AI accelerates deal flow in commercial real estate when you give a general assistant like ChatGPT or Claude one stage of your pipeline, the documents that stage runs on, and a written standard for what a good answer looks like. Start with screening. Check every figure against the source. Then move upstream to sourcing and downstream to packaging.

Where deal flow actually slows down

Ask any acquisitions lead or investment sales broker where their week goes and the answer is rarely "underwriting." It goes to the work around underwriting. Opening the package, pulling the trailing financials out of the data room, rebuilding the rent roll into something the model accepts, writing the two-paragraph summary a Partner reads before deciding whether the deal deserves a real look. Multiply that by every package that comes in, and the pipeline slows for a reason that has nothing to do with judgment.

Deal flow has four stages, and each one leaks time in a specific way.

Sourcing leaks time in research. Who owns the building, how long they've held it, whether the debt is coming due, what the last trade in the submarket looked like. Most of this is public or semi-public and most of it gets done by hand.

Screening leaks time in reading. A package arrives, someone reads all of it to find the six numbers that matter, and then decides whether to spend a day on it. Most packages fail the buy box early and still get read in full.

Packaging leaks time in writing. The deal memo, the OM, the investor update. Same structure every time, rebuilt from scratch every time.

Follow-up leaks time in memory. Which broker owes you the rent roll, which owner said "call me in Q4," which deal went quiet three weeks ago and nobody noticed.

A general assistant is good at exactly one kind of work: a repeatable first pass on documents you already have, against a standard you already know. That describes screening better than any other stage, which is why screening goes first. It's the stage with the most volume, the clearest standard, and the cheapest mistakes, because a screen that's wrong gets caught on the next read.

What you'll end up with

By the end of this guide you'll have a screening workspace in ChatGPT or Claude that holds your buy box and a screening standard, a prompt that takes any inbound package and returns a one-page screen in a format your team already recognizes, a checked result on a live deal, and a clear view of which stage to hand over next. You'll also know exactly where the output can be wrong, because you'll have watched it be wrong at least once.

What you need first

  • A paid individual account on ChatGPT or Claude. The standard $20-a-month plan on either is enough. The free tiers throttle file uploads, and this workflow is all file uploads. Either tool works; the prompt below runs on both.
  • Your buy box, written down. Asset type, size range, markets, return targets, deal killers. If it only exists in a Partner's head, step 1 fixes that.
  • One live deal package you haven't screened yet. The OM as a PDF, plus the T-12 and rent roll if the broker sent them separately. A real deal, because a test deal teaches you nothing about where the output breaks.
  • Whatever holds your pipeline today. A CRM, a spreadsheet, a shared inbox. You won't change it; you'll feed it.
  • About 90 minutes, once. Every deal after the first takes minutes.

Before you upload anything

Everything in this guide involves putting deal documents into a tool run by a company that isn't yours. Four rules keep that from becoming a problem. They take ten minutes and you only do them once.

Turn off model training. On the individual plans, both ChatGPT and Claude use your conversations to improve their models unless you switch it off. In ChatGPT: Settings, then Data Controls, turn off "Improve the model for everyone." In Claude: Settings, then Privacy, turn off "Help improve Claude." Do this before the first upload, and confirm it stuck by refreshing the page. Team and Enterprise plans on either tool don't train on your data at all, which is the right move once more than one person on your team is doing this.

Read the confidentiality agreement. Most marketed packages come with one. If it restricts sharing the package with third parties, an AI vendor is a third party, and uploading the OM breaks the agreement. When that's the case, don't upload it. Type the six screening figures in by hand instead and let the assistant do the buy box comparison and the broker questions. That takes five minutes and stays inside the CA. When in doubt, ask the broker whether their CA allows it; more of them are being asked this question every month.

Strip names before uploading. The screen doesn't need them. Rent rolls: keep unit, size, rent, and dates; delete the tenant name column. Ownership and pipeline exports: keep the deal name, stage, and dates; delete contact names, phone numbers, and email addresses. Sample emails: remove the recipient. If a document has tenant names or contact details in it, it isn't ready to upload.

Some things never go in. Investor lists, lender term sheets, anything with a Social Security number, bank details, or a signature. None of that is needed for any step here, and no setting makes uploading it a good idea.

The setup, in order

Step 1: Write your buy box as a one-page document

If your buy box lives in conversation, the assistant can't use it, and neither can a new analyst. Write it as plain text under five headings: asset type and class, size and price range, target markets and submarkets, return targets (the metric your team actually uses, whether that's going-in cap rate, stabilized yield, IRR, or price per unit), and deal killers (anything that ends the conversation regardless of the numbers: ground leases, rent control, environmental flags, hold periods that don't fit the fund).

You can tell this step worked when someone who's never sat in your investment committee could read the page and reject a deal for the right reason.

Step 2: Create a screening workspace and load the standard

Both ChatGPT and Claude let you create a workspace (ChatGPT calls it a Project, Claude calls it a Project too) that holds files and instructions every conversation inside it can see. Create one named "Deal screening." Upload the buy box from step 1. Then paste the following into the workspace's instructions field (ChatGPT: "Add instructions"; Claude: "Set project instructions").

You are a screening analyst on a commercial real estate acquisitions team. Your job is to read an inbound deal package and produce a first-pass screen against the buy box in this workspace.

Rules you never break:
1. Every number you report comes from the documents I upload. Cite the page or sheet it came from. If a number appears in two places and the values differ, report both and flag the conflict.
2. Never estimate a figure that isn't in the documents. If NOI, occupancy, or in-place rent is missing, say "not provided" and list it as an open item.
3. Treat broker pro forma figures as claims, not facts. Label anything from a pro forma as "broker projection." Only trailing actuals (T-12, rent roll, operating statements) count as in-place.
4. Do not pull in market data from memory. If a comp or market rent is needed, say so and leave it for me to supply.
5. If the package is a scanned image and the numbers are unreadable, tell me instead of guessing.

Output format, every time:
- Deal name, address, asset type, unit or SF count, asking price if stated
- Buy box fit: PASS, FAIL, or REVIEW, with the specific criterion that decided it
- The six figures: in-place NOI, T-12 revenue, T-12 expenses, physical occupancy, average in-place rent, asking cap rate (each with its source page)
- Deal killers found: yes or no, and which
- Open items: what's missing from the package that we'd need before spending real time
- Three questions to ask the broker
Keep it to one page. No summary paragraph at the end.

You can tell this step worked when you start a new conversation inside the workspace, type "what's in our buy box?", and it answers from your document rather than from general knowledge.

Step 3: Run a live deal through the screen

Start a new conversation inside the workspace. Upload the OM and any separate financials. Then send this:

Screen the attached package. Follow the workspace instructions exactly. Before you produce the screen, list every document you received and what each one contains, so I can confirm you're reading the right files.

That opening request matters. The assistant lists the documents first, you confirm it's looking at the T-12 and not the pro forma tab, and only then does it screen. Skipping this is the single most common way a screen goes wrong quietly.

You can tell this step worked when you get back a one-page screen with a page citation next to every figure.

Step 4: Check the screen against the source

Open the OM next to the screen. For each of the six figures, go to the cited page and confirm the number. Then check three things the assistant is most likely to get wrong:

  • Does the unit count on the screen match the unit count on the rent roll? If the OM says 120 units and the rent roll has 118 rows, the screen should have flagged it. If it didn't, the assistant is reading the marketing summary instead of the roll.
  • Is the NOI trailing or projected? Find the source sheet. If it's a pro forma, the screen should say "broker projection." If it says "in-place," the rule failed and you now know to watch for it.
  • Did the buy box verdict cite the right criterion? A FAIL on price per unit should point to your price range, not to a general opinion about the market.

You can tell this step worked when you've corrected at least one thing. If the screen came back perfect on your first live deal, check harder; that's unusual.

Step 5: Move upstream to sourcing

Once screening runs cleanly, the same workspace can take the research load off sourcing, with one important limit. The assistant does not know who owns the building at 400 Main Street, and if you ask, it will sometimes tell you anyway. Sourcing work goes to it in the other direction: you bring the data, it does the structuring.

Practical uses that hold up: paste an export from your ownership data source, with owner contact details stripped out, and ask for a ranked list against your buy box with the reason each property made the cut. Give it a target list and ask for a first draft of outreach for each, in your voice, from a sample email you upload with the recipient removed. Give it the last ten deals you passed on and ask it to find the pattern in why, which tells you what to stop sourcing.

You can tell this step worked when a list that used to take an afternoon to sort takes ten minutes, and every row still traces back to data you supplied.

Step 6: Move downstream to packaging and follow-up

Packaging is writing from documents you already have, which is the assistant's strongest ground. Upload the screen from step 3, the underwriting summary, and your team's last deal memo as a format sample, and ask for a draft in that format. The same pattern produces OM drafts on the sell side, with the same rule: every figure cites a source, and the human reads every one.

Follow-up is a memory problem, so give it memory. Once a week, export your pipeline (a CRM report, or just the spreadsheet), delete the contact columns, and ask: "Which deals have had no activity in 14 days, who's the last contact on each, and what was the last open item?" It produces a call list. You make the calls.

You can tell this step worked when your Monday pipeline review starts from a list instead of from scrolling.

Where it fails

These are the specific ways the output goes wrong, and what each one looks like when it happens.

Scanned OMs. Older packages and some regional brokers send image-only PDFs. The assistant either reads nothing or, worse, reads the few numbers it can make out and fills the gaps. The tell: page citations that don't match, or figures with no citation at all. Run the PDF through your scanner's OCR first, or ask the broker for the Excel.

Pro forma treated as in-place. The most expensive failure. An OM's financial summary usually leads with the projected year, and the assistant reads it as current performance unless told otherwise. The tell: NOI that's suspiciously round, or higher than what the T-12 supports. The workspace rule covers this, but check it every time.

Unit counts that don't reconcile. The OM cover says one number, the rent roll shows another, and the screen reports the cover. The tell: no conflict flagged. Any figure that appears in two documents should be checked in both.

Invented comps. Ask for market rent without supplying data and you may get a confident, specific, wrong number. The tell: a comp with an address you can't verify, or a cap rate with no source. Never let market data into a screen unless you brought it.

The package went in before the settings did. The first upload happens on a fresh account with training still on, or on a package whose CA nobody read. The tell is that there isn't one, which is the problem. Nothing looks wrong on the screen. The fix is the ten minutes in "Before you upload anything," done before the first deal rather than after.

Stale context across deals. If you screen two deals in one conversation, figures bleed from one into the other. The tell: an address or unit count from the previous package showing up in this one. One deal, one conversation, always.

The confident screen. Sometimes everything is cited and formatted and the verdict is still wrong, because the assistant applied a buy box criterion loosely. The tell: a PASS on a deal you'd have killed in ten seconds. This is why the verdict names its criterion, and why you read the criterion, not just the verdict.

How to check the output

Every screen gets the same check before it reaches anyone who makes decisions:

  1. Six figures, six source pages, six confirmations. Open the page, find the number.
  2. NOI source identified as trailing or projected, and labeled correctly on the screen.
  3. Unit count reconciled between the OM and the rent roll.
  4. Verdict criterion matches the buy box document, word for word if possible.
  5. No market figures on the screen that you didn't supply.
  6. Nothing went in that shouldn't have: training off, CA checked, names stripped.

If any of the six fails, the screen goes back with the correction, and you add a line to the workspace instructions that would have prevented it. That's the whole improvement loop. After ten deals the instructions are a page longer and the checks pass on the first read most of the time. They still get run.

None of this makes the assistant a substitute for underwriting. Nothing it produces on a lease, a rent roll, or an operating statement should reach a model without a person confirming it against the source. What it removes is the reading, sorting, and drafting around the underwriting, which is where the time was going.

What to do next week

Run every inbound package through the screen for a week, checking each one. Keep a note of what you corrected. At the end of the week you'll have two things: a workspace that reflects how your team actually screens, and a real number for how long screening takes now compared to before. That number is what tells you whether sourcing or packaging goes next.

If you'd rather have someone run this for you, this is the kind of work a Property Research Analyst or Sales Operations Assistant placed by Scale Partner does every day, on your documents, in your tools, with our AI infrastructure already set up. How it works explains the placement process, and the FAQ covers the common questions. For more workflows like this one, see the rest of the guides, or the blog for the people side of running a support team.