BUILT-IN - NO SETUP REQUIRED

The first AI voice platform that knows when it’s talking to another AI.

Every other outbound AI agent bleeds margin on AI-to-AI loops. We detect within 10 seconds and convert the exit into a warm SMS lead.

Margin Defender call-flow: AI-to-AI detection and relayOUTBOUND AIMARGIN DEFENDERAI RECEPTIONISTDETECTEDrelay message → SMSOWNER CALLBACKAMDBehavioralLLM ClassifierLoop Detector3 of 4 = TERMINATE
THE PROBLEM IN NUMBERS

Your outbound AI is losing to the inbound AI. Both think they’re winning.

US small businesses are deploying AI receptionists faster than outbound platforms can detect them. Without controls, your agent runs a 12-minute conversation with their agent and books nothing.

  • 1 in 3

    US small businesses now answer with an AI receptionist

    Source: NextPhone - 347K-call dataset
  • ~12 min

    average AI-vs-AI call duration without any safeguards

    Source: SmartMoves internal model - plan §09
  • $50-$80

    of gross margin wasted per active tenant per day, unaddressed

    Source: SmartMoves cost model - plan §09
  • ~98%

    of wasted call COGS recovered with Phase A + Phase B controls

    Source: SmartMoves cost model - plan §09
HOW WE DEFEND IT

Six layered controls. Zero new ML required for the first five.

Phase A controls (caps, watchdog, loop detector, cost ceiling) ship immediately. Phase B detection (4-signal ensemble + relay message) is the differentiator.

Hard duration cap

8-minute ceiling on every call. Public cold-calling benchmarks show 99% of productive calls finish in under 8 minutes. Anything longer is waste - we end it.

Source: public cold-calling benchmarks (2025)

Dead-air watchdog

10 seconds of silence after agent speech triggers auto-hangup. No more sitting in voicemail trees or IVR hold loops burning per-minute COGS.

Loop detector

4-gram Jaccard overlap on consecutive agent turns. If your agent starts repeating itself - classic AI-vs-AI symptom - we stop the call. ~1ms CPU, no model required.

Per-call cost ceiling

$0.60 COGS hard cap per call (10 minutes at $0.06/min blended). A soft-warn prompt injects at 80% - the call closes gracefully before margin evaporates.

AI-receptionist detection

4-signal ensemble: answering-machine detection + behavioral heuristics (first-word latency, filler rate, turn-gap variance) + side-channel classifier + loop detector. 3-of-4 positive votes triggers a structured exit.

Structured relay message

On detection, the agent speaks a clean 5-8 second message designed for AI receptionists to relay via SMS to the business owner. Every wasted call becomes a warm inbound lead.

DETECTION IN ACTION

Call · vapi_call_abc123 · 14:32:08
Ended
AMD Resultai-receptionist
Terminated byai_receptionist
Agent Turns
4/ 20 max
Est. COGS
$0.04$0.72
saved $0.68
Ensemble Votes
AMDBehavioralLLM ClassifierLoop3/4 = TERMINATE
Action taken: Speak relay message → SMS pending owner callback
THE MATH

What one tenant recovers in a month.

Single tenant, 8 hours/day, 5 concurrent calls, 30% AI-receptionist hit rate - conservative for US HVAC and dental campaigns (research: 30-50%).

Metric
Without controls
With SmartMoves
Wasted COGS / day (AI-vs-AI loops)
$259
~$5
COGS recovered / day vs. uncontrolled
-
$254 (98%)
Gross-margin recovery / month (22 days)
-
$5,588

Estimates based on published industry-survey datasets (347K-call cold-calling benchmark, 2025 cold-calling statistics) and our internal cost model. Source: AI-to-AI Handling Plan §09. Your numbers vary. Real per-tenant COGS and recovery are surfaced in the dashboard for every active campaign.

WHY IT’S HARD

No vendor has a one-shot signal.

The ensemble is the honest answer.

Standard answering-machine detection (3-4 second window, ~94% accuracy on traditional voicemail) classifies modern AI receptionists as "human" - they open with "Thanks for calling Acme HVAC, this is Sarah, how can I help you today?" That phrase is phonetically and grammatically indistinguishable from a real receptionist. No single signal is reliable enough. We use 4 signals because that is the engineering answer, not a product trick.

Detection runs fire-and-forget.

All safeguard checks are async - they never sit in the audio path. The same architectural pattern powers our escalation detection, which is in production today. Latency on healthy human calls is unchanged.

The relay message is not optional - it's reputation insurance.

Silent hangups on detected AI receptionists look identical to spam-scoring bots to carrier-reputation systems. Our 5-8 second scripted close before exit prevents your phone-number DIDs from being flagged, which would erode call-completion rates 15-25% on the trunk.

Signal Stack — Ensemble DefenseLatency
L1 Hard Caps
8-min duration · 10-sec dead-air · $0.60 COGS ceiling
~0 ms
L2 Carrier AMD
~840 ms · 4-class output (human / machine / no-answer / unknown)
~840 ms
L3 Behavioral Heuristics
First-word latency · filler rate · turn-gap variance
~50 ms
L4 LLM Side-Channel Classifier
Small fast LLM · runs on transcript, not audio
~1500 ms
L5 Loop Detector
4-gram Jaccard overlap on consecutive agent turns · ~1 ms CPU
~1 ms
3-of-4 positive votes required to trigger exit — keeps false-positive rate on human receptionists very low.

No standards path in 2026. The IETF draft for AI-preference signaling is about policy, not identity - there is no in-band “I am an AI” signal on SIP today. Heuristic ensemble detection is the engineering answer for the foreseeable future.

QUESTIONS

Frequently asked about Margin Defender

Does this work with my dialer?
Yes. Margin Defender is drop-in middleware - no changes to your dialer. Any SIP-based dialer that can forward to our number works out of the box.
What if I want to NOT hang up on AI receptionists?
Fully configurable, per tenant and per campaign. Some verticals - medical answering services, legal intake - need the AI receptionist to relay the message rather than triggering a hangup. The behavior is a flag in your tenant settings, not hardcoded.
How accurate is the AI-receptionist detection?
The ensemble requires 3-of-4 signal positives before triggering an exit, which keeps the false-positive rate very low on real human receptionists. Behavioral signals alone (first-word latency, filler rate, turn-gap variance) are accurate on approximately 85% of the AI-receptionist call class.
Is the side-channel classifier required?
No. The system runs and defends margin without it - Phase A hard caps activate immediately with zero extra inference cost. The side-channel classifier is recommended for ambiguous edge cases (HVAC dispatchers who pick up in under 300ms can read as AI to the behavioral layer).
Does this affect human-call latency?
No. All safeguard checks run fire-and-forget - they never block the audio path. The same architectural pattern is used in our escalation detection, which is in production today.

Stop bleeding margin to AI receptionists.

Try free for 100 minutes. Margin Defender is on by default - no configuration required.

Start freeTalk to sales