CRE AI Automation: Prompt vs Skill vs Agent vs MCP

By Jude Lee · · Custom

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

Plenty of brokers are now genuinely good at prompting. Very few brokerages have moved from prompting to plumbing — getting the assistant connected to the systems where the deal data actually lives, with rules about who can do what.

The move from one to the other isn’t a single upgrade. It’s five distinct rungs, and most teams should not climb all of them.

The rungs, defined honestly

Rung 1 — the saved prompt. A block of instructions you paste into ChatGPT, Claude, or Copilot, usually with a document or a spreadsheet export. Zero setup. Works today. It breaks because nobody saves the good version, every broker runs a slightly different one, and the assistant has no access to your systems — you are the integration layer.

Rung 2 — the skill. A packaged, named set of instructions (plus reference files like your BOV template or your market-report voice guide) that teaches the assistant to do one job the same way every time. Anthropic’s Claude supports skills as a first-class concept; ChatGPT’s custom GPTs and project instructions occupy similar ground. A skill fixes consistency, not access. It still can’t read your CRM.

Rung 3 — off-the-shelf integration. Before anyone writes code, check what already exists: native integrations between your CRM and your email, vendor-published connectors, iPaaS tools (Zapier, Make, Power Automate) that move records between systems on a trigger, and — increasingly, as of 2026 — prebuilt MCP servers that software vendors publish for their own products. Check your CRM, email, and document-storage vendors’ current documentation; this list changes quarterly. Most teams should exhaust this rung before commissioning custom development. It’s cheaper, somebody else maintains it, and it solves a surprising share of “the data is in two places” problems outright.

Rung 4 — the agent. Multi-step work where the assistant plans, calls tools, checks its own output, and takes actions — sending the draft, creating the task, updating the record. Governance starts to matter here, because now a mistake leaves the building. Our view on how much autonomy to hand over is that most brokerage tasks belong at “drafts and waits” for a long time.

Rung 5 — the custom MCP server. MCP (Model Context Protocol) is an open standard for giving an AI assistant structured, permissioned access to data and tools. A custom MCP server is a small piece of software you or a developer build that exposes your listings, deal records, commission ledger, or document store as callable tools with scoped permissions. It’s for the cases where no vendor connector covers your actual workflow.

Rungs one and two make the assistant better at writing. The rest make it useful without a human doing data entry first.

One task, up the ladder: the Friday listing sweep

The task: every Friday, identify which of 40 active listings have had no inbound activity in 14 days, and draft the owner update for each.

Prompt or skill
An ops coordinator exports an activity report from the CRM and a listing roster from the marketing platform, pastes both into the assistant with a saved prompt, and gets 40 drafted owner emails in one pass. Quality is fine. The coordinator still spends the morning exporting, reconciling two exports that disagree about which listings are active, and pasting drafts into Outlook one at a time. A skill makes the drafts consistent in tone and structure across every broker — it does nothing about the exporting.
Agent on a custom MCP server
An MCP server exposes three read tools (list_active_listings, get_inquiry_activity, get_last_owner_contact) and one write tool (create_draft_email). The agent runs Friday at 7am, queries live data, flags the stale listings, drafts into each broker’s mailbox as unsent drafts, and posts a summary in the team channel. No exports. You can log every tool call — if you build and keep the logs; the protocol doesn’t hand you an audit trail for free. Cost: real development work, a permissions review, and someone who owns it when the CRM schema changes.

Notice what changed between the columns. The writing quality barely moved. The data handling changed completely. That’s the honest case for the upper rungs — and the honest reason to skip them when a task runs twice a quarter.

There is no single best AI tool for commercial real estate

It’s the most-searched question in this category, and the premise is off. “Best tool” depends on which of three categories you’re shopping:

Most brokerages end up with all three. Our tool selection framework walks the decision per job rather than per vendor.

Don’t let residential rules of thumb into your screening instructions

The 2% rule — monthly rent should equal or exceed 2% of purchase price — is a residential rental heuristic. It isn’t a commercial underwriting standard, and an agent that screens deals against it will throw away good assets and surface bad ones. If you’re building screening automation, encode your actual buy-box: asset type, submarket, going-in cap rate, debt yield and DSCR thresholds at a stated loan constant, lease term remaining, tenant credit floor. Generic internet heuristics are exactly the plausible-sounding filler an unanchored assistant supplies when your instructions are vague.

The compliance questions that actually bite

Using AI is not, in itself, prohibited for licensed brokers in the U.S. The constraints come from rules that already applied to your output regardless of who typed it:

Modeling the payback without making up numbers

There is no credible industry benchmark for what a CRE AI build saves. Build the model yourself:

hours × rate
The only ROI formula that survives scrutiny — your hours, your loaded rate
52 ×
Weekly tasks compound; quarterly ones rarely justify custom development
1 task
Start with a single repeating workflow, not a platform rollout

Plug in: hours the task takes today × loaded hourly cost × frequency per year. Subtract the hours the automated version still takes, including review. Then add the part people forget — the revenue work those recovered hours go to, and the errors avoided. To value errors without inventing a rate, go back through the last two quarters, count the misses that actually happened (the owner update that never went, the reminder that blew, the commission invoice sent late), and attach your own cost to each one. A handful of real incidents you can name beats any assumed error percentage.

How to pick your rung

  1. Write down the task and its frequency

    Daily, weekly, per-deal, or quarterly. Frequency is the strongest single signal for how far to climb.

  2. Run it as a prompt for two weeks

    If a saved prompt plus manual exports gets you 80% there and the task runs monthly, stop. You’re done.

  3. Package it as a skill if more than one person does it

    Consistency across brokers is the problem skills solve.

  4. Count the copy-paste steps

    If most of the effort is moving data between systems rather than thinking, you have an integration problem.

  5. Exhaust the off-the-shelf options first

    Check your CRM, email, and listing platform for native integrations and published connectors or MCP servers, then price an iPaaS workflow. Only move past this rung when you can name what it can’t do.

  6. Scope the smallest possible custom MCP server

    Two or three read-only tools over the one system that matters most. Our walkthrough on connecting Claude to brokerage systems via MCP covers the setup pattern.

  7. Keep a human on anything that leaves the building

    Drafts, not sends, until the error rate is something you’ve measured on your own data.

The practical takeaway: stop asking which AI tool is best for commercial real estate and start asking which rung a specific task deserves. Most tasks stop at rung two or three. The handful that run every week, touch three systems, and currently eat an ops coordinator’s Friday are the ones worth building for.

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