How to leverage your firm's own data with AI in CRE

To leverage your firm's own data with AI, connect the assistant your team already uses to the places your work actually lives, such as SharePoint or Google Drive, the CRM, and your deal folders, after you have fixed who can see what. Then build specific workflows on top of that connection, like comparable deal lookups or call prep, instead of asking people to search it by chat. Keep licensed market data out unless the license allows it, because the major providers' terms now address AI directly.

This article is for the managing director, team lead, or operations lead at a brokerage or investment firm who suspects the firm's real advantage is sitting in old folders and wants to know what it takes to use it.

What counts as your firm's data?

Every CRE firm has two kinds of data, and AI treats them very differently.

The first is data the firm created: closed deal files, offering memoranda you wrote, BOVs, lease abstracts, rent rolls you reconciled, underwriting models, proposals, pitch books, the CRM's record of every owner call, and the email threads behind all of it. This is the record of what your firm knows, and nobody else has it.

The second is data the firm licenses: comps and property data from providers such as CoStar or CompStak, tenant and ownership data, and research subscriptions. You pay for access to it, and the license decides what you may do with it.

Most of the value from AI on a brokerage team comes from the first kind, and most of the legal risk comes from the second.

Why the firm's own record matters more than the model

Every brokerage can buy the same models. What a model cannot buy is your team's history: which owners you have called and what they said, what similar buildings traded for in deals you ran, how your senior brokers framed a BOV for a skeptical owner, and the rent escalations you negotiated last year. A model that can read that history drafts with your firm's knowledge. One that cannot is working from the open web.

Most firms are not there yet. JLL's 2025 Global Real Estate Technology Survey of more than 1,500 decision makers found that 88% of investors, owners, and landlords had started piloting AI, while only 5% of corporate real estate occupiers piloting AI reported achieving all their program goals. In Deloitte's 2027 commercial real estate outlook, 92% of the 950 respondents were still piloting or researching AI.

How the major assistants connect to company data today

All four of the assistants a CRE firm is likely to use can now read company files, and all four say they respect the permissions already set on those files.

Assistant	How it reaches firm data	What it says about permissions
Claude	Connectors for Google Drive, Microsoft 365, HubSpot, Box, and others, plus custom connectors over the Model Context Protocol	An admin enables a connector, and each user then authenticates with their own account
ChatGPT	Company knowledge, introduced October 2025 for Business and Enterprise, drawing on connected apps such as SharePoint and Google Drive	Answers come from what the user can already access in each connected app
Microsoft Copilot	Microsoft Graph, which covers mail, chats, and SharePoint and OneDrive files	Microsoft states that Copilot only surfaces data a user has at least view permission to
Gemini in Google Workspace	Drive, Gmail, and other Workspace data	Google states Gemini has the same access to Workspace data as you do

The right choice usually follows the systems the firm already runs. A Microsoft shop has a short path to Copilot, a Google Workspace shop to Gemini, and a firm that wants Claude or ChatGPT can connect either one to Microsoft or Google files. For document-heavy work, Claude's Projects also hold a team's reference files directly, and Anthropic says that when a project's files approach the context limit, Claude switches to retrieval to expand capacity on paid plans.

Step one: fix permissions before you connect anything

"It respects existing permissions" is reassuring only if the existing permissions are right. In most firms they are not. A deal folder shared with "everyone in the company" years ago was harmless when nobody could find it. Once a connected assistant can search it, a junior associate's question about comps may return a confidential client's financials.

Microsoft's own guidance on preparing SharePoint for Copilot says that SharePoint sharing settings default to the most permissive option, and recommends running an oversharing assessment, reviewing reports of content shared with everyone in the organization, and restricting sensitive sites before rollout. The same logic applies to Google Drive and Box. Before connecting anything, a firm should:

• Find folders shared with the whole company or by open links, and narrow them.
• Separate client confidential deal rooms from general reference material.
• Decide which teams should be able to search which folders, and set that in the file system, not in the AI tool.
• Repeat the review on a schedule, because new folders get shared broadly all the time.

We cover the wider set of controls in how to build secure AI workflows for confidential client data.

Step two: organize the record so the model can find the right thing

Connected assistants search your documents and add the relevant passages to the prompt before answering, which is the technique called retrieval-augmented generation. Anthropic describes it as retrieving relevant information from a knowledge base and adding it to the user's prompt, and its research on contextual retrieval shows how much the retrieval method itself affects whether the right passage is found.

What a firm controls is the material being searched. A few habits make a large difference:

• One final version. A folder with "OM_v7_FINAL_revised" beside four earlier drafts gives the model five conflicting answers. Mark or move the final.
• Consistent folder structure by deal. Property, deal type, year, and status in the folder name or metadata let a search narrow to the right deal.
• Abstracts and summaries next to the long documents. A one-page lease abstract or deal summary is easier to retrieve accurately than a 90-page lease. If your team builds abstracts with AI, the method in how to abstract a commercial lease with AI produces ones with page citations.
• CRM fields that are actually filled in. If last-contact dates and owner notes live in brokers' heads, no connector can read them.

Step three: build workflows on the connection, not a search box

Connecting the data is the plumbing. The return comes from specific workflows that use it, each with a fixed output and a person who reviews it. Some that brokerage teams get the most from:

• Comparable deal lookup. "Find every office deal we ran in this submarket since 2022 with the sale price, cap rate, and buyer type, with a link to each file."
• Call prep. Before an owner call, pull the CRM history, past correspondence, the property's own record in your files, and any open proposals into a one-page brief. We describe how to build one in Claude skills for commercial real estate.
• Proposal and BOV drafting from precedent. Start from the firm's own best past examples rather than a blank page, as in AI proposal drafting from a precedent library.
• Documents into the CRM and Excel. Extract terms from new leases and OMs into the CRM and the model, covered in deal documents to Excel and CRM with AI.

Every one of these should tell the user which files it used, so a broker can open the source and check it. Anthropic's guidance on reducing hallucinations recommends instructing the model to use only the documents provided and to quote them, and that instruction belongs in the workflow, not in each person's memory.

Licensed market data: read the terms before it goes near AI

This is where firms get into trouble without meaning to. Comps and property data are licensed, and the license terms have caught up with AI.

CoStar's terms of use, last revised July 6, 2026, prohibit users from exposing the licensed product "to artificial intelligence tools outside of the CoStar Product," or using it or data generated from it to train or improve any AI system, "including as input to generative artificial intelligence or retrieval-augmented generation." They also bar sharing passcodes with third-party AI services, including AI agents.

CompStak's terms, effective July 31, 2026, bar using its data to ground or create embeddings for AI, and say AI agents may access it only through CompStak-provided integrations such as Model Context Protocol connectors.

In practice that means a CoStar export saved into a deal folder may become searchable by a connected assistant, which the license appears not to allow. Firms should keep licensed exports in a location the AI connection cannot reach, use the provider's own AI features or approved integrations where they exist, and have counsel read the specific license the firm signed, since enterprise agreements can differ from the public terms.

Does a firm need a data platform?

Most teams do not need one to start. Connecting the assistant to well organized folders and a maintained CRM covers call prep, comparable deal lookup, and drafting from precedent.

A data platform or warehouse earns its cost when the firm wants to analyze across many years of deals or many properties at once, such as rent trends across a managed portfolio or win rates by deal type, where the answer depends on structured fields and not on reading documents. That is a larger project with its own owner, and it works better after the firm has learned which questions it actually asks.

A sensible order for the first ninety days

1. Pick one team and one workflow, such as call prep for a leasing team.
2. Review and fix permissions on the folders and CRM that workflow needs, and nothing else yet.
3. Clean up those folders: final versions marked, consistent naming, abstracts beside long documents.
4. Move licensed data exports out of reach of the AI connection and confirm the license terms with counsel.
5. Connect the assistant on a business plan, build the workflow with source citations in its output, and have the team use it for a month.
6. Measure whether people use it and what they correct, then extend to the next folder set and workflow.

Connecting a firm's own record to AI, with the permissions and structure done properly, is the core of what we build with brokerage and investment teams. If you want to talk through what your firm's data could do, you can get in touch.

Common questions

How do I connect my company's data to Claude or ChatGPT?

Both offer connectors on business plans. Claude's connector directory lists Google Drive, Microsoft 365 (SharePoint, OneDrive, Outlook, and Teams), HubSpot, and Box, and its help pages say an admin enables a connector for the organization and each person then signs in with their own account. OpenAI introduced company knowledge for ChatGPT Business and Enterprise in October 2025 to answer from connected apps such as SharePoint and Google Drive.

Will Microsoft Copilot show people files they should not see?

Microsoft states that Copilot only surfaces organizational data a user already has at least view permission to. The risk is that permissions are often broader than anyone realizes, and Microsoft's own guidance notes that SharePoint sharing settings default to the most permissive option. Fix oversharing before rollout.

Can I use CoStar data with ChatGPT or Claude?

Read your license with counsel first. CoStar's terms of use, last revised July 6, 2026, prohibit exposing the licensed product to AI tools outside CoStar's own product, including as input to generative AI or retrieval-augmented generation. CompStak's current terms allow AI access only through its own integrations.

What is retrieval-augmented generation in plain terms?

It is a way for an AI model to look up the relevant passages in your documents before answering, instead of relying on what it learned in training. Anthropic describes it as retrieving relevant information from a knowledge base and adding it to the prompt. The quality of the answer depends heavily on the quality and organization of the documents it searches.

Do we need a data warehouse before we can use AI on our data?

Usually not to start. Most brokerage teams get real value from connecting the assistant to the folders and CRM they already use, once permissions and folder structure are in order. A warehouse or data platform becomes worth it when you want to run the same analysis across many years of deals or many properties.