After-Hours Emergency Detection: Vapi Can't Match CallSphere 7-Agent Stack

TL;DR

CallSphere's After-Hours Escalation product (escalation.callsphere.tech) monitors both email (IMAP) and Dialpad voice/voicemail between 12am-7am EST and runs 7 specialized agents (EmailTriageAgent, DialpadAgent, VoicemailAnalyzerAgent, VoiceAgent, SmsAgent, AckMonitorAgent, HeadAgent) to detect and escalate emergencies. Vapi.ai is voice-only — there is no email channel, no IMAP listener, no multi-channel orchestrator. To match CallSphere on Vapi you would build the entire email pipeline, voicemail analyzer, ACK monitor, and head orchestrator yourself. This post breaks down the emergency detection cascade with a Mermaid diagram and a worked example of a 2:18am water leak.

The After-Hours Property Reality

Property management firms — multi-family residential, HOAs, commercial portfolios — face a pernicious problem: the worst emergencies happen between midnight and 7am. The National Apartment Association 2025 emergency benchmark report shows:

• 47% of after-hours calls are non-emergencies (lockouts, noise complaints, billing).
• 22% are urgent but non-life-safety (cooling failure in summer, no hot water, minor leaks).
• 11% are genuine life-safety (fire, gas leak, flooding, security breach).
• 20% are spam or wrong numbers.

The cost structure is brutal. A 24/7 dispatcher service costs $3,800-5,200 per month per portfolio. Missing one of the 11% genuine emergencies costs the property an average of $47,000 in damage and liability per incident, per Marsh McLennan property risk benchmarks. Acting on a non-emergency costs $190 in unnecessary plumber/electrician dispatch.

The job is detection — separating the signal from the noise — and then escalation — paging the right person at the right urgency level until someone acknowledges. Both require multi-channel listening (email and phone, because residents use both) and a structured escalation ladder.

Why Vapi Cannot Match This

Vapi is a voice platform. By design, by branding, by architecture. It has:

• No email ingestion. Period. No IMAP, no API, no inbox.
• No multi-channel orchestrator. You can call Vapi from voice; you cannot fan in voice + email + SMS into one decision.
• No emergency-keyword classifier. You write the prompt.
• No escalation primitives. You write the ladder, the timer, the ACK detector.
• No after-hours window logic. You write the time-zone math.

You can absolutely build this on Vapi. Many people have. They build a Python service that polls IMAP, calls Vapi for outbound voice, sends SMS via Twilio independently, runs a Postgres state machine, and stitches it together. It works. It is also six months of engineering and an ongoing maintenance burden that breaks every time Vapi or Twilio updates an upstream API.

CallSphere's after-hours product is the answer pre-assembled.

The 7-Agent Stack

Agent	Role	Channel
EmailTriageAgent	Score incoming emails 0-1 for emergency	IMAP poll
DialpadAgent	Listen to inbound Dialpad calls + voicemail	Dialpad webhook
VoicemailAnalyzerAgent	Transcribe + score voicemails	Whisper + scorer
VoiceAgent	Outbound voice with TTS scripts	Voice
SmsAgent	Outbound SMS with templated alerts	Twilio SMS
AckMonitorAgent	Detect acknowledgments, stop ladder	Multi-channel
HeadAgent	Orchestrator, owns event lifecycle	All

Plus the data model: users, on_call_rotations, escalation_contacts, events, escalation_logs, acknowledgments, admin_alerts, emergency_keywords, daily_metrics, escalation_ladder_config.

The platform runs on Python FastAPI with the OpenAI Agents SDK using gpt-5.2.

Comparison

Capability	After-Hours Escalation	Vapi
Email IMAP listening	Built-in	None
Dialpad voicemail capture	Built-in	None (Twilio only)
Multi-channel correlation	HeadAgent orchestrator	Build yourself
Emergency keyword scorer	emergency_keywords table	Build it
Escalation ladder	escalation_ladder_config	Build it
ACK monitoring stops cascade	AckMonitorAgent	Build it
12am-7am after-hours window	Configurable per portfolio	Write time math
Specialist agents	7	0
Time to deployment	Days	Months

The Emergency Detection Cascade

```mermaid graph TD A[12am-7am EST Window Active] --> B[Multi-Channel Listening] B --> C[EmailTriageAgent on IMAP] B --> D[DialpadAgent on Voice] C --> E[Email Arrives] D --> F[Call Connects or Voicemail] E --> G[EmailTriageAgent Scores 0-1] F --> H{Voice or Voicemail?} H -->|Voice| I[Live Conversation: VoiceAgent] H -->|Voicemail| J[VoicemailAnalyzerAgent: Transcribe + Score] I --> K[Capture Symptoms + Score] J --> K G --> L{Score >= 0.6?} K --> L L -->|No| M[Log Event Status: Non-Emergency] L -->|Yes| N[HeadAgent: Create Event] N --> O[Build Escalation Ladder] O --> P[Primary Contact + Secondary + 6 Fallbacks] P --> Q[VoiceAgent + SmsAgent: Simultaneous] Q --> R[Wait 120s] R --> S{ACK Received?} S -->|Yes| T[AckMonitorAgent: Stop Ladder + Log] S -->|No| U[Advance to Next Contact] U --> V{Last Fallback?} V -->|No| Q V -->|Yes| W[admin_alerts: Page Property Manager] T --> X[Update daily_metrics] M --> X W --> X ```

Every transition writes to events and escalation_logs. The dashboard surfaces every event in real time, with the full audit trail of who was contacted, on what channel, with what response.

Emergency Keyword Scoring

The EmailTriageAgent and VoicemailAnalyzerAgent both consult the emergency_keywords table. Sample rows:

Keyword	Weight	Category
flooding	0.85	water
water everywhere	0.85	water
fire	0.95	fire
smoke alarm	0.65	fire
no heat	0.55	hvac (winter only)
gas smell	0.95	gas
break-in	0.90	security
medical	0.80	medical
dead body	1.00	medical
locked out	0.15	non-emergency

The agent computes a score by summing matched-keyword weights, capped at 1.0, and then applies LLM judgment to adjust for context (e.g., "smoke alarm beeping every 30 seconds" scores higher than "smoke alarm needs new battery").

Score ≥ 0.6 → Event → escalation cascade.

Worked Example: 2:18am Water Leak

A tenant emails the after-hours inbox at 2:18am: "Water gushing out from under my kitchen sink, soaking into the floor — what do I do?"

Step 1 (EmailTriageAgent): Keywords matched: "water gushing" (0.78), "soaking" (0.45). Capped at 0.9. Above 0.6 threshold.

Step 2 (HeadAgent): Creates Event with severity "high", category "water". Looks up the property, builds the ladder from on_call_rotations: Primary maintenance (Jose), Secondary (Maria), Fallback 1 (regional manager Tom), Fallback 2 (emergency plumbing service), and 4 more.

Step 3 (VoiceAgent + SmsAgent): Simultaneously calls Jose's mobile and texts him: "URGENT: Active water leak at Unit 412, 1234 Main St. Tenant needs immediate response. ACK with REPLY YES or callback."

Step 4 (Wait 120s): Jose does not pick up.

Step 5 (Advance): Maria. Same simultaneous voice + SMS. Maria texts back "YES". AckMonitorAgent detects, stops the ladder, logs ACK.

Step 6 (HeadAgent): Replies to the original tenant email: "Maria from maintenance is on her way, ETA 25 minutes. If you can safely turn off the water shut-off under the sink, please do." Logs the closure.

Total time from email arrival to ACK: 4 minutes 41 seconds. The water damage is contained to the kitchen rather than spreading to the apartment below — a savings of $8,000-30,000 in remediation cost.

Why Email Matters as a Channel

Property management has migrated half of tenant communication to email and portal forms. Tenants under 35 default to typing, not calling. The 2025 NAA Resident Communication Survey found that 53% of after-hours emergency reports from millennial and Gen-Z residents arrive by email or portal form, not phone.

Vapi-based systems miss every one of these. The email arrives, sits in the inbox, gets read in the morning. The water has been gushing for six hours. The damage is total.

CallSphere's IMAP listener polls inboxes every 30 seconds during the after-hours window (every 5 minutes during business hours, since the office handles those). Multiple inboxes per portfolio are supported — most properties run a maintenance@, an emergency@, and per-property aliases.

Categorizing What Comes In

The EmailTriageAgent's classifier is multi-dimensional:

• Severity: 0.0-1.0 score from keyword-weighted summing + LLM judgment.
• Category: water, fire, gas, medical, security, hvac, lockout, noise, billing, other.
• Urgency window: now / today / tomorrow / next-business-day.
• Property + unit identification: parsed from email body and signature.
• Tenant verification: cross-referenced with property's resident roster.
• Spam/wrong-number detection: low-confidence emails get parked, not escalated.

The output is a structured Event row that downstream agents consume. The structured output is what makes the system reliable. Vapi's PDF-based knowledge approach has nowhere to put structured outputs; you build the schema.

Voicemail Analysis: The Subtle Detail

Many emergency calls become voicemails because the live VoiceAgent line is occupied or the tenant hangs up before the agent answers. The VoicemailAnalyzerAgent is purpose-built for these:

1. Whisper transcribes (multi-language).
2. Keyword scoring runs.
3. LLM judgment refines.
4. Tenant identification parses the caller-ID and references the resident roster.
5. Reply path: the system can call back the tenant from the agent line, or text them.

Voicemails that score above 0.6 trigger the same escalation cascade as live calls. Below 0.4 they queue for the morning office. Between 0.4 and 0.6 they get a courtesy callback at the start of business hours.

We have measured cases where the live line missed an emergency call but the voicemail analyzer caught the gas leak. That single recovered case saved a client roughly $340,000 in averted property damage and liability.

Spam, Pranks, and Adversarial Inputs

After-hours emergency systems are juicy targets for pranksters and bad actors. We have seen:

• Repeated false fire alarms from a hostile former tenant.
• AI-generated voicemails attempting to trigger expensive plumbing dispatch.
• Spam emails with "fire" in the subject to bait the keyword scorer.

CallSphere's defenses:

• Caller verification: cross-reference incoming caller-ID with resident roster. Unknown numbers get a verification challenge ("can you confirm your unit number?") before triggering escalation.
• Rate limiting: a single phone number cannot trigger more than 3 events per 24 hours without manual review.
• LLM cross-check: the keyword score is always rechecked by an LLM that looks for "this is a real emergency vs this is suspicious."
• Audit trail: every event has a full classification rationale logged for forensics.

Vapi has none of this. Spam and false-alarm detection is your problem.

FAQ

Can it monitor more than one inbox?

Yes — multiple IMAP accounts per portfolio. Common pattern: a maintenance@ address per property, all listened to.

What if the tenant calls instead of emails?

DialpadAgent picks up the call. If they reach a person, the VoiceAgent runs a structured triage. If they leave voicemail, VoicemailAnalyzerAgent transcribes (Whisper) and scores. Same threshold.

What about false positives?

Emergency_keywords are tuned per portfolio. If "no heat" generates too many false positives during winter, the threshold for that keyword can be raised, or the LLM judgment layer can downweight it without an action sentence ("no heat for past 3 days" vs "no heat right now and there is a baby").

Can the on-call rotation handle PTO?

Yes — on_call_rotations supports overrides, swaps, and PTO blocks. The HeadAgent reads the rotation as of the event timestamp.

What if no one in the ladder responds?

After all 8 ladder positions exhaust, admin_alerts fires, paging the property manager directly with the full escalation log. This is rare (under 0.7% of events in our data) but always logged.

Can we report on resolved emergencies?

daily_metrics rolls up: events created, severity distribution, time-to-ACK, time-to-resolution, by-category breakdowns. Property managers get a weekly digest by email.

What about insurance reporting?

The escalation logs are formatted for insurance claim documentation. When an incident becomes a claim, the property manager exports the event timeline as a PDF with timestamps, channel, and contact records. Several insurers have begun discounting premiums for properties using documented emergency-response systems.

Does this work in non-US markets?

Yes — the platform supports international IMAP, international Twilio numbers, and per-region ladder configs. UK, Canada, Australia, and Singapore are deployed. Local emergency-line integration (e.g., 999 in UK) is configurable.

Stop Missing the 11%

If your after-hours system is one tired dispatcher and a phone tree, you are missing the calls that matter. Book a demo at /demo — we will simulate the cascade on your real on-call rotation and emergency keywords.