DEV Community

Cover image for Automatic Calls: How Automated Calling and AMD Work
Piyush Sahoo
Piyush Sahoo

Posted on

Automatic Calls: How Automated Calling and AMD Work

An automatic call is a phone call that software places without a person dialling it. The system takes a number from a list and rings it. When someone answers, it plays a message, starts an AI agent, or connects a human. Answering machine detection (AMD) can choose that path.

This guide follows one call through the system, sizes a campaign with real arithmetic, and helps you decide whether AMD belongs in your flow at all.

Note: Key takeaways

  • Two limits govern outbound. Calls per second (CPS) caps how fast calls start. Concurrency caps how many run at once, and ringing calls count.
  • AMD saves agent time, not channels. In our worked example, it frees 20 agent hours but only two channels.
  • AMD is a trade, not a default. When machines are rare, more of its machine verdicts land on real people, so measure your list first.
  • Hold an AI agent silent during detection. An agent that greets on answer puts its own voice into the sample.
  • Rules can reach the method as well as the message. Some regulators now look at how software places a call, not only at what it says, so check the rules in each country you call.

What is an automatic call?

An automatic call is any call where software, not a thumb on a keypad, starts the dial. What happens after pickup varies, and that difference decides how much an answering machine costs you.

Form Who the callee hears Common use If a machine answers
One-way message A recording or text-to-speech Delivery slots, outage notices The message may still land
Keypad flow A recording that takes input Confirm or reschedule an appointment Nothing gets pressed, so the task stays open
AI voice agent A speech model in real time Reminders, qualification, surveys You pay for speech and model time on a recording
Software dials, human talks A live agent Sales, collections A paid person listens to a greeting

The word robocall usually means the first form, sent in bulk. AI doesn't escape that label in the US, where the FCC confirmed in 2024 that AI-generated voices count as "artificial" under the TCPA. If you're in India, TRAI's A2P definition also reaches calls that end with a live agent.

How does an automatic call work?

One attempt passes through seven stations. Each station has its own failure, so it helps to know where a problem sits before you tune anything.

Flow of one automatic call: list and consent, pacing with the CPS limit, dial, ring. Unanswered calls go straight to the hangup webhook. Answered calls fetch instructions, may pass through detection, then follow a person path or a machine path to the hangup webhook. A concurrency channel is held from dial to hangup.

  1. List and consent. You decide who may be called, for what purpose, and from which number series.
  2. Pace. A worker releases attempts at a rate your account and your numbers can sustain.
  3. Dial. Your system asks the platform or trunk to start the call.
  4. Ring. The network reports ringing, busy, no answer, or failure. You don't need AMD to learn any of this.
  5. Answer. Your server gets a request and returns instructions. On Vobiz, those instructions are VobizXML.
  6. Detect. If enabled, AMD classifies the answer from its opening audio.
  7. Record. A webhook (callback) reports the outcome, and your system closes the contact or schedules a retry.

Station 7 is the one teams underbuild. Without a reliable outcome per attempt, you can't tell a list problem from a number problem, and retries turn into repeat dialling.

How many calls can you run at once?

Concurrency equals attempts per second multiplied by the average seconds each attempt holds a channel. Unanswered calls belong in that sum, because they occupy a channel while they wait to be answered.

Here's a campaign with illustrative inputs:

  • Volume: 30,000 attempts across a 6-hour window, or 21,600 seconds. That's about 1.4 attempts per second.
  • Answered: 30%, or 9,000 calls. Answer rates vary widely by list quality; see the number health guide.
  • Unanswered: 21,000 calls ringing for 25 seconds each. The same guide suggests a ring timeout of 25 to 30 seconds.
  • Who answered: 7,200 people talking for 120 seconds, and 1,800 machines. We assume a 20% machine share here purely for the sum.
  • Machines without AMD: each holds the line for 40 seconds before the agent gives up.
  • With gated AMD: every answered call spends 4 seconds in detection, and machines end there.

Stacked bars of channel seconds for 30,000 attempts. No detection uses 1,461,000 seconds, or 68 channels. Gated detection uses 1,425,000 seconds, or 66 channels. Ringing and talk time dominate both. Agent time on machines falls from 20 hours to close to none.

Divide channel seconds by the 21,600-second window and you get 68 channels without AMD and 66 with it. The two-channel gap surprises most people. Machines end sooner, but every person now waits 4 extra seconds, and that nearly cancels the saving.

So don't buy AMD to shrink a channel bill. Its value sits one layer up: 1,800 machines × 40 seconds is 20 hours of agent or AI time, gone almost entirely.

Steady-state math also flatters you. Each attempt here holds a channel for about 47.5 seconds, so a burst of 2 attempts per second needs about 95 channels while it lasts. Platforms cap both concurrency and calls per second, so pace evenly and leave headroom. To try your own figures, use the CPS and concurrency calculator.

What does answering machine detection listen for?

AMD classifies the first seconds of an answered call. It typically relies on two inputs: tones such as beeps and fax signals, plus the timing and pattern of speech. A person tends to say something short and wait. A recording keeps talking.

That one behavioural gap does most of the work, and you can draw it:

Bars on a zero to eight second axis. A person's greeting is a short hello, a business greeting adds a name and company and runs longer, and a machine greeting often runs 5 to 15 seconds. Markers show typical greeting allowances and analysis windows used by detectors.

Voicemail greetings commonly run 5 to 15 seconds. A person's hello is usually far shorter, and a business greeting with a name and company falls in between.

The chart also shows why tuning matters. A business receptionist's greeting can run past a 1.5-second allowance, so calling offices with default settings invites false machine results. Raise the greeting allowance for that audience.

Treat published accuracy claims with care. A figure measured in one country may not carry over to another, because tones and greetings differ. We found no independent benchmark that holds across countries, so your own recordings are the only figure worth trusting.

Should you turn AMD on?

Decide this per campaign. AMD helps when an expensive listener would otherwise sit on a recording, and it hurts when it hangs up on people you wanted to reach.

The deciding number is your machine share, because it changes what a machine verdict means. Take one detector and three lists:

Illustrative chart for 1,000 answered calls with a detector that catches 85 percent of machines and wrongly flags 3 percent of people. With 5 percent machines it catches 43 and hangs up on 29 people. With 20 percent it catches 170 and hangs up on 24. With 50 percent it catches 425 and hangs up on 15.

The detector didn't change, but at a 5% machine share about four in ten machine verdicts hit a real person. This is the classic base-rate effect, and it's the reason a setting that works in one market can quietly lose customers in another. The rates in the chart are assumptions for the arithmetic, not measurements of any product.

How to measure your own list

  1. Run detection without acting on it. Log every verdict, and let every call continue.
  2. Listen to a sample. Label each recording as person, machine, or unclear. Stratify by carrier and hour, because one configuration may not suit every case.
  3. Count both errors. A false machine means a person lost. A false human means an agent talks to a recording.
  4. Price them. Compare agent seconds saved against the value of the people you'd lose.
  5. Switch on automatic hangup only if the sum works. Re-test after any change of list, audience, or region.

The two errors aren't equal. A false human wastes seconds. A false machine ends a call with someone who picked up, so weight your thresholds against it.

What should the callee hear during detection?

The analysis window is the same for everyone who answers, so the real design question is what a person experiences while it runs.

Approach What the person hears Best fit Main risk
No detection The agent or message at once One-way messages, warm lists An agent or AI can talk to a recording
Blocking detection Silence until detection finishes, on platforms that hold the call Machine-heavy lists with a wait you can accept People hang up during the gap
Asynchronous, gated agent A held line, then the agent Live AI agents A few seconds' wait before the first word
Asynchronous, no gate The agent at once Rarely a good choice The agent's voice corrupts the sample

Gating an AI voice agent

Voice AI makes timing harder, because the agent's first line and the detector's sample collide. Vobiz's tuning guide lays out why: a greeting spoken on answer talks over the callee, and the silence-then-speech pattern never forms.

The fix has four parts:

  1. Start detection in the background. Request an asynchronous result when you place the call.
  2. Hold the line without words. Return call instructions that speak nothing the classifier could pick up. Hold music or a wait sized to your detection window works.
  3. Branch on the result. A person gets the agent. A machine gets your recorded message or a hangup.
  4. Set a deadline. If no result arrives shortly after the window closes, treat the call as human. Leaving a live person on hold is the costlier mistake.

Who should own the decision?

Some voice AI platforms detect at the agent layer. Retell runs voicemail detection for the first 3 minutes of a call and can hang up or leave a message. Vapi combines model-based detection with optional beep detection.

These listen to the conversation, while telephony-layer AMD decides from the opening seconds. In our view, you should pick one owner for the hangup decision. Two detectors acting at once make failures hard to diagnose, and neither one's error rate tells you the combined result.

If you're in India, TRAI's Third Amendment requires advance declaration of A2P calls and their caller ID ranges, doesn't exempt calls that include a live agent, and expects promotional calls on the 140xx series. The India calling regulations summarise what applies.

Does AMD create a problem in the US?

It can, for telemarketing that connects to a live agent. Under 47 CFR 64.1200, a call is abandoned if it doesn't reach a live sales representative within two seconds of the person's completed greeting. Abandoned calls can't exceed 3% of live-answered calls, measured per campaign over 30 days. The rule also bars disconnecting an unanswered telemarketing call before 15 seconds or four rings.

A detection window that runs past the callee's greeting eats into those two seconds. If you dial US consumers, have counsel review your timing before you enable AMD.

How do you keep numbers answerable?

AMD only matters on calls people pick up, and answer rates depend on how your numbers behave. Vobiz's guides go deep on this, and three habits matter most:

Vobiz docs for this topic

The setup details for this are in the Vobiz docs:

Anything that places calls or sends messages is billed and reaches real people, so run it only with your own credentials after confirming the details.

Frequently asked questions

What is an automatic call?

An automatic call is one that software starts without a person dialling it. On answer, the system might play a recording, collect a keypad choice, start an AI agent, or connect a human. Reminders, delivery updates, and payment alerts are common examples.

Is an automatic call the same as a robocall?

Not always. Robocall usually means a prerecorded or artificial voice message sent in bulk. An automatic call also covers software that dials and then connects a live agent. Some regulators, such as India's TRAI, define application-initiated calls broadly enough to include robocalls and agent-connected calls.

Are automated calls legal?

Generally yes, when you follow the local rules on consent, calling hours and caller ID. In the US, the TCPA restricts prerecorded and artificial-voice calls made without consent. If you're in India, TRAI's rules add advance declaration of A2P calls and set the number series to use. Check the current text and take legal advice for your use case.

How accurate is answering machine detection?

No single figure carries across markets. Accuracy depends on the settings, the destination, and how people answer. A low machine share also makes each machine verdict less reliable. Label a sample of your own call recordings, then count both error types.

Should an AI voice agent use answering machine detection?

Usually, if machines answer often enough to waste real agent time. Use asynchronous detection, keep the agent silent during the window, and start it when the result says a person answered. If the result is late, treat the call as human.

How can I stop unwanted automated calls?

Register your number on your country's do-not-call list, such as the National Do Not Call Registry in the US or the TPS in the UK, and report unwanted calls to your carrier or regulator. If you're in India, set your preferences on the DND registry, through the TRAI DND app, or report a call by calling or texting 1909.

Further reading on Vobiz

Sources

Indian regulators

US rules

Vendor documentation and reports

News

Build outbound on Vobiz — Provision a number and place your first detection-gated call.


If a voice AI agent is placing your automatic calls, Vobiz provides the SIP trunks and numbers it calls from, priced pay as you go per minute.

Top comments (0)