AI for commercial real estate brokers: where to start
Which brokerage workflows are worth building properly with AI, why most pilots stall, and a sensible order for a leasing, sales, or management team.
King & Company
In short
AI is dependable on a brokerage team when it produces the first draft of a document the team already makes (a lease abstract, underwriting inputs, a BOV, a survey, a call sheet, a market report) and a named person reviews the fields that carry money or liability. Pilots tend to stall when the team's definitions, template, and review step were never written down, which is a different problem from the quality of the model. Start with one workflow built against your own documents inside your firm's own AI workspace, then add the next.
AI for commercial real estate brokers works best when it is pointed at a specific document the team already produces every week, with a named person reviewing the parts that carry money or liability. The dependable way to get there is to rebuild one workflow at a time against the team's own leases, rent rolls, and templates, and that can be done inside a general AI workspace on a business plan.
That is a narrower claim than most of what is written on this subject, which tends to be a list of products sorted by category. A broker's week is organized by deliverables, so this guide is organized the same way: what to build for leasing, capital markets, and property management teams, what a person still has to review, and what order to build in.
Why do most brokerage AI pilots stall?
The industry's own surveys describe a wide gap between trying AI and getting a result from it. JLL's 2025 Global Real Estate Technology Survey of more than 1,000 senior corporate real estate decision-makers found that 92% of corporate real estate teams had started piloting AI or planned to that year, and that only 5% reported having achieved most of their program goals. Deloitte's 2027 Commercial Real Estate Outlook, a survey of 950 executives at real estate owners and investment companies, found 92% in the piloting or research phase and 8% with integrated AI solutions.
Those two surveys cover corporate real estate teams and owners. Brokerages have their own version of the problem. Bisnow reported in June 2026 that companies across the industry are struggling with "shadow AI," meaning employees using AI without official approval and parameters, while technology costs rise.
The most useful finding is about the constraint. JLL's Future of Work Survey 2026, which polled over 2,200 executives and corporate real estate leaders, says that for the first time in 15 years of the research, skills gaps overtook budget as the key constraint on real estate transformation.
The surveys do not say what goes wrong inside a single brokerage team. Our view is that reading a lease is well within what current models can do, and that a pilot usually stalls for one of three plainer reasons:
- Nobody wrote down the team's own definitions and template, so the output looks like a generic summary and gets rewritten by hand.
- Nobody designed the review step, so either everything is rechecked from scratch or nothing is.
- The output stops in a chat window and never reaches the Excel file or CRM record where the work continues.
Which documents eat a brokerage team's week?
Start from the deliverable. Six of them involve a great deal of reading, retyping, and formatting, and each has a natural reviewer.
| Deliverable | What AI drafts | Who signs off |
|---|---|---|
| Lease abstract | Dates, rent schedule, options, expense terms, with a page reference for each | The broker or lease administrator who owns the account |
| Offering memorandum and rent roll intake | Underwriting inputs placed into the team's own model | The analyst, then the senior broker |
| Broker opinion of value | Property summary, comp table, narrative first draft | The broker whose name is on it |
| Survey and proposal comparison | A workbook row per building or proposal, from flyers and landlord responses | The broker running the requirement |
| Call list and call prep | Ownership and tenant research, a one-page brief per call | The broker making the call |
| Market report | A recurring draft in the team's format from the data the team is licensed to use | The research lead or team lead |
What should each kind of team build first?
Leasing teams
The lease abstract is the usual first build, because the source document is long, the output format is fixed, and every field can be checked against a page. Our guide to abstracting a commercial lease with AI covers the field list and the checks. Surveys and proposal comparisons come next, since they are the same extraction problem pointed at flyers and landlord proposals, and the output belongs in a workbook. That build is described in getting deal data into Excel and your CRM.
Capital markets teams
The time goes into getting the rent roll, the trailing twelve months, and the offering memorandum into the team's model. AI should fill the inputs of the model the team already trusts and flag where the documents disagree with each other. It should leave the assumptions to the analyst. The BOV first pass builds on the same inputs and adds the comp table and narrative draft.
Property management teams
Abstracts matter here as well, because they feed the critical dates the team is accountable for. After that, the best candidates are recurring items with a fixed format: tenant and inspection notices queued for a person to approve before they go out, and monthly or quarterly owner reporting drafted from the same sources each cycle.
Prospecting, call prep, and market reports
The work around the deal suits packaged, repeatable instructions that every broker on the team runs the same way. In Claude these are called skills, which Anthropic describes as folders of instructions, scripts, and resources that Claude loads dynamically. Our guide to Claude skills for commercial real estate lists the ones a brokerage team tends to want first.
What does a person still have to review?
Every workflow needs a short list of fields that a named person checks against the source every time: rent and escalations, commencement and expiration dates, option and termination terms, anything that sets a price. The workflow should make that check fast by showing the page or cell each value came from and by flagging values it was unsure about. Fields outside the list can be spot-checked.
AI should not be doing three things on a brokerage team:
- Final valuation and pricing judgment. It can assemble the comps and draft the narrative. The opinion belongs to the broker.
- Sending anything to a client or counterparty unreviewed. Drafts are queued for approval.
- Carrying licensed market data into tools the license does not cover. Read the data agreement before building a report on top of it.
The reason for the first two is practical. In the Bisnow article above, Walker & Dunlop described a test where Claude finished an ownership research task in about ten minutes that took an analyst about a week, and the AI missed several results. A first pass that fast is worth having, and it still needs someone who knows what a complete answer looks like.
General assistant, CRE point solution, or a workflow built for your team?
| Option | Suits | Watch for |
|---|---|---|
| General assistant on a business plan | Teams willing to write down their template and definitions once | Without that setup, every broker prompts differently and results vary |
| CRE point solution | One narrow job where the vendor's format is acceptable | Another subscription, another place client documents live, output that may not match your template |
| Workflow built for your team | Deliverables where your format and review standards are part of what you sell | Someone has to build and maintain it, inside or outside the firm |
A team does not need a new platform subscription to start. On the question of ChatGPT or Claude, we would compare each vendor's current business terms and test both on five of your own documents before deciding. The assistant matters less than whether your definitions and review step are built into it.
Where are client documents and licensed data allowed to go?
Use the firm's own workspace on a business plan, and keep client documents out of personal accounts. Anthropic states that by default it does not use inputs or outputs from its commercial products to train its models, with exceptions for feedback a user explicitly submits and for data a customer otherwise chooses to share. Consumer plans are governed by a separate model improvement setting. Other vendors publish their own terms, and they should be read the same way.
Vendor terms are one of three things to check. The others are the confidentiality agreement on the deal and the license for any market data. Deloitte's survey found that 72% of respondents had completed preliminary data mapping while fewer than half had implemented more advanced process and security controls, which suggests that many firms still have this work ahead of them. Confirm the specifics with your own counsel or compliance lead. Our guide to secure AI workflows for confidential client data goes further.
What order should a team build in over the first twelve weeks?
This is the sequence we would suggest for a team starting from scattered experiments. Adjust the first workflow to whatever your team produces most often.
- Weeks 1 to 2. Settle the workspace and the rules for what may go into it. Pick one deliverable and collect ten real examples with the finished versions your team produced.
- Weeks 3 to 5. Build that workflow against the examples. Write down the template, the definitions, and the list of fields a person checks. The person who owns the deliverable corrects each draft.
- Weeks 6 to 8. Connect the output to where the work continues, usually an Excel template or a CRM record. Start a second workflow that reuses the first one's extraction.
- Weeks 9 to 12. Add the work around the deal (call prep, prospecting, a recurring report) and train the rest of the team on what exists.
Building one workflow at a time is slower than buying a platform license for everyone, and it leaves the team with something it uses on live deals. If you would like a second opinion on which workflow to start with, you can get in touch.
Common questions
What is the best AI for commercial real estate brokers?
For most teams the choice of assistant matters less than the plan it is bought on and the work built on top of it. A business plan from a major vendor, with your team's own template, definitions, and review step built in, will do more for a brokerage team than a better model used ad hoc. Compare the vendors on their data terms and on how well each reads your own documents.
Can AI replace a commercial real estate analyst?
No. It can take the first pass of reading, extraction, and formatting, and the analyst still checks the inputs, owns the model, and forms the judgment. Bisnow reported a Walker & Dunlop test where Claude finished a research task in about ten minutes that took an analyst about a week, and still missed several results.
Is it safe to upload leases and offering memoranda to an AI tool?
It depends on the plan and on what your client agreements allow. Anthropic states that by default it does not train on inputs or outputs from its commercial products, and consumer plans work under different settings. Confirm the vendor's terms, any confidentiality agreement on the deal, and your data licenses with your own counsel or compliance lead before client documents go in.
Do we need CRE-specific AI software or can we use Claude or ChatGPT?
A general assistant on a business plan is enough to start, provided someone builds your template, definitions, and review step into it. Point solutions make sense when you want a vendor to run one narrow job and you accept their format. Neither removes the need to decide who reviews the output.
Which brokerage tasks should not be handed to AI?
Keep final valuation and pricing judgment with the broker, and have a person read anything before it goes to a client or counterparty. AI can prepare the draft for both. Licensed market data should stay out of any tool the license does not cover.