools — list_active_listings, get_recent_activity, search_inbox_by_listing — an assistant can answer it in one prompt, cite the underlying records, and draft the follow-ups for a human to send. That’s the concrete pattern we walk through in connecting Claude to brokerage systems via MCP.

Add skills — packaged, reusable instructions that teach the assistant to do a job the same way every time — and the output stops drifting. A “deal status update” skill that always pulls the same fields in the same order beats a brilliant one-off prompt that nobody can reproduce. Same logic applies to standardized property marketing and follow-up.

Where each route wins, job by job

  1. Marketing copy and listing descriptions — buy

    Your listing platform’s built-in generation is good enough and already has the property record. A custom build here is engineering spend with no defensible payoff.

  2. Comps and market data — buy

    The value is in the licensed dataset, not the AI layer. Evaluate on data quality and coverage first; see the CRE data platform comparison for how those differ.

  3. Lease and rent-roll abstraction — mixed

    Off-the-shelf abstraction tools are mature. Build only if your asset class or lease forms are unusual, or if you need output written straight into your own underwriting model.

  4. Pipeline hygiene and cross-system questions — build

    This is the strongest custom case: multi-system reads, your definitions of ‘stalled’ and ‘qualified’, actions taken on your behalf.

  5. Anything with a signature, a fiduciary duty, or a number a client will rely on — human

    Agents draft; licensed people decide and send. Underwriting assumptions, LOI terms, and disclosure language stay reviewed.

Modeling the cost without pretending to know your numbers

Don’t start from a vendor’s ROI headline. Start from a formula you can defend to your managing broker:

Annual recovered hours = (hours per week the task consumes across the team) × 48 × (share the agent can realistically absorb — assume well under 100%).

Value = recovered hours × your blended loaded hourly cost, plus whatever revenue those hours produce when reallocated to tours, calls and pitches — which for producers is usually the bigger number and the harder one to estimate honestly.

Cost = build or subscription cost + integration effort + ongoing maintenance of the integration + the hours someone spends checking outputs, which never goes to zero.

hours/wk × 48 × capture rate
Recovered hours — fill in your own inputs
Worked example, not a benchmark
3 systems
Typical minimum an agent must touch before custom beats off-the-shelf
Author's rule of thumb
0%
Realistic reduction in human review for client-facing output
Editorial position

Our payback calculator walkthrough expands this with the assumptions spelled out. Any figure you see quoted as an industry average — including in vendor case studies — should be treated as directional at best.

Is free AI enough to start?

For a solo broker or a small team, genuinely yes for a while. Free tiers of the major assistants will summarize a lease, restructure a marketing email, or sanity-check a rent roll you paste in. What free tiers won’t do: connect to your systems, enforce who can see which deal, or take actions. The moment you want an agent that reads your CRM rather than one you paste into, you’re in paid territory with real data-handling decisions.

The honest verdict

If you’re a mid-size brokerage, don’t try to build JLL GPT. Build the two or three connections that let a good off-the-shelf assistant see your listings, your CRM and your inbox — then wrap the recurring jobs in skills so the output is consistent. Buy everything generic. Keep a licensed human on anything a client relies on.

That’s not a smaller version of what the majors did. It’s the same idea with the expensive, non-differentiating parts left out. If you’re still weighing the two directions broadly, our build-vs-buy framework for CRE software walks the decision tree end to end.

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