Outbound voice AI has a problem that doesn't get talked about enough: answering machines. In India, voicemail penetration is lower than in Western marketsbut DND voicemails, operator-side IVR intercepts, and Truecaller-blocked calls result in a meaningful percentage of outbound calls that reach something other than a live human.
Without proper Answering Machine Detection (AMD), your AI agent launches into its full pitch to a voicemail promptburning per-minute cost, producing no outcome, and polluting your analytics with false "answered" records.
What AMD Is and Why It's Hard
Answering Machine Detection is the process of determining, within the first 1–3 seconds of a call being answered, whether the call was picked up by a human or a machine. The challenge is that this determination must be made in real timeyou can't wait for the voicemail greeting to finish before acting.
The acoustic signatures of human and machine pickup differ, but subtly:
- Humans typically say something within 0.5–1 second of pickup"hello," a greeting, background noise
- Machines have a characteristic silence gap (0.8–1.5 seconds) before the recorded message begins
- Operator intercepts ("The number you have dialed is currently busy") have a distinct synthesized voice signature
In India, the challenge is compounded by background noise (vehicles, market sounds, construction) that can make human pickup sound machine-like, and by the wide variety of operator announcement voices across Airtel, Jio, Vi, and BSNL.
Agni AMD accuracy: 94.2% accurate across Indian operator environments in production testing. False positive rate (human incorrectly classified as machine) is below 1.8%the threshold at which false positives become a meaningful campaign problem.
How Agni's AMD Works
Agni's AMD system runs a three-stage classification in the first 2.5 seconds of call pickup:
Stage 1: Energy and Silence Analysis (0–500ms)
The first 500ms of audio after pickup is analyzed for energy profile. A characteristic machine silence or operator intercept pattern triggers immediate classification. Human pickup with voice energy in this window moves to Stage 2.
Stage 2: Phoneme Pattern Classification (500ms–1.5s)
The first 1 second of speech is classified against acoustic models of human speech vs. synthesized/recorded voice. Indian-language operator announcements have been specifically trained into the model"Aapka call abhi..." is immediately classified as an operator intercept.
Stage 3: Semantic Confirmation (1.5s–2.5s)
For ambiguous cases, partial transcription is used to confirm classification. If the first words are a greeting ("hello," "haan," "ji"), human classification is confirmed. If they match voicemail prompt patterns ("Please leave your message"), machine classification is confirmed.
What Happens After AMD Classification
Once AMD classifies the call:
- Human detected: The conversation begins immediatelyno gap, no re-introduction
- Machine detected: Three configurable options: (a) hang up silently, (b) leave a pre-recorded voicemail drop, (c) log the attempt and schedule a retry at a different time
- Operator intercept detected: Immediately logged as non-contactable; call not counted against retry limits
The Campaign Economics
At a 15–20% machine/intercept rate on typical Indian outbound campaigns, AMD saves significant per-minute cost. For a campaign making 10,000 calls per day at an average call duration of 2.5 minutes:
- Without AMD: 1,750 machine calls × 2.5 min × ₹8/min = ₹35,000 wasted per day
- With AMD: Machine calls terminated at 3 seconds = ₹1,400 total AMD cost
- Daily savings: ₹33,600. Monthly: ₹10.1 lakh.
Beyond direct cost, AMD also cleans your analytics: completion rate, sentiment data, and outcome tracking are based only on actual human conversationsnot machine pickups that skewed everything.
"We thought our completion rate was 62%. After enabling AMD, we discovered our real human completion rate was 78%the difference was machine pickups being counted as incomplete human calls."Ops Head, Telesales Company (Delhi)
Voicemail Drops: The Indian Context
In markets like the US, voicemail drops (pre-recorded messages left on answering machines) are a significant marketing channel. In India, voicemail usage is much lowermost consumers don't listen to voicemails. Agni's default AMD action for Indian campaigns is therefore a silent hang-up + retry scheduling, not a voicemail drop. This can be configured per campaign if voicemail drops are relevant to your use case.
Ready to get started?
AMD is enabled by default on all Agni outbound campaigns. Start your campaign at app.ravan.ai or contact us at info@ravan.ai.
