Field Stories10 min read

2 a.m., Water Through the Ceiling: A Property Manager's Night

By Ergini, Software & AI Developer

A composite story. The company and the people in it are invented. The problem, the rules and the system are real, and the full blueprint is in the use case library.

TL;DR

A composite story: a four-person property management firm in Linz with 1,200 flats puts its emergency number on a single on-call phone, so a dripping tap wakes the same person as water through a ceiling. The agent I would build takes WhatsApp and email requests in the tenant's language, applies an emergency protocol the firm wrote, phones the on-call person only for real emergencies and lines up everything else with a trade by morning. Staff approve costs above the limit, and liability and rent stay with people.

21:40, a dripping tap on Stiege 2

Lena is on call this week. The notice in every stairwell of 45 buildings gives one number for emergencies and WhatsApp, and this week it rings the on-call phone on her coffee table. It is a Thursday in November, she is on the sofa in Linz, and the first message arrives at 21:40:

Küchenhahn tropft seit gestern. Top 5.

Kitchen tap dripping since yesterday, flat 5. No building, no photo. Lena recognizes the name, so she knows the building. She replies that someone will call tomorrow and writes it on a notepad, because the work orders live in the property management system on a PC in the office, not on her phone.

The firm is invented: four staff managing about 1,200 flats in 45 buildings, rental buildings for private owners and condominium buildings for owners' associations. Three of them take the on-call phone a week at a time, and it goes home with whoever has it.

At 22:55 a tenant writes that he is locked out, key inside. Lena sends him a locksmith's number from memory and reminds him, gently, that the call-out is at his own cost.

23:30, a voice note in Turkish

The next message is forty seconds of voice note, in Turkish. Lena does not speak Turkish. She writes back in German asking the tenant to put it in writing, and gets no reply. In the morning it will turn out to be no hot water in one flat on Stiege 1: urgent in November, not dangerous, and possibly one boiler reset away from fixed.

At 01:15 the phone buzzes again, an email to the repairs address about the intercom on Stiege 4, forwarded to the on-call phone because nobody wants to miss the one that matters. Lena turns the phone face down. She has an owners' meeting at nine.

This is the whole problem in one evening. The number is built to reach a person, so everything reaches the person. A dripping tap feels exactly the same in the hand as water coming through a ceiling, and after the third message of the night, nobody can tell them apart before they have read them.

02:04, water through a bathroom ceiling

The phone rings face down and keeps ringing. It is Ioana, Top 9 on Stiege 3, who wrote on WhatsApp first and called when nobody answered. Water is coming through her bathroom ceiling, around the light fitting.

Lena is awake at once. She tells Ioana to keep the bathroom light off, stay away from the wet switch and close the flat's main water valve if she can find it. Then she needs three facts she does not have in bed: who lives above Top 9, their phone number, and which plumbing firm holds the emergency contract for that building. The first two are in the property management system, which she reaches through a slow remote desktop session on her laptop. The third is in a binder at the office. She calls the plumber she remembers, who covers other buildings and comes anyway.

The tenant in Top 13 answers on the fifth call, at 02:31. By then the stain on Ioana's ceiling is the size of a tea tray. The plumber arrives at 03:20, finds a burst hose under the sink upstairs and closes it off.

Lena is back in bed at four. In two hours she made one decision that needed her: send a plumber, now. Who pays can wait for the plumber's report. Everything else she did was looking things up.

07:30, three reports of one leak

By half past seven the repairs inbox has three more messages about the same water: a damp wall in the flat below Ioana's, the cleaning firm reporting a wet stairwell, and Ioana asking whether anyone is coming to look at the ceiling. They become three tickets, picked up by three people. The Turkish voice note is still unanswered.

That afternoon Martin, who owns the firm, books a video call with me and puts it in one sentence: only real emergencies should reach whoever is on call, and everything else should be lined up with a trade by morning. It is simple to say and hard to build, and most of the hard part has nothing to do with a model.

I ask him to hold the binder up to the camera. It is the most important document in the firm: which plumber, heating firm, electrician and lift company holds the contract for which building, with their night numbers. Lena needed it at 02:10, and it was on a shelf across town. Before an agent reads a single message, that binder becomes data, together with who is responsible for what in each building and the cost limit in each management contract.

Their property management system has no open API. That settles one design question early: the agent keeps its own work-order record and exchanges files with the system on a schedule, and nobody has to change the software they manage the buildings with.

A protocol written in daylight

The emergency protocol comes next, and the firm writes it, not me and not a model. It lists the situations that are always emergencies and the exact words a tenant receives in each, agreed in plain German and translated once into the languages the tenants write in. A smell of gas: leave the flat, use no switches, flames or lift, and call the gas emergency number 128 from outside. Fire or smoke: call the fire brigade on 122, and the first reply says nothing else. Water through a ceiling: close the main valve if it is safe, keep away from wet sockets and switches, send a photo.

The model never composes safety advice. It recognizes the situation, and code sends the approved text for it, word for word. Nor is the model the only thing listening. A tenant who writes "probably nothing, but the kitchen smells a bit of gas" sounds calm, and a model that weighs tone could file it as routine. So a keyword list in every supported language raises any mention of gas, smoke or sparks, however it is phrased. The model or the list can raise a message to emergency. Only the on-call person can lower it again.

The on-call person gets a phone call, not a text, because a text at two in the morning is easy to sleep through. If nobody answers within a few minutes, the next person on the rota is called. And every tenant's first reply says that an automated assistant is answering and how to reach a person, as Article 50 of the EU AI Act requires.

The same Thursday, run again

Here is Lena's night once more, message by message, as the agent would take it.

TimeWhat came inThat nightWith the agent
21:40Dripping tap, Top 5Lena replies and writes it on a notepadTenant matched by phone number, photo requested, work order for the morning
22:55Locked out, key insideLena sends a locksmith's numberAcknowledged, with the locksmith list at the tenant's own cost
23:30Voice note in TurkishNo reply possibleTranscribed: no hot water in one flat. Urgent, not an emergency. The approved checks in Turkish, first slot next working day
01:15Intercom emailForwarded to Lena's phoneWork order for the morning; the phone stays silent
02:04Water through Ioana's ceilingA call, after a WhatsApp nobody sawApproved leak text in Romanian at once. Lena phoned, with the flat above, its tenant's number and the contracted plumber on her screen
07:30Damp wall below, wet stairwell, Ioana askingThree tickets, three peopleLinked to the same work order; each reporter told it is known

In that version Lena is woken once, by the one message that needed her, and she wakes to a summary instead of a search. The model's part is narrow: understanding a Turkish voice note, describing Ioana's photo as a wet patch above the shower with water dripping from the light fitting, for Lena to confirm, and writing the updates. Urgency, the protocol, the phone call, the duplicate check and the contractor's details are rules anyone at the firm can read.

Lena approves the plumber, and the agent sends the firm the address, access details, Ioana's number and the photos, and nothing else from her file. Lena still calls Top 13 herself; waking a neighbor at two in the morning is a person's job. The full design is in the blueprint for maintenance request triage.

Friday, 10:12: "Do I still pay full rent?"

The morning after, Ioana writes again. Her ceiling is stained, the bathroom light is off until an electrician checks it, and she wants to know whether she still owes the full rent this month. In Austria, as in Germany, tenants can reduce rent for serious defects, and a repair request can turn into that argument in one message.

The agent does not take part in it. It does not discuss rent, liability or legal positions at all. Ioana is told that a person will reply, and the message reaches Lena with a two-line summary and the work order attached.

The same line runs through the money. Who pays for the ceiling depends on the building type, the lease and the law, the MRG for many rental flats and the WEG for condominiums, and the firm's staff write those rules per building. The agent is never asked who is liable. It is asked what happened and where, and it shows which rule it applied. Here the cause stays unclear until the plumber's report, so a person allocates the cost afterwards, and anything above a management contract's limit goes to the owner or the owners' association as an approval request with photos and a quote.

Every dispatch has a clock. If a routine job is not confirmed in time, the agent tries the second firm on the list; for anything urgent it tells staff. A work order closes only when the tenant says it is fixed and the contractor has sent a completion report. Photos of homes stay in EU hosting, and a tenant's mention of illness is kept as a priority flag, not as text.

If the emergency number is someone's mobile

In US multifamily, buy: EliseAI, AppFolio's own AI features and Property Meld cover much of this. In the UK, Fixflo is the established repair-reporting tool, and in Germany and Austria, Casavi gives tenants an app and staff a ticket queue. If your tenants use those tools and your rules are standard, start there.

A build earns its place in German-speaking residential management, where condominium law, owners' associations and privately let flats in administered buildings change who is responsible for each repair, where tenants will not install an app but already write on WhatsApp in six languages, and where the management software has no open API. It is usually a single-purpose AI agent build, and the price moves with the channels, the languages and the number of buildings and contracts the rules must cover. The full blueprint has the protocol table and the responsibility rules, and the same intake can later answer the phone out of hours through an after-hours call agent.

Frequently asked questions

Can an AI agent handle tenant repair requests on WhatsApp at night?

Yes, for intake, triage, dispatch and follow-up. The agent says it is automated and how to reach a person, asks for photos and missing details, and applies an emergency protocol your team wrote in advance. Real emergencies get the approved safety text and a phone call to the on-call person; routine requests become work orders for the morning. Updates sent after Meta's 24-hour window use approved templates.

How does an AI agent decide what counts as a maintenance emergency?

Not on its own. A protocol table agreed with your team lists the situations, such as gas, fire, active leaks, electrical danger and building-wide heating failures, and code applies it. A keyword list in every supported language raises any mention of gas, smoke or sparks, however calmly it is phrased. The model or the keyword list can raise a message to emergency; only the on-call person can lower it.

Who approves repair costs when an AI agent dispatches contractors?

Your staff, through rules they write per building and contract. Within a management contract's cost limit, the agent dispatches the contracted firm and shows which rule it applied. Above the limit, it prepares the approval request with photos and a quote for the owner or the owners' association. It is never asked who is liable, and it does not discuss rent reductions or legal positions with tenants.

What does it take to build a maintenance triage agent?

Usually a single-purpose agent, the first tier of AI agent development. The effort depends on the channels, on whether the property management system has an open API or needs a scheduled file exchange, on how many languages the safety texts are translated and approved in, on how many buildings and contracts the responsibility rules cover, and on the on-call rota and phone escalation.