AI Emergency Triage for Property Management: Architecture and Lessons

Property management has an interesting AI challenge: you need to classify incoming tenant calls into urgency levels where false negatives have legal consequences.

A burst pipe logged as "routine maintenance" = flooded flat + legal liability. A noise complaint escalated as "emergency" = unnecessary after-hours contractor callout at 3x rates.

Here's what I've learned about building/evaluating these systems:

The Classification Problem

Tenant calls roughly break down into:

Category	% of Calls	Response SLA	Examples
Emergency	8-12%	Immediate (< 5 min)	Flood, gas leak, break-in, fire
Urgent	15-20%	Same day	Heating failure (winter), lock problem, no hot water
Routine	40-50%	2-5 business days	Dripping tap, minor repair, appliance issue
Informational	25-30%	Instant self-serve	Rent due date, bin collection, lease terms

The tricky part: tenants are unreliable narrators. "There's water everywhere!" might mean a catastrophic pipe burst or a small puddle from a leaky dishwasher seal.

The Architecture That Works

Tenant Call → STT → Intent + Urgency Extraction (LLM)
                          ↓
                   Keyword Safety Net (rule-based)
                          ↓
              Clarifying Questions (if ambiguous)
                          ↓
                   Classification Decision
                     /        |        \
              Emergency    Urgent     Routine/Info
                 ↓           ↓            ↓
           Immediate      Same-day     Log + 
           alert to PM    queue        auto-respond

Key design decisions:

1. Keyword Safety Net

Even if the LLM classifies something as routine, certain keywords always trigger emergency escalation:

• "fire", "smoke", "gas smell", "flooding", "break-in", "carbon monoxide"
• This is a hard override — no LLM interpretation allowed

2. Clarifying Questions

For ambiguous situations, the AI asks targeted questions:

• "Is the water actively flowing or is it a small drip?"
• "Can you see the source of the water?"
• "Is there any electrical equipment near the water?"

This takes 20 seconds and dramatically improves classification accuracy.

3. Conservative Bias

The system is intentionally biased toward over-escalation. In property management, the cost asymmetry is massive:

• False positive (unnecessary urgent alert): PM loses 2 minutes reading a text
• False negative (missed emergency): potential property damage, tenant harm, legal liability

Real-World Performance

From what I've seen testing VoiceFleet with property management clients:

• Emergency detection accuracy: ~97%
• False positive rate (over-escalation): ~8%
• Average time to PM notification for emergencies: 45 seconds
• Tenant satisfaction with AI handling: surprisingly high (tenants prefer instant answers over voicemail)

Integration Considerations

Property management software (Arthur, Propman, Reapit) have varying API quality. The main integration points:

1. Tenant identification — match incoming number to tenant record
2. Property lookup — which property, which unit, which contractor list
3. Maintenance logging — create work order with category, urgency, description
4. Contractor dispatch — for emergencies, auto-notify the relevant trade

The API maturity in proptech is about where dental PMS was 3 years ago — functional but clunky.

---

If you're building in this space, the key lesson is: get the triage right or don't ship it. A property management AI that misclassifies an emergency is worse than no AI at all.

Interested in hearing from anyone working on similar triage problems in other verticals (healthcare, facilities management, etc.).