AI for Property Management in the Bay Area

Eight files land in a folder by 4am. One of them is a detail ticket report that runs anywhere from 53 to 200 pages.

It's a hotel group, near 10 properties, and every morning one person opens all of it and checks the previous day's revenue by hand, property by property. Card settlements against the property management system. No-shows. Cancellations that posted late. Credits that need a question sent back to the property. They boil it down to a short list and send it.

Then the file gets binned. The owner's line: "Two days later it becomes irrelevant."

So credits get issued for the wrong reason, nobody catches it, and by the time anyone thinks to look there is nothing to look at. That is a job nobody has time to do properly and everybody is doing anyway.

What I actually built

A folder the nightly reports drop into. A rules document that spells out what counts as revenue, which figure comes from which report, and what to flag. And a task that reads yesterday's files each morning and writes the summary.

The rules document is where everything lives. It names the properties in scope. A negative posting on the detail ticket is a credit to the guest, tax lines are never room revenue, and no-shows post up to 48 hours late so today's get paired against yesterday's file. Above a set dollar amount, an adjustment gets a question sent back to the property.

And it has four lines in it under the heading of what the thing must never do. The first one is "Invent a number that is not in a file." The second is "Pick one figure when two files disagree." If two reports disagree it has to show both and say so.

What the first run actually got

I put the output next to their handwritten list and went through it line by line.

It caught 16 of their 22 items. It missed 6. And it added about 8 things they would never list, like tax voids and wifi adjustments at $9.95 a go that nobody needs to see at seven in the morning.

Anyone showing you a demo where it works perfectly first time is showing you a demo.

Seven fixes came out of that comparison. Early departure fees got their own line. Balances got restricted to money owed to the hotel. No-shows take the posted amount only. Future-stay cancellations came out entirely. The fee discrepancy questions, which were technically correct and useless, came out too.

Every one of those was an edit to the rules document. No rebuild, no new tool, nobody touching code. Then I re-ran it against the same day's files and it matched her list 100%, line for line. The doc went from v2 to v4 in about two weeks and the build underneath it never changed.

Halfway through I moved it onto Claude Opus 5.5, which runs about 40% cheaper than the model it replaced. Same rules file, same output, cheaper every single morning. Of everything in the stack it's the one cost that falls without anybody touching it.

When the output is wrong the instinct is to assume the AI is wrong, and nine times out of ten it is reading the wrong file or missing a rule nobody wrote down. The rules document is everything, I can’t stress this enough.

I watched a construction firm spend around $250,000 finding that out the hard way, which I wrote up in AI for construction in the Bay Area.

And when the franchise brand pushes out new systems and the report formats change, that's a rules file edit per property, which is the answer to the thing that stops most groups from starting.

It does not go live until five clean mornings

I use the same gate on every install now. It’s the same rule I used with an interior design studio in Tiburon: run it by hand until it’s boring, then automate it.

The client runs it himself, by hand, every morning. It is finished on the fifth consecutive morning where the output matches their list and needs no correction. Until then nothing goes on a schedule.

It's slower than any vendor will promise you, and it's the only reason the thing survives a real week.

Bay Area property management companies have the same hole

Around the same time I'm on a call with the owner of a property management firm, out of state this one, two partners and about a thousand multi-family units between them. I ask where the day goes expecting leasing or tenant complaints. He says work orders. Twenty or thirty a day that nobody ever closed out.

A vendor turns up, does the job, submits the invoice and gets paid, then never goes back into the portal to tick the last box. So somebody on his side chases a plumber who has already been paid to confirm he did the thing he already did. His words: "maybe you have 20 or 30 work orders in a day, and that carry on next because the process was unfollowed appropriately by an external party."

Different trade, same hole.

He'd already run the maths on replacing a property manager with an assistant service, out loud, without being asked: "We won't have to pay someone, you know, $70,000 a year versus $3,000 a month." A Bay Area landscape contractor did the exact same maths with me, same $70k a year, different trade entirely. Owners weigh this against a salary line, never against software.

Start with one and find out 

I went into that hotel group asking about one process at one property. A night audit summary. Small, finite, easy to walk away from.

The owner scoped it himself out loud and then widened the starting scope from one hotel to all eight without me asking. On how he wanted it priced: "this is the number of hours and this is the price per hour ... And then we'll make that commitment and then we'll see. But we won't get any commitment beyond that."

Fair play to him, because that is a better buying process than most of the proposals I have sent. Prove one process, pay for the hours it takes, commit to nothing beyond it. One property is now done and running. The list for what comes next is four more property summaries, daily revenue checks across five more, and a morning note telling the principal which five emails actually need him today. None of that was on my proposal. It came out of the first one working. 

If you can’t pick the one, I wrote about what to automate first in a small business. 

One thing Marin and East Bay property managers can do this week

If you run buildings or rooms anywhere in Marin or the wider Bay Area, whether that’s a handful of units in San Rafael or Novato, or a portfolio across Oakland, Walnut Creek and San Jose, pick the one report somebody on your team rebuilds by hand every morning and then throws away. Not the big system you hate. The small daily one nobody has ever written down.

Send me a week of it and I'll tell you what the rules document for it looks like and what it takes to get it running clean. I’m in Tiburon, twenty minutes from most of North Bay and across the bridge in under an hour, so if it is worth doing I'll come to your office and build the first one beside you. And I'll give you the price in the room rather than emailing you a proposal three days later.

As always, if you would like to organize a walkthrough, just visit me at Jaimate.com 

FAQ

What can AI actually do for a property management or hotel group? Close the daily loops: the revenue or occupancy report somebody rebuilds by hand every morning, unclosed work orders, vendor invoices nobody chased.

Do I have to replace my property management system? No. The right build reads the files it produces and writes a summary. Migrating a portfolio off its system of record costs more than the problem.

How long before it is reliable? The hotel build above runs by hand every morning and only goes on a schedule once it matches the client's own list five days running. Expect weeks, not an afternoon. Now, it’s done in under a minute each morning. 

Do you work with Bay Area property management and hospitality operators? Yes. Marin first because that's where I'm based, and across the nine Bay Area counties from there, in person.

Do you work with property managers in San Francisco, Oakland and San Jose?

Yes. I’m based in Tiburon and work in person across Marin, San Francisco, the East Bay and the South Bay. Marin and San Francisco jobs I can usually be at the same week.