AI commercial real estate market report: a weekly setup

A weekly commercial real estate market report can be assembled and drafted by AI, as long as the model never supplies a number of its own. Every figure in the issue comes from a public source pulled by code, from the team's own records, or from a licensed provider that a person entered and attributed, and a named broker approves the issue before it is sent.

Why is a regular market update hard to keep going?

A regular market update is worth sending, and the difficulty is the assembly. Someone has to look up where rates moved, find the latest employment release, collect what closed, ask three brokers what they are hearing, and then write it up in a way that sounds like the team. That work lands on whoever has a free afternoon, and when a deal heats up, the afternoon goes away and the issue slips.

The writing is the smaller part of the job. Gathering inputs the same way every time is the larger part, and it is the part that can be handed to software.

What makes a market report worth opening

A market report is credible for two reasons. The reader can trace every number to where it came from, and the report contains something the national research desks do not publish, which is your team's read of your submarket.

A single prompt asking a chatbot for "a summary of the office market in our city" cannot deliver both. Whatever comes back, the reader has no way to trace each figure to a dated source, and the tool has no knowledge of the tour you gave on Tuesday. A build that respects those two reasons looks different from a single prompt.

Three kinds of content, handled three different ways

Sort everything that goes into the report into three groups, because each one needs different handling.

Kind of content	Where it comes from	How it gets into the report
Public data	Federal statistical agencies and the Federal Reserve	Pulled by a script on a schedule, with the source and date stored beside each value
The team's own activity	CRM records and a short weekly note from each broker	Exported or read from the CRM, plus the notes as written
Licensed market statistics	The data provider your firm subscribes to	Entered by a person, attributed, and only as your agreement allows

Which public sources can be pulled the same way every week?

Four federal sources cover the economic backdrop and are published on a predictable basis.

• FRED. The St. Louis Fed's API lets a program retrieve economic data series from FRED, and every request needs an API key tied to a FRED account. The 10-year Treasury yield, for example, is a daily series there.
• County employment. The BLS Quarterly Census of Employment and Wages publishes a quarterly count of employment and wages covering more than 95 percent of U.S. jobs at the county, metro, state, and national levels by industry. BLS also provides CSV files meant for programmers to retrieve the published data. It arrives with a lag: the release calendar shows first quarter 2026 data published on August 28, 2026, so this belongs in the report when a new quarter lands and should be labeled with the quarter it describes.
• Building permits. The Census Bureau's Building Permits Survey provides monthly statistics down to the county and place level. It covers new privately owned residential construction only, so it is context for multifamily and retail readers and says nothing direct about office or industrial supply.
• Regional commentary. The Federal Reserve's Beige Book gathers anecdotal information on current conditions from each Federal Reserve District and is published eight times per year. A sentence from your District's section, quoted and linked, is a useful outside voice.

The script that pulls these should save the value, the series or table name, the period it covers, and the date it was retrieved. That record is what makes the report traceable later.

How do you capture the team's own read without adding a meeting?

The second group of content is what your clients cannot get elsewhere. Part of it already exists in the CRM: deals closed, proposals out, tours given, new listings. If those records are kept current, they can be read on a schedule. If they are not, the report is a good reason to fix that, and we describe one way to do it in getting deal data into Excel and your CRM.

The rest is judgment. Ask each broker for a short note once a week, written or dictated, answering the same two or three prompts: what landlords are saying, what tenants are asking for, and what surprised them. A few sentences each is enough. The notes go into the draft as the team's commentary, edited for length and never invented.

Smaller markets have a particular reason to do this. Steve Zacher, SIOR, wrote in 2018 that "brokers in smaller markets do need to create their own research in order to communicate to tenants, landlords and investors," while brokers in large markets usually have a research department or reliable third-party providers. A team that maintains its own survey of buildings, vacancies, and asking rates owns the data in its report outright. Keeping that survey current is the work an automated assembly step frees up time for.

Licensed data: read your agreement before it goes into any AI tool

Vacancy, asking rent, and absorption figures for a submarket usually come from a subscription data provider, and that data is governed by the license your firm signed. Before any licensed figure goes into an AI tool or a client email, read the agreement for two things: whether the data may be entered into software outside the provider's own product, and what the terms are for quoting it to people outside the firm. Ask your account representative or your counsel if the language is unclear. We are not characterizing any provider's terms here, and yours are the ones that matter.

Until that question is answered, the cautious design keeps licensed statistics out of the automated steps entirely. A person types the handful of headline figures into a clearly marked section of the report, with the provider and the period named. Published research handles attribution the same way. The National Association of REALTORS, for example, states that its commercial metro market dashboard of net absorption, vacancy, rent, and other indicators is built on CoStar market data.

Let code do the arithmetic and the model do the drafting

A calculation done in code can be rerun and checked line by line, and a figure typed out by a language model cannot. Split the work so that code owns every number and the model owns the prose.

1. A script pulls the public series and reads the CRM export.
2. The same script computes every comparison: change since last period, change since a year ago, counts and totals of the team's activity.
3. The results are written to one data file, each value with its source, period, and retrieval date.
4. The model receives that file, the brokers' notes, and the manually entered licensed section, with an instruction to use only the figures provided and to attach the source to each one.
5. The model writes the draft, and every number in it is copied from the data file.

If a figure is missing from the data file, the draft should say so and leave the gap for the reviewer. A vacancy rate or rent figure recalled from the model's training data has no source and no date, so it has no place in the issue.

How do you make it sound like your team?

Give the model your past issues. Five or six that the team is proud of, saved in one folder, do more than any description of tone. Add a short list of house rules: how you refer to submarkets, which terms you avoid, how long the issue runs, how you sign off. Packaged together with the data script, this is the kind of job that suits a skill, which we cover in Claude skills for commercial real estate.

When the reviewing broker rewrites a paragraph, save the final version back to the folder. The style reference improves with every issue sent.

Definitions matter: say what you mean by absorption and vacancy

Readers compare your figures with others they have seen, so state your definitions once and keep them fixed. NAIOP's Commercial Real Estate Terms and Definitions (2024) is a workable standard. It defines net absorption as "the net change in occupied space over a specified period," gross absorption as the total space occupied over a period without subtracting space vacated, and vacancy rate as vacant space divided by total inventory. It also separates vacant space (inventory that is not currently occupied) from available space, which is everything being marketed for lease, including space that is still occupied but will be vacant in the near term. Where sublet space is excluded, it recommends adding the word "direct."

Put a one-line definitions note at the foot of each issue, and write the same definitions into the instructions the model works from so that it never swaps "available" for "vacant."

The approval step and the send

One named broker reads every issue before it leaves. The review has a checklist: each number matches the data file, each source link opens, the licensed section is attributed, the commentary says what the brokers meant, and nothing identifies a client or a deal that is confidential. Our guide to designing the human review step goes further into what a reviewer should be shown.

The send itself goes through the firm's normal email platform and approval process. The FTC's CAN-SPAM compliance guide says the law covers all commercial messages and "makes no exception for business-to-business email," and it lists requirements that include a valid physical postal address, a clear way to opt out, and honoring opt-out requests within 10 business days. Confirm with your compliance lead or counsel how those and your brokerage's own policies apply to your list.

What the routine looks like once it runs

As an illustration, a weekly cycle could run like this. On Friday afternoon each broker sends a short note. Over the weekend the script pulls the public series, reads the CRM, and computes the changes. On Monday morning the model drafts the issue from the data file, the notes, and the past issues. The reviewing broker enters the licensed figures the agreement allows, reads the draft against the checklist, edits the commentary, and approves. Marketing sends it through the usual platform.

A monthly cadence uses the same steps with more room for a deeper section, such as an update to the team's own building survey. Pick the frequency your team has something new to say at, and keep it.

A market report is one of several builds that fit a brokerage team, and AI for commercial real estate brokers covers where it sits among the others. If you would like help setting one up against your own data and past issues, get in touch.

Common questions

Can AI write a commercial real estate market report?

It can assemble the inputs and write the first draft, provided every figure comes from a named source or from the team's own records. It should not be asked to supply vacancy, rent, or absorption figures from memory, and a broker should read and approve each issue before it goes out.

Where can I get free commercial real estate market data?

Federal sources cover the economic backdrop at no cost: FRED for rates and economic series, the BLS Quarterly Census of Employment and Wages for county employment, the Census Bureau Building Permits Survey for residential permits, and the Federal Reserve's Beige Book for regional commentary. None of them publishes submarket vacancy or asking rents, which come from a licensed provider or from the team's own survey.

Can I put CoStar data into ChatGPT or Claude?

That depends on the license agreement your firm signed, so read it and ask your account representative or counsel before any licensed figure goes into an outside AI tool. The cautious design keeps licensed statistics out of the automated steps and has a person type them in with attribution.

How do I stop AI from making up market statistics?

Give the model a data file in which every figure has a source and a date, instruct it to use only those figures, and have code calculate every change and percentage. Then have the reviewer check each number in the draft against that file before approving.

Can a market report be fully automated?

The data pulls, the arithmetic, and the first draft can run without anyone touching them. The team's read of the submarket, the licensed figures, and the approval to send all need a person, and the report is only worth opening because of the first of those.