AI for Commercial Real Estate Market Reports: 4 Options

By Jude Lee · · Comparison

Two CRE research analysts reviewing a printed quarterly market report in a brokerage conference room

The report is a data-assembly job wearing a writing costume

Ask a research director where the quarter goes and you rarely hear “drafting.” You hear: pulling the submarket export, reconciling it against last quarter’s numbers, chasing brokers for deals that never made it into the CRM, rebuilding the absorption chart because the vacancy definition changed, and formatting.

That matters for AI strategy. Generative AI is genuinely good at the writing layer — turning a clean table into three paragraphs of readable market commentary in your house voice. It is far weaker at the part that actually consumes your week: reconciling inconsistent sources and deciding which number is right. If you automate only the prose, you save the smallest slice of the job.

If AI writes your market commentary but a human still rebuilds the spreadsheet, you automated the easy part and left the expensive one in place.

Where AI actually shows up in a brokerage research workflow

People searching “how is AI used in commercial real estate” usually get a list of futures. Here is the narrow, present-tense version for a research function:

Notice what’s missing: deciding what the market is doing. That stays human.

Four production models, compared

1. Analyst-only (no AI)

A research analyst in Excel, a designer in InDesign or Canva, a review cycle. Highest control, highest consistency of judgment, and for a firm producing two reports a year it is often still the correct answer. The failure mode is capacity: the report ships late, or ships thin, or the analyst leaves and the methodology leaves with them.

2. Platform exports plus built-in AI summaries

Several platforms in this category now ship AI summarization or natural-language query features over their own data; check the current product documentation for whichever subscriptions you actually hold, since these features change release to release. Fastest path to something on the page, zero build cost, and the data is already normalized inside the vendor’s definitions.

Two real constraints. First, you are limited to that vendor’s coverage and their vacancy/absorption definitions, which may not match your firm’s published methodology. Second — and brokers underweight this constantly — subscription data is licensed, and redistribution or bulk extraction is typically restricted by the terms you signed. Before you pipe anything into a model or a public report, read your agreement and confirm with the vendor; we walk through the patterns in what CoStar and MLS data your AI agent can legally use.

3. A generalist AI assistant driven by a written skill

Claude, ChatGPT, or Copilot, plus a skill: a reusable, packaged set of instructions that teaches the assistant to do one job the same way every time. For market reports, the skill encodes your methodology — how you define net absorption, which submarket labels are canonical, that asking rents are quoted full-service in this market and NNN in that one, your house tone, your required disclaimer, your chart conventions.

The analyst still assembles and verifies the data table. The assistant turns that verified table into the report, consistently, across twelve submarkets, without the voice drifting between the office and industrial sections. For most brokerages this is the cheapest place to start — and the approach generalizes, which is why we treat skills as the standardization layer rather than a one-off prompt.

Its downsides are real and worth naming. Skills drift: the document goes stale the moment someone changes a submarket boundary or a quoting convention and doesn’t update it, and the assistant will keep applying the old rule with total confidence. Voice can regress silently when the underlying model is updated — output that matched your tone in March may read subtly off in September with no warning and no changelog you control. And most importantly, a skill does nothing about the assembly hours that this article identifies as the actual bottleneck. It makes the last mile consistent; it does not touch the first mile.

4. A custom agent connected through MCP

MCP (Model Context Protocol) is an open standard published by Anthropic for giving an AI assistant governed access to external tools and data. A custom MCP server exposes your systems — your comps database, your CRM deal records, your signed-lease archive, your licensed data warehouse — as tools the agent can query with permissions you control.

Now the agent can assemble the table, not just narrate it: pull every signed industrial lease in the quarter from the CRM, join it to your comps table, compute absorption, flag the three deals with missing square footage, and draft the section. That’s the full job. See connecting Claude to your systems via a custom MCP server for the build mechanics.

Before any of that touches confidential material, settle the data-handling question. Signed leases, CRM records, and broker email threads carry confidentiality obligations to sellers, landlords, and tenants that do not evaporate because the recipient is software. Confirm in the provider’s own data-usage documentation whether your inputs are retained, for how long, whether they are used for model training, and whether your plan or workspace offers a training opt-out or zero-retention configuration — then check that against the confidentiality and NDA language in your own agreements. Where a specific lease or listing agreement has unusual confidentiality terms, have counsel read it rather than guessing.

The other honest caveat: an MCP agent is only as good as the data behind it. If your comps live in four analysts’ spreadsheets with inconsistent submarket spellings, the agent will produce confident garbage faster than a human could.

Skill on a generalist assistant
Days to stand up. No engineering. Automates drafting and formatting only. Needs an owner who keeps the skill current. Good when your numbers are already reliable and consistency of voice is the pain.
Custom MCP agent
Weeks to months, with real maintenance. Automates assembly and drafting. Requires a clean, queryable, licensed data source and someone who owns it. Worth it when you publish many reports on a fixed cadence.

Sizing the payback without inventing a number

Don’t start from a vendor’s savings claim. Start from your own timesheet, then model it:

Hours per report × reports per year × fully-loaded hourly cost = current production cost. Then split those hours into assembly, drafting, formatting, and review, and apply automation only to the buckets a given option actually touches.

A worked example with assumptions you should replace: assume 40 hours per report, a fully-loaded analyst cost of $85 an hour, and four reports a year — that’s 160 hours and roughly $13,600 of annual production cost. If you log that drafting and formatting are a quarter of those hours, a skill-driven assistant is bidding against about 40 hours a year, not 160. Run the same split before you scope anything custom.

40 hrs
Illustrative hours per report — substitute your logged time
Worked example, not a benchmark
$85/hr
Illustrative fully-loaded analyst cost — substitute yours
Worked example, not a benchmark
4×/yr
Illustrative publication cadence — substitute yours
Worked example, not a benchmark

The larger prize usually isn’t the hours — it’s cadence. A brokerage that can publish monthly instead of quarterly gives brokers a fresh reason to call owners twelve times a year. Model that as additional qualified conversations, then apply your own historical conversion rate — not as a dollar figure someone else calculated.

Rules of thumb your AI will happily parrot — and shouldn’t

Two things turn up in AI-drafted market commentary that should never survive review.

The first is the 2% rule — the retail-investor heuristic that a property’s monthly rent should equal roughly 2% of its purchase price. It originated in small residential rental investing and has essentially no application to institutional commercial assets, which are priced off cap rates, NOI, lease term, and credit. If a draft offers it as CRE analysis, the model is pattern-matching on consumer blog content. Cut it.

The second is the so-called 30% rule in AI. There is no established, standardized definition of this term — it gets used loosely to mean anything from “AI can handle about 30% of a task” to “expect to rewrite 30% of AI output.” Treat any such figure stated as a law as unsourced. Your own measured edit rate on your own last ten drafts is the only version of that number worth acting on.

A realistic thirty-day path

  1. Time the current report, by bucket

    Log the hours for one full cycle, split into assembly, drafting, design, and review. You cannot choose between these four options without this.

  2. Write the methodology down

    Definitions, submarket boundaries, quoting conventions, tone, disclaimers. This document is the skill. If it doesn’t exist on paper, no tool will make your reports consistent.

  3. Clear the licensing and confidentiality questions

    For each data source, confirm in writing what you may extract, store, and redistribute — and confirm the assistant’s retention and training settings before confidential documents go anywhere near it.

  4. Pilot the drafting layer

    Run one submarket section through an assistant using the written skill and a verified table. Compare it to the analyst’s version side by side.

  5. Measure the edit rate and have the research director sign off

    Track how much of the draft survives review. The sign-off belongs to the research director or whoever owns published methodology — not the analyst who ran the pilot, who is too close to it. My own stop signal, offered as judgment rather than a benchmark: if more than roughly a third of the draft still needs rewriting after two rounds of skill revision, the methodology isn’t written clearly enough yet, and more prompting won’t fix it.

  6. Only then consider MCP

    If assembly is still the bottleneck and your data is clean enough to query, scope a custom agent. If your comps are a mess, fix the data first — that work is not wasted either way.

So which one is best?

There is no single best AI tool for commercial real estate, and anyone who names one is selling it. For market report production specifically: a firm publishing twice a year should probably keep an analyst and a good template. A firm publishing quarterly across three property types should write the skill and run it on a generalist assistant, with someone accountable for keeping that skill current. A firm publishing monthly across many submarkets, with a real comps database behind it, is the case where a custom MCP-connected agent earns its maintenance cost.

The constant across all four: the methodology document, the traceable sources, and the human who signs off. Those are the things clients are actually paying 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