You've been using AI like a very confident intern who has never seen your portfolio. It writes decent email copy. It has no idea that Building 3 has eleven vacant one-bedrooms and a comp two blocks away just dropped to $1,485.
That changes the second you start feeding it your actual data. Not "our occupancy is around 93 percent" typed into a chat box. The real export. The rent roll, the ILS lead report, the Google Ads CSV. Once the model can see the numbers, it stops writing marketing and starts answering questions you'd otherwise wait three days for someone to pull.
Here's how to do it without handing resident PII to a stranger.
Step one: pick a workspace that doesn't train on you
Before any file gets uploaded, check one setting. In ChatGPT, that's a Team or Enterprise workspace (business plans exclude your data from training by default). In Claude, the paid consumer and Team plans don't train on your conversations. Free tiers and personal accounts are where policy gets fuzzy and your legal team gets loud.
If you're not sure which plan your company pays for, ask IT before you upload anything. That's a two-minute Slack message, not a project.
Then: turn off chat history retention if your plan allows it, and never paste anything into a random AI tool you found on LinkedIn last Tuesday. The good ones are boring and cost $25 to $30 a seat.
Step two: strip the file before it leaves your laptop
This is the part people skip and it's the only part with real downside.
Your rent roll has names, phone numbers, emails, sometimes lease guarantor info and balances. None of that helps the model answer a marketing question. So take it out.
Open the export in Excel. Delete these columns entirely:
- Resident name, co-resident, guarantor
- Phone, email, emergency contact
- Social security or ID fields (if your rent roll has these, that's a separate conversation with your PM software rep)
- Ledger balances and payment history
Keep these:
- Unit number, floor plan, square footage, bed/bath
- Market rent, actual rent, concession amount
- Lease start, lease end, move-out notice date
- Occupancy status, days vacant
- Building or phase
What's left is a marketing dataset. Unit 204 is a 1x1, 720 square feet, market rent $1,550, leased at $1,425, expires 3/31. No human being is identified. Save it as "Property_RentRoll_Marketing.csv" so nobody grabs the wrong version later.
Do the same to your ILS export: keep source, date, unit interest, tour status, lease status. Drop the prospect's name and phone. You're looking for patterns, not people.
Ad reports from Google, Meta, or your ILS dashboards are already aggregate. Upload those as-is.
Step three: give it context, not just a file
A CSV with no explanation gets you a summary nobody asked for. The model needs to know what it's looking at and what decision you're trying to make.
Structure it like this: what the file is, what your situation is, what you want out of it.
"Attached is the marketing rent roll for a 240-unit garden community in suburban Charlotte, exported 3/12. We're at 91 percent occupied with 34 leases expiring between now and June. I need to know which floor plans are exposed and where I should be pushing ad spend."
Now it can actually work. Ask it to build an expiration curve by floor plan and flag any month where more than 8 percent of a plan type rolls at once. Ask which units have been vacant longest and what they have in common. Ask it to calculate your average concession as a percentage of market rent, by plan, and tell you where you're discounting hardest without knowing it.
That last one is a personal favorite. Concessions creep. Nobody decides to give away six weeks free on 2x2s. It just happens one approval at a time, and then it's in your effective rent.
Step four: the multi-file questions are the real prize
One file is useful. Two files stacked is where you stop being a marketer and start being the person in the room with the answer.
Upload the ILS lead export and the ad spend report together. Ask for cost per lead, cost per tour, and cost per lease by source. Then ask which source has the worst tour-to-lease conversion despite decent lead volume, because that's the contract you should be renegotiating in Q3, not the one with fewer leads.
Upload the rent roll and the ILS export. Ask which floor plans are getting the most inquiries relative to their share of upcoming availability. If your 1x1s generate 60 percent of leads and represent 25 percent of your exposure, you're spending money on demand you can't house.
Upload two months of lead reports. Ask what changed. Not "summarize," ask what changed and what could plausibly explain it. Models are good at spotting the drop you missed because you were looking at the portfolio total.
Step five: verify, always, and it takes ninety seconds
Models do arithmetic wrong. Not often, but enough that presenting an unchecked number to asset management is a bad career plan.
Two habits fix this. First, add "show your work: list the rows you used for each calculation" to every prompt involving math. Second, spot-check one number yourself. If it says average effective rent on 2x2s is $1,712, filter your spreadsheet and check. If that one's right, the rest usually are. If it's wrong, the whole output is suspect and you re-prompt.
Also worth knowing: models sometimes read a partially loaded file and confidently analyze the first 200 rows of a 340-row export. Ask it how many units it counted. If the number doesn't match your unit count, make it recount.
The fair housing note
When you move from analysis into copy, watch what the model describes. Ask it to write ad copy off your rent roll data and it will happily produce "perfect for young professionals" or "ideal for a growing family." That's describing who should live there, which is exactly the thing fair housing law prohibits in advertising.
The fix is one line in your prompt: describe the unit and the community amenities only, never the type of person who would live here. Add it to every copy prompt you write. Then read the output before it ships, because the instruction works about 90 percent of the time and you own the other 10.
Same rule applies to targeting. If you're using data analysis to inform audience segments, keep it to geography, search behavior, and unit features. Never demographics.
What to do this week
Export one rent roll. Strip it. Upload it with real context and ask three questions you'd normally email a regional about. Time how long it takes.
Then do the ILS and ad spend pair, because that one changes your renewal conversations.