An agent takes every maintenance text, call, and photo, sorts emergencies from routine work, collects what the vendor needs, and hands a complete work order to a human who approves the dispatch.
For a small or mid-sized property manager, maintenance is the job that never closes. Requests arrive by text, voicemail, email, and the portal, usually missing the one detail the vendor needs — which unit, which fixture, how long it has been leaking, whether anyone can let the tech in. Somebody on staff spends the day turning "the sink is broken" into a work order, deciding whether it can wait until Monday, and calling a plumber who then calls the tenant back to ask the same questions again. After hours it is worse: the on-call phone gets every request, and the manager has to judge from a two-line text whether a "leak" is a drip under a sink or water coming through a ceiling. Large multifamily operators buy this as a product — EliseAI, for example, says its assistant "interprets messages related to maintenance requests, captures essential details, and categorizes work orders," walks residents through troubleshooting for minor issues, and follows preset emergency routing after hours. A manager running a few hundred scattered units usually is not that buyer, and can build the same core loop: an agent on the tenant-facing phone and SMS line that classifies urgency first, asks the missing questions second, attaches the photos, and writes a work order a human approves with one tap. The agent never decides that an emergency can wait, never gives repair instructions for gas, electrical, or fire hazards, and never commits a vendor or a cost — those stay with people.
Sit down with whoever takes the on-call phone and write three lists. EMERGENCY, dispatch now: smell of gas; fire or smoke; carbon monoxide or smoke alarm sounding; sparking or arcing; burning or electrical smell; hot, buzzing, or scorched outlet, switch, or electrical panel; exposed wiring; water on or near outlets, switches, light fixtures, or the electrical panel; water coming through a ceiling or wall or covering a floor; sewage backing up; no heat when outdoor temperature is below [40F]; no power to the whole unit; exterior door or ground-floor window that will not lock; tenant says they feel unsafe. HAZARD PRECEDENCE: if water and any electrical item both apply, hazard_type is electrical and only the ELECTRICAL script is sent. URGENT, within 24 hours: no hot water, a single toilet not working when it is the only one, refrigerator out, a slow leak that is contained. ROUTINE: everything else. These are your rules, not the model's — the agent applies them, it does not invent them, so the EMERGENCY list here and the one in the classification prompt must be word-for-word identical. Check your lease and local housing code for any response timelines you are bound to and write them next to each list.
Give tenants one number for maintenance and route its texts and after-hours calls to the agent. A Twilio number gives you full control of the SMS flow in n8n; register it for A2P 10DLC before go-live, because Twilio requires registration for application-to-person SMS to US numbers over 10DLC and unregistered traffic incurs additional carrier fees. For calls, record the voicemail and transcribe it before classification, so the classifier always reads text. If you already run Quo as your business phone, its built-in AI agent, Sona, can answer after-hours calls, but every Quo plan includes 1,000 automation credits, which Quo equates to about 10 calls (as read on Quo's pricing page in early November 2026), so check that allowance against your call count before relying on it as the only path. Whatever the channel, the first message says it is an automated assistant and how to reach a person.
In n8n, send each new inbound message (and any photo) to the Claude API with the urgency rules and the classification schema from the assets. Classification happens first and alone, so an emergency is never delayed by a questionnaire. On EMERGENCY, the agent sends the safety script immediately — for gas, fire, or carbon monoxide that means leave the unit and call 911 or the gas utility, and for an electrical hazard it means stay away from it and do not touch it, and nothing else. When water and electricity both appear — water through a ceiling at a light fixture, a leak onto an outlet — the hazard is electrical and only the ELECTRICAL script is sent, never the FLOODING script that asks the tenant to reach a valve. On every EMERGENCY, the agent also pages the on-call person in Slack and by phone. When the model is unsure between two levels, the rule is to pick the higher one.
For URGENT and ROUTINE requests, the agent asks only the questions on the checklist for that category (plumbing, electrical, appliance, HVAC, pest, doors and locks, other): location in the unit, when it started, whether it is getting worse, a photo, permission to enter, pets, and best access windows. One question per message, no more than six. It may suggest one safe, non-technical check from your approved list — press the reset button on the underside of the garbage disposal, or check that the thermostat is set to heat — and nothing that involves tools, outlets, electrical panels, gas, or ladders.
The agent writes a structured work order to Airtable (or to your property management system if it accepts inbound work orders): unit, tenant, category, urgency, summary, answers, photos, access notes. A Slack message shows the work order with Approve, Edit, and Hold buttons and your preferred vendor for that category and property. A coordinator approves; the agent never assigns a vendor or a not-to-exceed amount on its own. Approval sends the vendor a complete work order by text or email. Many trade vendors schedule in a field-service tool such as Jobber or Housecall Pro, so send them something they can enter in one step rather than trying to integrate with their system.
When the vendor confirms a window, the agent texts the tenant the date, the vendor name, and the access arrangement. After the visit, it asks the tenant one question: is it fixed, yes or no. A no reopens the work order and alerts the coordinator. The vendor's invoice and completion note attach to the same record, which is the history you want when the same unit reports the same leak in three months.
Pull ten work orders a week and check three things against what actually happened: was the urgency right, were the questions enough for the vendor to come out once, and did any request wait longer than your rules allow. Any emergency classified lower than it should have been is a rule change the same day. Track repeat requests per unit and per category; that list is your capital-planning input, and it comes free once the work orders are structured.
Use these templates as-is or customize for your business.
You classify residential maintenance requests for [Company]. You do not give repair advice and you do not schedule anything.
URGENCY RULES (apply exactly; when unsure between two levels, choose the higher):
EMERGENCY: smell of gas; fire or smoke; carbon monoxide or smoke alarm sounding; sparking or arcing; burning or electrical smell; hot, buzzing, or scorched outlet, switch, or electrical panel; exposed wiring; water on or near outlets, switches, light fixtures, or the electrical panel; water coming through a ceiling or wall or covering a floor; sewage backing up; no heat when outdoor temperature is below [40F]; no power to the whole unit; exterior door or ground-floor window that will not lock; tenant says they feel unsafe.
HAZARD PRECEDENCE: if water and any electrical item both apply, hazard_type is electrical and only the ELECTRICAL script is sent.
URGENT: no hot water; the only toilet not working; refrigerator not cooling; contained leak; HVAC out outside the emergency conditions; pest sighting of rodents or bed bugs.
ROUTINE: everything else.
CATEGORIES: plumbing | electrical | appliance | hvac | doors_locks | pest | structural | exterior | other
Return JSON only:
{
"urgency": "EMERGENCY | URGENT | ROUTINE",
"category": "...",
"hazard_type": "gas | fire_smoke | co_alarm | electrical | flooding | sewage | no_heat | no_power | security | none",
"confidence": "high | medium | low",
"reason": "one sentence quoting the words that drove the decision",
"photo_observations": "what is visible in any photo, or none"
}
If confidence is low, set urgency one level higher than your best guess.FIRST REPLY (every new thread): "[Company] Maintenance: this is an automated assistant. I'll get the details to our team. Reply PERSON at any time to reach staff. If anyone is in danger, call 911 now." GAS / FIRE / CO (send immediately, then page on-call): "Please leave the unit now and keep everyone out. Call 911, and for a gas smell also call [Gas Utility] at [number] from outside. Do not use switches, flames, or your phone inside. Our on-call team is being alerted and will contact you." ELECTRICAL (sparking, burning smell, hot outlet or switch, exposed wiring, water near outlets or the panel; send immediately, then page on-call): "Please stay away from it and don't touch it, or anything plugged into it. If you see smoke or flames, or smell burning, leave the unit and call 911. Our on-call team is being alerted now and will contact you." FLOODING / SEWAGE: "Our on-call team is being alerted now. If you can do so safely and the water is nowhere near outlets, light fixtures, or the electrical panel, turn off the water valve under the fixture or the main shutoff at [location if known]. Do not touch electrical outlets or devices near water." ROUTINE CONFIRMATION: "Thanks — work order #[id] is created for [summary]. Our team will confirm a vendor and time within [1 business day]. Reply here if anything gets worse." SCHEDULED: "[Vendor] is scheduled for [date, window] for work order #[id]. Access: [arrangement]. Reply CHANGE if this doesn't work." COMPLETION CHECK (day after visit): "Is the issue from work order #[id] fixed? Reply YES or NO."
work_order_id | autonumber property / unit | linked records tenant | linked record (name, phone) received_at | timestamp of first message channel | sms | voice | email | portal urgency | EMERGENCY | URGENT | ROUTINE (agent-set; coordinator can change) urgency_changed_by | blank or coordinator name — every change is a training signal category | plumbing | electrical | appliance | hvac | doors_locks | pest | structural | exterior | other summary | two sentences, agent-written answers | location in unit, started, worsening y/n, entry permission, pets, access windows photos | attachments suggested_vendor | lookup from vendor table by category + property approved_by | REQUIRED before dispatch vendor_assigned | set on approval only not_to_exceed | set by coordinator only scheduled_window | from vendor confirmation completed_at | from vendor tenant_confirmed_fix | yes | no | no_reply invoice | attachment repeat_flag | auto: same unit + category within 90 days VIEWS: Open emergencies (should always be empty or owned) · Awaiting approval · Awaiting vendor · Repeat issues by unit
Get a new AI workflow every week. Prompts, tool stacks, and ROI math included.
AI does the categorization or first-draft work, a human approves before action is taken. The pattern of choice for anything irreversible, externally visible, or financially sensitive.
Learn the agentic glossary →Where this workflow tends to break in production — and what to put in place before you ship it.
An emergency is classified as routine and waits
Mitigation: Classification runs alone on the first message; low confidence rounds up; any coordinator downgrade or upgrade is logged and reviewed weekly; rule changes ship the same day.
Agent gives unsafe repair advice, or misses an electrical hazard
Mitigation: Safety scripts are fixed text, not model output; the prompt forbids repair advice; only checks from an approved list (disposal reset button, thermostat setting) may be suggested; electrical hazards get the fixed electrical script.
Emergency page goes unanswered after hours
Mitigation: Page by Slack and phone with escalation to a second person after 10 minutes; open emergencies view must be empty or owned at all times.
Vendor arrives without what they need and a second trip is booked
Mitigation: Category checklists built from vendor feedback; track second-trip rate by category and add the missing question.
Tenant cannot reach a human
Mitigation: PERSON keyword routes to staff in business hours and to on-call outside them; tested monthly from a tenant phone.
Skip the build if you run a large multifamily portfolio on a platform whose maintenance AI you can switch on — buy EliseAI or your platform's equivalent and spend the effort on vendor management. Do not use it if you cannot staff the human approval step and the emergency page every hour of every day; an agent that classifies an emergency correctly and then pages nobody is worse than a voicemail box, because tenants believe someone was told. Never let the agent troubleshoot anything involving gas, electrical panels, fire, or work at height, never let it promise a repair time your vendors have not confirmed, and never let it assign a vendor or approve a cost — the value is a complete, correctly prioritized work order, not an autonomous dispatcher.
A phased approach to get this workflow running and delivering ROI.
Days 1–30
Foundation
Days 31–60
Optimization
Days 61–90
Scale
Maintenance requests eat 40% of a property manager's time. Here is how AI is handling triage, scheduling, and tenant communication automatically.
Only Fin publishes a per-unit price ($0.99 per outcome). Decagon, Sierra and Parloa are quote-only, so this comparison focuses on what each one bills for and the break-even rates to bring to the negotiation.
AutoGen is archived and Semantic Kernel is now Microsoft Agent Framework. A comparison of the open agent frameworks that are left, plus a worked hosting-cost example from each vendor's own price list.
One practical AI workflow per week. No fluff.
Get the full guide with step-by-step setup, workflow templates, and copy-paste assets.