I Graded Every Message My AI Property Manager Sent to NYC Tenants

Every morning I do something that sounds slightly insane for a software founder.

I read every single message my AI property manager sent to tenants the day before, and I mark each one perfect or not perfect. Two grades. No middle. No “pretty good,” no “close enough,” no “well, in context.”

The blunt scale is deliberate. The second I allow myself a B-minus, I stop fixing things, and the whole product drifts toward the thing every landlord in New York already has too much of, which is software that technically works.

Maya is an AI property manager built for NYC owner-operators. She chases rent arrears over SMS, collects renters insurance certificates, takes maintenance work orders, and answers the ordinary building questions that eat a super’s afternoon. Half Ave builds her. I grade her.

I want to walk you through what I will call The Not-Perfect Pile, because everything I now believe about AI property management software came out of that stack. I have written before about reading 100 tenant conversations end to end. This is the narrower, meaner version of that exercise, run every day.

1. The worst failure was answering at all

Roughly 20 percent of my not-perfect grades were a single note: response not needed.

Picture it. A tenant in a walk-up in Astoria texts back “ok thanks.” Maya replies with a warm, correct, helpful paragraph about how happy she is to help and to reach out any time. Every word of it is accurate. Every word of it is a failure, because a good property manager reads “ok thanks” and does absolutely nothing.

General-purpose AI is built to answer. That is the entire product. Ask, receive. But a person who texts you back every time you say thank you is exhausting, and on SMS, where each message spends a little of your tenant’s attention, one unnecessary reply is how a system announces itself as a bot.

So Maya got a short circuit I call The Silence Rule. Acknowledgments, thumbs-up tapbacks, and thread-closers end the conversation. Silence became a valid output, and it is now one of the highest-value things she does all day.

Ask yourself how many property management platforms you have seen that measure how often the software stayed quiet.

2. Canned responses make tenant communication sound like a form letter

My first instinct was templates. Situation X gets message X. Predictable, auditable, safe. I was wrong, and I was wrong fast.

Real tenant messages refuse to sort into buckets. One arrived last month, lightly redacted: “sorry I’m sick, give me a few, I’ll send the screenshot.” That is an apology, a delay, and a promise, all inside eleven words. A template grabs one of the three and drops the rest, and you get a reply that is on-topic and obviously machine-written.

So the templates came out and the brain went in. Maya reads the whole picture before she writes: the ledger balance, the payment history, what this tenant said three weeks ago, whether the super ever closed out the radiator ticket. Then she writes the message herself, in her own words, for this one person.

What she does not get to decide is whether the message sends at all.

3. Rules live in code. Words live in the brain.

This is the sentence everything else in the product hangs off, so I am going to be tediously specific about it.

Code owns the decisions. Should this send. Has it already sent. How many times this week. Is this lease in housing court. Is the rent ledger fresh enough that we can safely quote a number to a tenant. Those live in deterministic application logic with an audit trail.

The brain owns the wording. Tone, what to acknowledge, what to leave out, how gently or how firmly to push on a past-due balance.

Get that backwards and you get six duplicate rent reminders at two in the morning, which is the exact story every landlord has heard about AI and the exact reason they hesitate. When a language model decides whether to send, your send cap stops being a cap and becomes a suggestion. The off-switch has to sit at the actual send, downstream of everything, rather than at the thing that triggered it.

Boring architecture. It is the reason I sleep.

4. The Living File, because a tenant told me she had already said this

One afternoon a tenant replied to Maya with something close to: I already told you that.

She had. Different thread, three weeks earlier. Maya had no memory of it, because every conversation started from zero. That behavior is normal for a chatbot and unforgivable for anyone whose job title includes the word “manager.”

Now every resident has what we call a Living File, an append-only set of notes Maya writes about them as she goes. CityFHEPS voucher status. Who is home during the day for a plumber. That 4B pays late every month and pays every month. That the ceiling leak got fixed in March and she is still nervous about it. Your own property manager notes live in that same file, and Maya can read them, though she can never overwrite them.

That single feature is the whole distance between a call center and a super who has known your building for four years.

5. The moat is New York housing law

Anybody with a weekend and an API key can wire a language model to a texting service. I could teach you. The hard part sits underneath, and it is the layer nobody is building for small owner-operators: RPAPL 711(2) rent demand sequencing, CityFHEPS voucher timing, rent stabilization, and the HSTPA rules that govern what you may and may not say to a tenant about money.

So Maya knows to leave a tenant alone when the voucher portion already covers the balance she would otherwise be chasing. She knows that a lease in active litigation means she stops talking entirely and hands over your office number, and that rule is enforced by a data switch rather than a line of prompt instruction, which means no clever tenant and no clever model can talk their way around it.

A national platform will not ship that for a landlord with 40 units in Queens. The math never works for them, which is the whole argument in our comparison with EliseAI, and it is the entire reason Half Ave exists.

If you want the long version of the voucher rules, we wrote a full CityFHEPS guide for NYC landlords.

What the grading actually changed

The scorecard made Maya quieter, more specific, and much harder to catch out. Her message volume went down. Her usefulness per message went up. Both of those are the point.

And the lesson underneath is one I keep repeating to anyone evaluating AI for property management in New York. Sounding human has very little to do with the writing. It is restraint, memory, and knowing the rules of this city well enough that you never have to hedge.

Go read yesterday’s messages from whatever software you are using. Mark each one perfect or not perfect. I suspect you will learn something uncomfortable, and I suspect you will fix it, because you now have a pile to look at.

Frequently asked questions

What is an AI property manager?

An AI property manager is software that handles the recurring communication work of running a building: rent reminders and arrears follow-up, renters insurance collection, maintenance intake and status updates, and general tenant questions. Maya works over SMS, which is where tenants actually answer, and escalates anything outside her rules to a human. She supplements your staff rather than replacing your super, your handyman, or your attorney.

Can an AI property manager collect rent in New York?

She can chase it, which is where most of the money is lost. Maya references the live ledger, sends the reminder, answers questions about the balance, and passes along payment instructions. Legal notices, including any RPAPL 711(2) rent demand, stay with you and your attorney. Maya follows the sequencing so her outreach never conflicts with a notice already in flight.

What happens if a tenant is in housing court?

Maya stops. A lease flagged as in litigation trips a data switch that halts all automated outreach to that tenant and returns your office number instead. The switch sits in application code rather than in a prompt, so it holds no matter what the tenant writes or how the conversation goes.

How does an AI property manager handle renters insurance compliance?

She tracks which units have a current certificate, texts the ones that lapsed, answers the predictable questions about what coverage the lease requires, accepts the document when it arrives, and keeps nudging the units that go quiet. Renters insurance compliance is pure follow-up work, which makes it one of the tasks AI closes out fastest.

What happens when a tenant reports a repair?

Maya takes the work order in conversation, asks the follow-up questions a good super would ask, records it, routes it, and then circles back to the tenant with status until the ticket closes. The follow-through matters more than the intake. Most maintenance complaints in New York escalate because nobody ever told the tenant what happened next.

Is this safe for rent-stabilized buildings?

Rent stabilization is the reason Half Ave is built for New York specifically. Maya works from stabilization status, voucher status, and lease state, and her rules reflect HSTPA. A national platform tuned for market-rate lease-ups in Texas has no concept of any of that.

How is this different from a chatbot on my website?

A website chatbot waits for someone to visit and start typing. Maya starts conversations over SMS, remembers each resident through a Living File that persists across months, operates under rules enforced in code, and stays silent when silence is the correct move. She is also built around existing tenants rather than around leads, which is where nearly every other AI leasing product points.

Do tenants know they are texting an AI?

Yes. Maya identifies herself, and every tenant can reach a human. The goal is a system tenants trust enough to answer, and pretending to be a person is how you lose that on day one.

See it on your own buildings

If you own or manage buildings in New York and you want to watch Maya work on your portfolio, email hello@halfave.co. Bring your worst arrears list.