Use caseAI agentWorkflow automation
Maintenance requests triaged, dispatched and chased, with a real emergency path
An AI agent for tenant repair requests on WhatsApp and email: a fixed emergency protocol, contractor dispatch within your limits, and follow-up until closed.
A blueprint, not a client story. The business described is illustrative; the architecture, integrations and trade-offs are real, and this is how I would build it. By Ergini, .
The short version
An AI agent for property managers that takes maintenance requests from tenants on WhatsApp, email and the tenant portal, asks for photos and details, and applies an emergency protocol written in advance: gas, active leaks, electrical danger and heating failures reach the on-call person at once. Everything else becomes a work order in the property management system, dispatched to the contracted trade within cost limits and followed up until closed. Staff confirm responsibility, costs above the limit and every emergency decision.
- Best for
- Property managers with a small team and a few thousand units, whose tenants report repairs on WhatsApp, email and the phone at all hours.
- Connects to
- WhatsApp Business Platform, Property management system, Outlook, Contractor scheduling, Twilio voice and SMS
- The AI does
- Understands requests in the tenant's language, asks for the missing details and photos, describes what the photos show, and keeps tenants and contractors updated.
- People do
- Handle emergencies on call, confirm responsibility and approve costs above the limit, and take every dispute about liability or rent.
- Built as
- AI Agent Development, usually $12K - $30K
Four people, 1,800 flats, and a phone that rings at night
Take a property management firm in Vienna with about 1,800 units across 60 buildings: rental buildings managed for private owners, and condominium buildings managed for owners' associations. Four staff handle everything from service-charge statements to broken intercoms. Repairs arrive by phone during office hours, by email, through the tenant portal and, more and more, on the WhatsApp number posted in every stairwell.
Each request starts the same loop. Someone asks for a photo, asks again, works out whether it is urgent and who is responsible, finds the contracted plumber for that building, calls, waits for a date, tells the tenant, and a week later asks whether it was fixed. Tenants write in German, English, Turkish, Bosnian, Serbian, Croatian, Polish and Romanian. The same leak gets reported by three neighbors.
Nights and weekends are worse. The emergency number rings a manager's personal mobile, so a dripping tap at 23:00 wakes the same person as water pouring through a ceiling. What managers ask for is simple to say and hard to build: only real emergencies should get through, and everything else should be lined up with a trade by morning.
The emergency protocol, written before the first message arrives
The protocol is agreed with your team before launch, in plain language, and translated once. The model never composes safety advice. It recognizes the situation, and code sends the approved text for it, in the tenant's language.
| What the tenant reports | Treated as | What the tenant is told at once | Who is alerted |
|---|---|---|---|
| A smell of gas, a gas alarm | Emergency, always | Leave the flat, use no switches, flames or lift, and call the gas emergency number 128 from outside | The on-call person, by phone call, immediately |
| Fire or smoke | Not a triage case | Call the fire brigade on 122 now. The first reply contains nothing else | The on-call person, once the tenant has the number |
| Water coming through a ceiling, a burst pipe | Emergency | If it is safe, close the flat's main water valve, keep away from wet sockets and switches, and send a photo | The on-call person, with the contracted emergency plumber ready to call |
| Sparks, a burning smell, exposed wires | Emergency | Touch nothing, and switch off at the fuse box only if it is dry and safe to reach | The on-call person, then the electrician on contract |
| No heating or hot water in the heating season | Emergency for a whole building, urgent for one flat | For one flat, the approved checks: the boiler's error code, a single reset, bleeding the radiators | Whole building: on-call person and heating contractor. One flat: first slot the next working day |
| Someone stuck in a lift | Emergency | Press the lift's alarm button, which connects to the lift company's 24-hour service | The on-call person, who confirms the lift company is on its way |
| Locked out, dripping tap, mold, broken intercom | Routine | An acknowledgment and the next step, plus a locksmith list for lockouts at the tenant's own cost | Nobody tonight. A work order for the morning |
From first message to closed work order
The model appears where language and images are involved. Everything that decides urgency, responsibility, money or who gets woken up is a rule your team can read.
01 Trigger · WhatsApp Cloud API, Graph, portal webhook
A tenant reports a problem
On WhatsApp, by email, through the tenant portal, or as a voicemail that is transcribed. The first reply says in the tenant's language that an automated assistant is answering and how to reach a person, as the EU AI Act's Article 50 requires.
02 Plain code
Identify the tenant and the unit
The phone number or email address is matched to a tenancy in the property management system, which gives the building, the unit and the language on file. An unknown number is asked for the address and flat number, and a person confirms the match before anything is dispatched.
03 AI model
Understand the problem and ask for what is missing
The model extracts what, where and since when, plus any safety signals, under a fixed schema. It asks for a photo and the missing details in one message rather than five, and describes the photos it receives for staff to confirm.
04 Decision
Emergency, urgent or routine?
Decided by the protocol table and a keyword safety net in code, never by the model's judgement alone.
- Gas, fire, active water, electrical danger, a building-wide heating failure then the approved safety text to the tenant and a phone call to the on-call person, at once
- Urgent but not dangerous then the first available slot, with staff told in the morning digest
- Routine then straight into the work order flow
- Not maintenance: rent, noise, a legal letter then handed to staff with a summary, and no reply on substance
05 Plain code
Check for duplicates and who is responsible
Open work orders in the same building with a similar problem in the last two days are linked instead of duplicated. Responsibility and the cost limit come from rules your staff configured per building and contract.
06 Person
Staff confirm what the rules cannot
A person confirms photos and responsibility before dispatch for anything unusual, above the cost limit, or with an unclear cause. In an emergency the on-call person decides, with everything the agent gathered in front of them. The approval patterns are in human in the loop AI.
07 System · PMS API, contractor portal or email
Create the work order and dispatch the trade
The work order goes into the property management system through its API where there is one. The contracted firm for that building and trade receives the address, access details, the tenant's phone number and the photos, and nothing else from the tenant's file.
08 Result
Follow up until it is closed
The agent chases the contractor for an appointment, tells the tenant when it is, asks the tenant afterwards whether the problem is fixed, and asks the contractor for the completion report. A work order closes only when both sides have answered.
A leak through a bathroom ceiling, Saturday 22:47
The tenant wrote in Bosnian; the exchange is shown here in English. The safety text is the approved one for active leaks, sent word for word.
WhatsApp, Saturday 22:47, January
Tenant, Stiege 2, Top 14 · WhatsApp, in Bosnian
Water is coming from the bathroom ceiling, a lot, near the light.
- identify_tenant(phone: "+43 660 ...")building B-031, Stiege 2, Top 14 / rental flat / language on file: Bosnian
- classify_request(message: "...")water, active, near electrics / keyword net agrees / emergency
- send_protocol(template: "leak_active_v3", lang: "bs")approved safety text sent
Agent · WhatsApp
This is the automated assistant of your property management, and a person is being called right now. If it is safe, close the main water valve in your flat, usually next to the water meter. Stay away from the bathroom light and any wet switches. Could you send a photo of the ceiling?
- page_on_call(building: "B-031", summary: "active leak, Top 14 bathroom ceiling")call placed 22:48 / answered 22:51 by Markus
Tenant, Stiege 2, Top 14 · WhatsApp, in Bosnian
(photo) It is still dripping.
- describe_photo(image: "wa_media_...")wet patch about 50 cm across above the shower, water dripping from the light fitting / for Markus to confirm
- find_open_orders(building: "B-031", category: "water", hours: 48)none open / flat above: Top 18, tenant on file
- Markus confirms, calls Top 18 and approves the emergency plumber.
- create_work_order(priority: "emergency", trade: "plumbing", vendor: "24h contract")WO-2026-0412 / plumber confirms arrival by 23:40
Agent · WhatsApp
A plumber is on the way and should be with you by 23:40. Markus from the office has called your upstairs neighbor. Please keep the bathroom light off until the plumber has checked it.
- Sunday 08:10: reports from Top 10 and the cleaning firm linked to WO-2026-0412.
Who is responsible, and who approves the cost
Who pays depends on the building type, the lease and the law: in Austria the MRG for many rental flats and the WEG for condominiums, in Germany the BGB, the WEG and the lease's small-repairs clause. Your staff write the rules per building. The agent applies them and shows which rule it used.
| Situation | Usually responsible | Who approves the cost | What the agent does |
|---|---|---|---|
| Common parts: roof, risers, facade, stairwell, lift | The building owner or the owners' association | Staff within the management contract's limit; above it, the owner or the association | Dispatches the contracted firm within the limit; above it, prepares the approval request with photos and a quote |
| A defect inside a flat the firm lets on an owner's behalf | The landlord, where the lease and the law put it on them | Staff, within the limit the owner has set | Dispatches within the limit and informs the owner |
| A privately let flat in a condominium the firm only administers | The flat's owner, unless the cause lies in the common parts | Not the firm's to approve | Tells the tenant to contact their landlord, and opens an inspection if a common riser could be the cause |
| Damage the tenant caused, or small items the lease assigns to them | The tenant | The tenant | Explains politely, citing the lease, and offers the contractor's contact |
| Source unknown, such as water from the flat above | Unclear until inspected | Staff | Orders the inspection first; staff allocate the cost after the report |
Cases that break a naive maintenance bot
One leak, three reports
The tenant with the wet ceiling, the tenant below with a damp wall and the cleaner who found water in the stairwell all report the same event. Reports are matched on building, location, problem type and time and linked to one work order. Each reporter hears that it is known and being handled, and the tenant upstairs is asked to check their flat.
A tenant who plays danger down
'Probably nothing, but the kitchen smells a bit of gas' reads as low urgency to a model that listens to tone. It is an emergency, which is why a keyword list in every supported language raises any mention of gas, smoke or sparks, however calmly it is phrased.
A photo that shows the wrong thing
A close-up of a stain says nothing about its cause. The model describes what is visible and never diagnoses; a person confirms the description before it justifies a dispatch, and the tenant is asked for a wider shot when the location is unclear.
The contractor who never confirms
Every dispatch has a clock. If the contracted firm has not confirmed an appointment in the agreed time, the agent tries the second firm on the list for routine jobs and alerts staff for anything urgent. Response times per contractor build up into a record the next contract negotiation can use.
Rent reduction, liability and angry letters
Tenants in Austria and Germany can reduce rent for serious defects, and a repair request can turn into that argument in one message. The agent does not discuss rent, liability or legal positions at all. Those messages go to staff with a summary, and the tenant is told a person will reply.
Photos of homes and vulnerable residents
Photos show the inside of people's homes, and some messages mention illness or disability, which is special category data under GDPR. Data stays in EU hosting under a processing agreement, contractors receive only what the job needs, and health details are kept as a priority flag rather than as text. More on that in GDPR-compliant AI development.
What tenants get from the agent, and what stays with staff
The AI model
Understand requests in eight languages
Tenants write the way they talk; the model turns that into what, where and since when.
Ask for the missing details and photos in one message
One clear question beats five rounds of back and forth.
Describe what a photo shows
Staff need not open every image, and a person confirms before the description counts.
Plain code
Decide emergency, urgent or routine
The protocol and the keyword net are rules, so urgency never depends on how calm a message sounds.
Send safety instructions
Only approved, translated texts go out, word for word.
Match duplicates and apply responsibility and cost rules
Written by staff per building and contract, and applied the same way every time.
A person
Emergency decisions at night
The on-call person decides what to dispatch and whom to call, with the facts already gathered.
Costs above the limit, liability, rent and disputes
Money and legal positions are never delegated to the agent.
EliseAI, Fixflo, Casavi, or a build for your buildings?
In US multifamily, buy. EliseAI raised $250 million at a $2.2 billion valuation in August 2025 to automate resident communication and maintenance for apartment operators, AppFolio has built AI into its own platform, and Property Meld specializes in coordinating maintenance with vendors. In the UK, Fixflo is the established repair-reporting tool. 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. I have not found a product built around those rules. It also wins when tenants will not install an app but already write on WhatsApp in six languages, and when the property management system is a legacy package without an open API, where the agent keeps its own work-order record and syncs by file.
The two combine well: keep Casavi or your property management system as the record, and put a WhatsApp agent in front of it that writes tickets into it. Built as an AI agent, the same intake can later answer the phone out of hours through an after-hours call agent.
How you would know it is working
A blueprint has no results to report, so here is what I would measure from the first week instead, on your own data.
- Time to on-call acknowledgment
- From the tenant's first emergency message to the on-call person answering the call, for every emergency, reviewed weekly.
- Complete at first staff touch
- The share of requests that reach staff with the unit, the problem, the location and a photo already there.
- Night calls that were not emergencies
- Calls the on-call person rated as unnecessary. The protocol is tuned until these are rare.
- Time to close, by category
- Days from report to confirmed completion, split by trade and by contractor.
- Tenant chasers
- Messages asking 'any news?' per work order. Good follow-up pushes this toward zero.
What a build like this costs
This is built as AI Agent Development, which runs $12K - $60K overall. A build like this one usually lands in the single-purpose agent tier: $12K - $30K, 2-5 weeks. The first working version runs on your real data well before the end of that window.
What it costs to run
Model costs are cents per request, a little more when photos are described. Tenants' WhatsApp messages and replies within 24 hours cost nothing; later updates use utility templates that Meta bills per message. On-call phone calls and SMS are billed per minute and per message by the telephony provider, usually a small monthly sum.
What moves the price
- Which channels: WhatsApp and email are the baseline, while voicemail transcription and a voice line add work
- The property management system: an open API, a partner-gated one, or none at all
- How many languages, plus the one-off translation and approval of the safety texts
- How many buildings, contracts and cost limits the responsibility rules must cover
- The on-call setup: rota, phone escalation and acknowledgment tracking
Who this is for
- Property managers in Austria and Germany running rental and condominium buildings with a small team
- Firms whose tenants already report repairs on WhatsApp, in several languages
- Managers on a property management system with no open API who still want intake automated
- Teams whose after-hours emergency line rings someone's personal mobile
Questions people ask about this
Can an AI agent handle tenant maintenance requests on WhatsApp?
Yes, for intake, triage, dispatch and follow-up. Meta's rules allow a business to use AI for its own customer service, and a property manager answering its tenants is exactly that. The agent must say it is automated, offer a clear route to a person, and use approved templates for updates sent more than 24 hours after the tenant's last message.
What happens if a tenant reports a gas smell or a burst pipe at night?
Within seconds the tenant gets the approved safety instructions for that situation in their language, and the on-call person gets a phone call, not a text. If nobody answers within a few minutes, the next person on the rota is called. The agent never tells a tenant to wait for the office instead of calling the emergency services.
Our tenants want a human, not AI. Will they accept this?
What tenants dislike about bots is being stuck with one. Here a person is always one message away, the first reply says so, and emergencies reach a human faster than before. What tenants notice is that someone asked for the photo straight away and told them when the plumber is coming. Tenants who prefer the phone keep the phone.
Does it work with Casavi, Aareon, Karthago or AppFolio?
It works with any property management system that can create or import work orders. With an open or partner API, work orders are written directly. Without one, the agent keeps its own work-order record and exchanges files with the system on a schedule. Which route applies to your system is one of the first things I check on a call.
Can the agent decide who pays for a repair?
It applies the rules your staff wrote per building and contract, and shows which rule it used. It never settles a disputed case: when the cause is unclear, the tenant contests a charge, or the lease and the law point different ways, a person decides. The model is asked what happened and where, never who is liable.
Sources