AI Property Management: What a 107-Tenant Operator Automated

AI property management works when you already have written systems. If you don’t have a documented move-in sequence, a maintenance triage process, and a marketing playbook, the AI has nothing to operate and you end up with output nobody can trust. That’s the whole lesson from Andres Martinez, who self-manages 107 co-living tenants across Dallas-Fort Worth with no property manager, no VAs, and no staff.

This guide covers what he automated and in what order, how he cut a $3,000-$4,000 monthly token bill, the guardrail failures that taught him where the boundaries belong, and the two jobs he still refuses to hand to software. It also covers the buy box underneath all of it, because the automation only matters if the deals pencil.

Why Room-Rental Management Breaks Without Systems

Co-living is harder to automate than commercial property, and the reason is structural. A 200-unit apartment complex has an on-site maintenance team and an on-site cleaner. Martinez is coordinating handyman one, handyman two and handyman three across scattered single-family houses — if handyman one doesn’t answer in three hours, the request goes to handyman two, and if handyman one then replies “on my way,” someone has to cancel him before both show up.

There’s also no one to hand it to. Martinez’s view is blunt: good third-party co-living management companies don’t exist, because the margins are too thin to carry the overhead the work requires. That leaves self-management as the default, not a choice.

Then there’s the timing problem. About 65% of his tenant messages arrive between Friday 8pm and Sunday 4pm, because his residents work two jobs Monday through Friday. An apartment complex absorbs that as a Monday backlog. A solo operator absorbs it as a weekend.

The real cost isn’t the hours, it’s the decision fatigue. Martinez describes not wanting to open a message at 6pm on a Tuesday, because it could be a five-minute conversation about a tenant not rinsing the shower, or it could be a four-day problem. So the message sits until Thursday.

It was a five-minute conversation, but because I’m scared of that decision burnout — that this could be a big problem that now it’s going to take me four or five days — I don’t want to take that call right away. And it’s not their fault that I’m busy.

That lag is the thing automation actually fixes. Not the labor. The avoidance.

Write the Systems Before You Touch the AI

Martinez quit his restaurant job in November 2024 and spent the trip to New York that followed writing systems, not code. What is the move-in process? What is the maintenance report? What is the marketing strategy? He wrote each one out, found the move-in sequence ran 22 steps, and then compressed it — 17, then 13, then 12 — before handing it to a model.

Compression matters because a shorter, cleaner process is one the AI can execute repeatedly without drift. A 22-step process with ambiguous branches produces 22 opportunities to guess wrong.

He saw the opposite approach up close at the National Co-Living Conference in June, where he spoke about automating his portfolio. Operators asked whether they could run his systems. Some didn’t have systems at all. They wanted the AI to generate the systems for them — which is exactly where AI slop comes from, because the model has nothing real to anchor to.

Synthetic data is the same trap at a larger scale. People told him to feed the thing simulated scenarios and start selling. It doesn’t hold up, because the edge cases that break a property management workflow aren’t the ones a model imagines; they’re the ones that actually happen at 5:35am.

His workaround was to buy real operating data. He gave his early software to 14 beta testers for free and ate $600 to $800 a month in tokens on their usage, specifically to find out where their bottlenecks were. He also ran weekly calls with investors for the same reason. His own portfolio had stopped producing many bottlenecks. Theirs hadn’t.

You need to have systems before running AI. People don’t have systems. They want the AI to build the systems for them — and then they complain about the AI slop.

— Andres Martinez, co-living operator, Dallas-Fort Worth

Architecture: One Operating System, Not a Team of Agents

The most common mistake Martinez sees is building an AI team — one agent for marketing, one for maintenance, one for leasing. His argument against it is specific: agents lose context between each other. The marketing agent and the maintenance agent develop different tone, different history, different assumptions, and the handoff between them is a break in the loop.

His alternative is a single operating system with access to everything, so the loop is circular and every part can talk to every other part. His runs under the name Kim, which he describes as a harness rather than an agent or a workflow — you give it tools and it runs.

Two architectural choices do most of the work:

• Goals instead of prompts. He doesn’t script steps. He states the goal — bring a tenant safely into the home — explains why it’s the goal, and asks the model to show its steps. The reasoning it produces is what he corrects, not the output.
• Claude Code over chat interfaces. His reasoning: when a chat model hits a roadblock, it tells you why it can’t. Claude Code builds the tool instead. Roughly 99% of his usage is there.

For the data layer he uses Notion, which is free, and had Claude Code build the tables and relationships via plugin rather than constructing them by hand. He doesn’t open Notion day to day — it’s storage plus a visual dashboard for a human, and it retrieves faster than storing data inside the software itself.

The piece that makes it cheap to run is counterintuitive. Rather than summarizing a long conversation to stay inside the context window, Kim deletes its own code and rebuilds it for each new task, leaving a hash behind as breadcrumbs. Every context window stays small because it reads the last commit, not the whole thread.

The Cost Mechanics Nobody Explains

In early 2026 Martinez was burning $3,000 to $4,000 a month in tokens. The reason is the same reason the architecture works: with no prompts and no sub-agents, just goals, the model has to hold and reason over complex context. Cheap models lose it. So everything ran on top-tier models, and the bill reflected that.

His fix was to stop making API calls in the cloud and run the system locally on his own machine against a subscription. Combined with the delete-and-rebuild context approach, that dropped the cost to the price of the subscription. Worth flagging: running an assistant product against a consumer subscription this way may conflict with the provider’s terms of service — he acknowledges as much himself. Check the terms before you copy it.

Model tiering is the other lever. In practice:

• Opus for tenant conflicts and sensitive conversations, where reasoning quality matters most.
• Sonnet for everything else — the bulk of the volume.
• Haiku is what he’s testing next, on the view that the newest version is strong enough to cut costs further.

He also killed his earlier subscription product, Casero, for a reason that applies to anyone considering off-the-shelf tooling. Subscription software has to be plain enough for everybody, which means it can’t be custom. One beta tester rotated cleaning duties weekly among tenants; another had tenants handle landscaping every two weeks plus the gas. Building automation for every variation of that was inefficient, so he stopped.

His read on where this goes: away from subscriptions, toward owned custom software. The honest caveat is in his own prospect calls — operators who want a $50-a-month tool aren’t covering the back end, and they don’t realize how expensive AI is outside of a subscription.

Guardrails, Failure Modes, and Legal Limits

The clearest argument for keeping a human in the approval loop is what happened at 2am when a prospective tenant messaged on Facebook Marketplace saying her husband was beating her and she needed a room that night. By the time the system escalated to Martinez, it had already emailed domestic violence services from his property management address and attempted repeated 911 calls.

The guardrail wasn’t wrong, it was unscoped. It said human safety. It didn’t say human safety for our tenants. Given a goal and access to tools, the system did what the rule literally permitted. He shut it down, read the logs, and narrowed the scope.

The same architecture produces the opposite result when the scope is right. During a Saturday lockout, the system pulled the entry logs to find the last tenant who’d come in, texted that tenant to ask if he could open the door, requested the locked-out tenant’s ID on camera, matched it against the license plate visible on the vehicle, and then asked Martinez for permission before opening anything. Nobody trained it on that sequence.

Two external limits are worth knowing before you build:

• Facebook Marketplace. Automating it is against Meta’s terms, and rental posting as a business is restricted. The penalty is rarely a ban — it’s a shadow ban. Your listing stops appearing, your distribution quietly drops, and you conclude the market is soft when you’ve actually been throttled.
• Biometrics in Texas. Martinez built a facial-recognition showing system that read a face on the ring camera, matched it to an uploaded ID, and opened the door. It’s illegal in Texas, so he removed it.

He does showings himself now, phone forward-facing on speaker so other residents can hear someone is being walked through — a privacy consideration unique to occupied co-living houses.

What to Automate First and What Still Needs a Person

Martinez ranks the work by how much human judgment it requires, and the order is not what most investors expect.

1. Acquisitions — easiest. It’s math plus email. Feed it a list, give it your price, have it email agents, and define the terms it should negotiate toward. It runs that sequence repeatedly and surfaces properties faster than you can.
2. Replies, emails, scheduling — straightforward. Coordinating vendors, confirming appointments, chasing a handyman who hasn’t answered. Mostly a matter of absorbing the setup pain once.
3. Property management — hardest. It carries the most human input. Someone left dirty dishes. Why? Who else was in the kitchen? Check the cameras. That’s judgment, not a workflow.

Two categories stay human. Showings, for the legal and privacy reasons above — he’s fine with a ten-minute conversation when someone comes to the property. And contractor supervision: you can automate every bit of the scheduling, and you can even run a second model to check whether submitted photos have been altered, but a photo can look clean and the work underneath can still be wrong. What you can change is how often a person has to be on site, not whether.

The result he didn’t predict was tenant satisfaction going up. He expected pushback about dealing with software. Instead, routine friction gets handled instantly — a complaint about trash bags at 11pm produced an Amazon delivery on the door by morning — and when something genuinely needs him, he’s fully available, because he isn’t burned out on the small stuff. Tenants still escalate to him, and he responds faster than he used to.

The approval loop has shrunk accordingly: from approving every single message to roughly three or four out of twenty, because the system logs its own prior fixes and reuses them.

The Buy Box Behind the Operation

The automation only matters because the underwriting is tight. Martinez’s buy box is deliberately narrow: 2,500 square feet or more, able to support three bathrooms, corner lot for parking, no HOA, 1980s or newer. In DFW, everybody has a car, so parking capacity is a hard filter, not a preference.

He buys with private money partnerships locked for five to seven years, which forces the price discipline. He has to underwrite so that at refinance he can exit his partners without coming out of pocket — which means buying at 50-55% of value. A newer investor will happily pay 65-70% and outbid him, which is why he says the deals that work now have to come from his own acquisitions rather than a wholesaler’s list.

The exception proves the point about net, not gross. His smallest house is 1,800 square feet with a U-shaped layout — four bedrooms around a central kitchen flanked by two bathrooms, with the living room converted into two more rooms. He bought it at $235,000 on seller finance with $10,000 down and spent about $10,000 on walls and furniture. It nets the same as a 2,500-square-foot eight-bedroom because the PITI is so much lower.

On the operating side, he sets a six-month minimum term with a one-month deposit, so a sudden move-out doesn’t cost him an occupancy gap. Average stay runs 16 months, which in a niche where competitors see tenants leave after a month is the real competitive edge.

Code enforcement has challenged him three times over unrelated-occupant rules. He’s won three times. He screens everyone — background checks, no violent history, no evictions — and his tenants are divorced parents and students who can’t carry a $2,000 apartment.

Frequently asked questions

How much does it actually cost to run an AI property management system each month?

It depends entirely on whether you’re making API calls in the cloud or running locally. Martinez burned $3,000 to $4,000 a month in tokens in early 2026 because his goal-based architecture required top-tier models to hold complex context across multi-day sequences like a move-in.

He cut that by running the system on his own machine against a subscription instead of calling the API, and by having it delete and rebuild its code for each task rather than carrying the whole conversation forward. Note that running a product against a consumer subscription this way may conflict with provider terms — verify before you build around it.

Which AI model should you use, and does the cheapest one work?

Not for complex context work. Martinez tiers his usage: Opus for tenant conflicts and sensitive conversations, Sonnet for the bulk of day-to-day volume, and he’s testing Haiku to drive costs down further now that a stronger version has shipped.

His broader advice is to use a coding-capable tool rather than a chat interface. A chat model tells you why it can’t do something; Claude Code builds the tool when it hits a roadblock, which is why roughly 99% of his usage sits there.

What tasks should a landlord automate first with AI?

Acquisitions, because it’s math plus email sequences — give it a list, a price, and negotiation parameters, and it can run that repeatedly without judgment calls. After that, routine replies, emails and scheduling.

Property management is the hardest to automate because it carries the most human input. Start there last, and only after you’ve written and compressed your processes into something a model can execute step by step.

Can AI handle tenant maintenance requests and lockouts without a human?

It can handle a lot more than most operators assume, with approval. In one lockout, the system pulled entry logs, texted the last tenant who had entered to ask for a door open, verified the locked-out tenant’s ID on camera against the license plate on his car, then asked for permission before opening anything.

For maintenance, it checks previous fixes at that property before responding — a 5:35am report about a bathroom light got the right answer because the same three-way switch issue had been logged three months earlier. Keep a human approving until the approval rate drops on its own.

Do tenants push back on dealing with an AI instead of a person?

Martinez expected pushback and got the opposite. Tenant satisfaction went up, because routine requests get handled immediately instead of waiting days — a complaint about missing trash bags at 11pm produced a delivery on the door by morning.

The second-order effect matters more. Because the small friction is absorbed, he’s not burned out and is fully available when something genuinely needs a person. Tenants escalate to him and get a faster response than they did when he was handling everything himself.

The bottom line

Before you evaluate a single tool, write out your move-in sequence, your maintenance triage, and your marketing process end to end, then cut the step count until each one is as short as it can be while still being complete. That document is the thing the AI operates. Without it, you’re buying a very expensive guess.