AI Over Your CRE Deal History: 4 Options Compared

By Jude Lee · September 6, 2026 · Comparison

Two commercial real estate brokers reviewing deal documents on a laptop in a glass-walled office

The question your brokerage can’t answer quickly

Every brokerage has the same failure mode. A broker asks: “Have we ever worked with the ownership group at 4200 Industrial? What did they push back on last time?” The answer exists — in a CRM note from 2022, a thread in someone’s Outlook, an LOI PDF in a SharePoint folder named 4200 Ind FINAL v3 (2).pdf. Finding it means someone opening four systems and guessing at filenames, or nobody bothers and the pitch goes out generic.

We’re not going to hand you a benchmark for how long that hunt takes. Time it yourself for two weeks — that measurement is the only one that should drive your budget.

That’s an institutional-memory problem, and it’s the one AI is genuinely good at: retrieval plus synthesis over documents you already own. It’s also where most “CRE AI tool” evaluations go wrong, because they compare draft quality instead of asking the harder question — can this thing see my actual data, and can I trust what it says about it?

Option 1: A chat assistant plus manual uploads

Claude Projects, ChatGPT Projects, or a custom GPT where you drop in the relevant PDFs and ask questions. Zero integration work. Genuinely excellent for a bounded task: “Here are eleven comparable leases we signed in this submarket — summarize the concession patterns and flag anything unusual in the TI allowances.”

Where it breaks: you have to already know which files matter. It can’t answer “search everything we’ve ever done with this ownership group,” because it only sees what you handed it. Long documents get truncated or silently summarized before the model reasons over them, and a rent roll pasted as raw text loses its table structure in ways that quietly corrupt numbers. Also: check your firm’s policy and the vendor’s published data-handling terms before uploading anything under an NDA or a confidentiality clause in a purchase and sale agreement.

Best for: solo brokers and small teams working one deal at a time. Free tiers cover a surprising amount of this — see our breakdown of what free AI tools actually cover for CRE brokers.

Option 2: Grounded notebooks

Google’s NotebookLM and comparable tools take a defined set of sources and answer only from them, with inline citations back to the source passage. That citation behavior is the real feature. When an assistant tells you a lease has a 2029 expiration, you want to click through to the page that says so.

Practical fit: a due-diligence data room, a single portfolio, a market study you’re assembling. You build a notebook per deal, the team asks questions against it, and every answer is traceable.

Limits: it’s a curated corpus, not a live system. It won’t know a deal moved stages this morning, and it doesn’t write back anywhere. This is a reading tool, not an agent.

Best for: deal-team Q&A over a fixed set of documents, where traceability matters more than freshness.

Option 3: Copilot or Gemini over the files you already have

Microsoft 365 Copilot searches across SharePoint, OneDrive, Outlook and Teams; Gemini for Workspace does the analogous thing across Drive and Gmail. If your brokerage is already standardized on one of those suites, this is the lowest-friction path to “ask a question across everything.”

Three honest caveats. First, results are only as good as your folder discipline — an assistant searching eleven copies of an OM with no naming convention will confidently cite the wrong version. Second, your CRM is usually outside the boundary: Buildout, Apto, ClientLook, Salesforce — unless a first-party connector exists, the assistant can’t see pipeline stage, commission splits or call logs.

Third, and the one firms underestimate: these assistants respect existing permissions, but existing permissions are usually wrong. Over-shared SharePoint sites and legacy “anyone with the link” Drive files mean the assistant will happily surface material a user technically had access to but never would have stumbled across — comp splits, a partner’s deal notes, an old payroll spreadsheet. Run a sharing audit before you switch it on, not after someone asks the wrong question.

Best for: firms already on Microsoft 365 or Workspace where the file store is the system of record — after the permissions cleanup.

Option 4: A custom MCP server over CRM plus documents

MCP — the Model Context Protocol, an open standard for giving an AI assistant governed access to specific tools and data — closes the gap between “my files” and “my systems.” You build (or commission) a small server exposing a handful of read operations: search_deals, get_contact_history, list_comps_by_submarket, fetch_document. Claude or another MCP-capable assistant calls those tools when a question requires them.

The difference in practice: the assistant chains steps. Ask about 4200 Industrial and it queries the CRM for the contact record, pulls the linked deal files, checks email threads tied to that contact, and returns a synthesis with sources — without anyone uploading anything. That’s an agent doing multi-step retrieval, not a chatbot. Our walkthrough of connecting Claude to CRE systems via MCP covers the build shape.

Before you scope anything, check what your CRM already ships. Buildout, Salesforce and other platforms have been adding first-party AI search and assistant connectors; if the vendor’s own documentation shows a connector that answers your top ten questions, the custom build is unnecessary and you should say so out loud in the meeting. Verify current capability in official product docs rather than sales decks — this category moves fast.

What a custom build costs you: someone owns scoping, permissions (a junior analyst’s assistant should not surface partner-level commission data), and maintenance when an API changes. It’s software. Treat it like software.

Best for: firms where deal history genuinely spans a CRM and a separate document store, that already have an engineering team or a standing contractor relationship, and that can name one internal owner accountable for scope and permissions. Missing any of the three? Stay on options 1–3 until you’re not.

Off-the-shelf assistant search

Live in days. Per-seat pricing you can cancel. No engineering owner. Scope limited to files in one suite; CRM usually invisible; answer quality tracks your folder hygiene and sharing settings.

Custom MCP server

Weeks to a first useful version. Reaches across CRM, email, document store and internal databases. Permissions are yours to define — and yours to get wrong. Requires a budget owner and someone on call when an API changes.

Picking by job, not by brand

The common search here is some version of which AI tool is best for commercial real estate, and the honest answer is that the question is scoped wrong. Best for lease abstraction is not best for market narrative writing is not best for portfolio Q&A. Pick per job, then consolidate where one vendor is genuinely adequate at several. Our 10-point scorecard for vetting CRE AI tools is built for exactly that comparison.

For deal-history recall specifically: if the answer lives in files, options 2 and 3 are usually enough. If it lives across CRM plus files plus email, and no first-party connector reaches it, option 4 is what’s left.

Your brokerage’s edge isn’t the market data everyone buys. It’s the ten years of your own deal history nobody can search.

General knowledge is free; your deal history isn’t

Brokers sometimes test assistants on the 2% rule — the heuristic that monthly rent should be at least 2% of purchase price. It comes from single-family residential investing and is rarely how commercial deals get screened; commercial pricing runs on NOI, cap rates, debt terms and comparable sales. An AI assistant will give you a fluent explanation from general knowledge, which is precisely the point: general knowledge is undifferentiated and available to every competitor. The valuable answers are the ones only your own data can produce.

Similarly, there is no formal “30% rule in AI” we can point to in any authoritative body of work. The phrase gets used loosely — sometimes about the share of tasks people expect to automate, sometimes as shorthand that AI drafts get you most of the way and a human closes the gap. Treat it as folklore, not a benchmark, and don’t build a budget on it.

Where the human stays

IBM’s Think coverage of research on the boundary of AI automation is a useful read on why full autonomy keeps hitting a wall. In brokerage terms, the boundary is judgment under liability: what you represent to a client, what goes into a signed document, what you tell an owner about pricing. Retrieval and summarization can be delegated. Assertion cannot — which is why read-only retrieval is the sane first rung for any firm here.

On the wider “what will AI do to commercial real estate” question, JLL published an insights piece on AI and its implications for real estate worth reading directly rather than through anyone’s summary. Our own view, stated as opinion: the task mix inside brokerage roles shifts faster than headcount does, and the work that thins out first is retrieval and reformatting — exactly the work discussed here.

Modeling the payoff without inventing numbers

Don’t accept anyone’s dollar figure, including ours. Build your own from three measurements you can actually take:

What to measure	How to calculate it	Where the input comes from
Time hunting through deal files	searches per week × minutes each	Log it yourself for two weeks
Recovered analyst capacity	hours saved × loaded hourly rate	Your payroll
Revenue effect of better-informed pitches	pitches per year × win-rate delta you believe	Your own pipeline — estimate conservatively

The recovered hours only count if they’re reallocated to revenue work — prospecting, tours, underwriting — rather than quietly absorbed. Our broker-ops payback calculator walks the arithmetic.

A four-week starting sequence

Log the questions

For two weeks, have the team paste every “does anyone know…” question into one channel. That list is your requirements document.
Try the cheap option first

Run the top ten questions through a grounded notebook, your existing Copilot/Gemini seat, or your CRM’s own AI search. If seven come back right with citations, stop — you’re done and you saved a build.
Identify the gap

The failures will cluster, usually around CRM state or email history. That cluster is your MCP scope.
Build read-only, permission-scoped

First version exposes search and fetch only, mirroring existing CRM roles. No writes until retrieval is trusted.
Set a review cadence

Sample answers monthly against source documents. Retrieval quality degrades silently as data grows.

Vendor capabilities in this category change quarterly; verify current connectors and data-handling terms in official product documentation before you commit, and loop in counsel where client confidentiality obligations are involved.

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