I Read 100 Conversations Between My Tenants and an AI. Here Is What Surprised Me.

I run my own buildings. For the past few months an AI has handled the tenant messaging in them: arrears, insurance, maintenance intake. So one weekend I did something slightly obsessive and went back and read every transcript.

Over 100 real conversations, across nearly 200 residents. Every single one with money or compliance riding on it.

I sat down expecting to grade a chatbot. That is not what I found, and the difference turned out to be the whole story.

Four things stuck with me.

1. Calling it a chatbot gets it exactly backwards

A chatbot is that little help bubble in the corner of a website. It waits for you. You have a question, you open it, it answers, it goes back to sleep.

Reactive by design. Its entire job is to be sitting there when you come looking. Order status, password reset, return policy. The product already works, and the bot tidies up the tail end.

What I was reading did the inverse.

The tell was sitting right there in the message log. It sent more messages than it received. Roughly 950 outbound against 590 inbound. It starts most of the conversations.

A chatbot cannot do that. A chatbot has nothing to say until spoken to. This thing texts the resident first, because chasing rent and lapsed insurance is something you go out and initiate rather than something you sit around waiting to be asked about.

That one ratio reframes everything. The tenant is answering a property manager who reached out first, and the property manager happens to be software.

2. What it is good at: the grind

I want to name the category of work that filled these transcripts, because I think the category is the actual product. I will call it The Relentless Eighty.

• 💵 Reaching every delinquent resident, every cycle, and running the demand sequence correctly each time
• 🛡️ Following up on renters insurance, chasing lapses, confirming coverage, logging it
• 🔧 Taking maintenance requests as they arrive, day or night
• ✅ Tracking who promised to pay what, and whether the money actually landed

None of this is hard in the clever sense. It is hard in the relentless sense: high volume, unforgiving on timing, and thoroughly thankless. It is the kind of work that drowns a capable person, and frankly the kind of work a capable person is wasted on.

This is where the machine shines. It does not get tired on the fortieth arrears message. It does not forget to circle back. It does not skip the resident nobody wants to call, and let me tell you, there is always a resident nobody wants to call.

Across 100-plus conversations, the routine middle of the job simply ran. Nobody touched it. That is why you hand it over.

3. What it is bad at, which I care about a good deal more

If I told you only the wins, I would be selling you a demo, and you would be right to close the tab.

So here is the honest half. The AI is good right up until the conversation stops being routine. The moment a case needs judgment instead of a rule, it should stop and hand the resident to a human. The system is built so that it does exactly that:

• 💔 A resident in genuine crisis, where the right response is a conversation rather than the next step in a sequence
• ⚖️ A legal gray area where the regulation refuses to apply cleanly
• 🏛️ A voucher or subsidy timing question where the true answer is it depends, and somebody has to decide what it depends on
• 🧭 Any moment where the correct move is deciding what the rule should be here, instead of following the rule

That last twenty percent belongs to a human being. Trying to automate it is how you end up as a viral screenshot of a bot saying something idiotic to somebody who is struggling.

The skill lies in knowing precisely where to stop.

One more thing it is bad at: pretending. The stakes in this business are a long way from a mistaken return policy. A botched demand sequence or a mistimed compliance notice is a legal liability.

Which means the rules that carry weight cannot live in a friendly prompt and a hopeful spirit. They have to be hard-coded and identical every single time. Sounding helpful is the easy part. Being correct about regulated, jurisdiction-specific work, every time, unsupervised, is the actual product.

The takeaway

A chatbot answers a question after the work is already finished. It is reactive, it serves the tail, and it is measured by how few humans it drags into the room.

An assistant property manager does the work itself. It reaches out first. It carries The Relentless Eighty that grinds people down. It hands you the twenty that requires a person.

Same two letters. Completely different machine.

After 100-plus conversations, here is the cleanest way I can put it. A chatbot is the last mile. Maya is the first eighty.