What Is an Auto Dialer? Preview, Progressive, Predictive

November 3, 2026 · By Piyush Sahoo An auto dialer is software that places outbound calls from a contact list without a person dialing each number. It then sends each answered call to an agent, a recorded message, or an AI voice agent. Preview, progressive, and predictive modes differ in one thing: when the dial happens relative to the agent. This guide explains how the machinery works, models what each mode does to agent time, and shows where predictive pacing breaks. It also sets out the Indian rules on auto dialers, including an abandoned-call limit that none of the vendor guides we read mention.

​

What is an auto dialer?

This kind of system starts calls from stored instructions rather than from a person’s fingers. India’s regulator offers a useful definition. TRAI’s consolidated commercial communications regulations describe equipment that calls stored or generated numbers and, on answer, either plays a recorded message or connects a live person. Those two outcomes explain why the words around this topic get tangled. Different people use the same term for different systems, so check which one a vendor or lawyer means: So a predictive dialer is a kind of auto dialer, not an alternative to one. The US case also shows that a legal label can be much narrower than the product label. A system can be an auto dialer in the sales brochure and not an ATDS in court.

​

How does an auto dialer work?

Every auto dialer runs the same six stages in a loop. The modes differ only at stage 3. The pacer needs clean signals, and they arrive in two layers. The signalling layer comes first. The telephony side sends a SIP INVITE, then reads the responses defined in RFC 3261. 180 Ringing means the phone is alerting, 486 Busy Here means a busy line, and 200 OK means something answered. That “something” is the catch, because the network treats a person and a voicemail greeting alike. The media layer comes second. Answering machine detection listens to the first seconds of audio to separate people from machines. That takes time, and time is exactly what an abandoned-call rule limits. Here is the conflict in numbers. US rules give you two seconds after the greeting to connect a representative. On any dialer, an agent-first flow that waits on a longer detection window can overrun that limit on every live answer, so size the window on purpose. On Vobiz, you set machine_detection_time from 2,000 to 10,000 ms, or supply machine_detection_url to detect in the background. Test this before you turn AMD on for a human-agent campaign. AMD errors also matter more than they look. Ofcom’s 2010 statement treats a person wrongly hung up on as a machine as an abandoned call, and asks you to estimate those in your rate. The dialer can’t see these errors in its own logs.

​

Preview, progressive and predictive compared

The clearest way to see the modes is to ask who waits for whom. In preview, the call waits for the agent. In predictive, the agent barely waits for anything, and sometimes the customer does. Vendors don’t use these words identically, so read their documentation rather than the label. Three differences are worth knowing:

• Preview can auto-dial. Some dialers offer a timer that places the call if the agent doesn’t act. That matters for the Indian A2P question below.
• “Power” is vendor-specific. Some products use it for several dials per idle agent, with no forecast of when busy agents will finish. Don’t assume another product means the same thing.
• Progressive isn’t always zero-risk. A blended agent can take an inbound call while an outbound call is connecting, and that call then gets abandoned.

​

How much does each mode change agent time?

Mode comparisons usually stop at “predictive is faster”. We built a small model to put numbers on it, including where that stops being true. The chart shows occupancy, the share of an agent’s logged-in time spent in talk and wrap-up, with every predictive option held to 3% of answered calls abandoned. Assumptions, all ours and illustrative:

• 25% of attempts reach a person, inside the 15–30% range our capacity calculator gives for Indian outbound dialers.
• An answered call rings for 10 seconds on average, and any other attempt holds a channel for 25 seconds.
• Talk time averages 180 seconds, with 30 seconds of wrap-up. Preview adds 20 seconds of reading per record.
• A person who answers when no agent is idle counts as abandoned straight away. There are no breaks and no shared agents.

How we calculated each line:

• Preview and progressive come from simple cycle arithmetic, so team size doesn’t change them.
• Fixed-rate dials at a steady pace with no feedback. We solved it exactly with the Erlang B formula.
• Reactive adjusts dialing at every call event. We simulated it over 8 hours and 8 runs per team size, and gave it the true answer rate and handle time. Real forecasts are worse, so treat this line as a ceiling.

Three things stand out:

1. A simple overdial loses to progressive on small teams. The fixed-rate line stays below 71% until about 25 agents. Below that, it pays with abandoned calls and still gets less agent time.
2. Feedback does most of the work on small teams. The reactive line pulls well ahead at 10 to 30 agents, but only if its inputs are good.
3. The gain flattens. Going from 100 to 200 agents adds a little over 4 points on the fixed-rate line and just over 1 on the reactive line.

Vendor guidance on the threshold varies. Our model suggests the answer depends on how good your pacer is, so test yours rather than trusting a number. Answer rate moves the picture too. With the same assumptions, progressive occupancy falls to 58% at a 15% answer rate and rises to 75% at 30%. List quality can be worth as much as a change of mode.

​

What each mode asks of your trunk

Faster modes need more telephony capacity, because unanswered attempts hold channels too. Here are the same assumptions at a team of 50, and for an AI campaign capped at 50 channels: These are averages from our model, not benchmarks. Note that predictive needs more channels than agents, because several calls are ringing for each conversation. Pacers also burst, so check your calls-per-second ceiling as well as the average. Our SIP trunking guide has the wider sizing method.

​

Why do predictive dialers abandon calls?

A predictive dialer is making a bet on the near future. It assumes a certain share of its calls will reach people, and that certain agents will finish on time. When reality beats the forecast, more people answer than there are agents, and someone gets dropped. The surprise is how fast this happens, and which teams it hits hardest. We tuned a fixed-rate dialer to exactly 3% at a 25% answer rate, then changed the real answer rate. A rise to 30% looks small, yet the 200-agent team reaches about 15% abandoned calls, against about 7% for the 15-agent team. Large pools run closer to full capacity, which is why they’re so efficient and also why they have so little slack. Scale makes predictive worthwhile, and it also shortens your reaction time. Answer rates can shift during a campaign when the audience or the hour changes, so build the controls in from the start:

• Measure per campaign, per window. A daily average across all your campaigns can hide a bad hour on one of them.
• Count AMD errors. Add an estimate for people wrongly classified as machines.
• Pace down fast and pace up slowly. The cost curve is lopsided, as the chart shows.
• Give abandoned callers something. Silence is the worst case. The US safe harbor and Ofcom’s policy both expect a short message with your identity, and it’s a sensible default elsewhere.
• Don’t re-dial abandoned numbers on autopilot. Ofcom treats repeat calls within 72 hours with no guaranteed live agent as an aggravating factor.

​

Which dialing mode should you choose?

Start with the slowest mode that meets your volume target, then move faster only when you can measure the cost. The 25-agent line comes from our model, so treat it as a starting point. A strong reactive pacer can justify a smaller team. AI voice agents change the question. With an AI session on every answered call, there’s no agent shortage to forecast, so overdialing has nothing to gain. The limits become channels, sessions, and how quickly the agent can speak after answer. Pacing turns into a simple concurrency cap, which is a much easier problem. The regulatory questions below still apply in full.

​

What rules apply to auto dialers in India?

India regulates auto dialers by name, and the rules tightened in September 2026. They cover what the call is about, which number it comes from, and whether software started it. They change often, so check the current text and take legal advice for your own case. Check when the A2P terms start. The Third Amendment sets regulations 4 and 18, the A2P declaration and charge, to begin 60 days after Gazette publication, and most other changes after 30. We could not confirm the Gazette date. If it matched the 18 September notification, by our count the 60 days end around 17 November 2026. Four details change how you build a dialer: A live agent doesn’t take you out of A2P. Stakeholders asked TRAI to exclude calls with a human on the line, naming agent dialers among their examples. TRAI declined, saying providers can’t technically tell the cases apart. On that reading, progressive and predictive calls are A2P calls. A preview call the agent places personally looks closer to “direct human dialling”, while a preview timer that dials by itself looks less so. Ask your provider how they treat your flow. India’s 3% uses a different denominator. The US rule and Ofcom’s 2010 criteria measure against calls answered by a person, while TRAI’s text measures against all attempts. At a 25% answer rate, 3% of attempts equals 12% of answered calls, so the Indian line is looser on paper. Don’t design to it, because 12% of your answered customers meeting a dead line will generate complaints, and complaints carry their own triggers. The definition of an abandoned call in TRAI’s text also has no two-second clock. It has no carve-out for agentless calls either, so ask your provider how recorded or AI calls count toward the ratio. Enforcement looks across your numbers. Under the September 2026 changes, five or more of a sender’s numbers flagged by provider AI systems in ten days starts graded action, and three complaints plus an AI flag can also trigger it. Rotating through more numbers spreads your load, but it won’t separate your numbers from you as a sender. The same changes say 140xx, 1600xx, and 1601xx numbers aren’t flagged as suspected spam to recipients. Consent has shorter clocks than many lists assume. The Third Amendment limits communication based on a customer’s inquiry to seven days, and requires that inquiry to be written or digital and verifiable. Dialers working old lead lists should check this first. Subscribers can also block by mode, with separate options for auto dialer calls with a recording and those with a live agent. TRAI’s explanatory note shows where its attention is. In April 2026, 15,736 of 20,591 disconnections for these violations were SIP or PRI connections, which TRAI describes as used for A2P calls.

​

How do India, the US and the UK compare?

If you call across borders, the same word can mean different things: Ofcom says plainly that its 2010 figure of 3% was never a safe harbor, only a way to prioritize cases. Under its policy from 1 March 2017, it may act in any case that meets the statutory definition of misuse. The word “silent” also differs: in the UK it means an answered call with nothing on the line, while TRAI uses it for short calls designed to get a callback.

​

How Vobiz fits into an auto dialer

Vobiz Campaigns is a bulk outbound dialer built around concurrency. You upload a contact list, attach a reusable set of answer and hangup URLs, set a concurrency ceiling, and Vobiz works through the list. That matches the agentless pattern directly. For preview, progressive, and predictive, agent state lives in your contact center software, which places calls through the Make Call API.

• Pacing. A campaign’s max_concurrent caps simultaneous calls. Create a campaign in the console or over the Campaign Manager API.
• Caller ID. Choose fixed, per_contact, or pool. Pools rotate with round_robin, least_used, or random, and support per-number daily and lifetime caps plus pool_cooldown_seconds.
• Calling windows and retries. Set a daily window in the campaign’s timezone, and retry_attempts from 0 to 5 with your own retry_delays.
• Feedback. Signed webhooks (callbacks) report campaign states and contact outcomes. Reasons for pausing include window_closed and caller_pool_daily_limit_reached.
• Capacity. The account capacity endpoint shows free capacity while a campaign runs. Size concurrency and CPS with the calculator. Going over either limit returns a 429 error, so pace your own dialer below them.
• Connecting an agent. On answer, Vobiz fetches VobizXML from your answer_url. Use <Dial> to bridge to an agent’s number or a registered endpoint, or <Conference> for a shared room.
• Machine detection. On a Make Call request, set machine_detection to true or hangup, or supply machine_detection_url to run it in the background. A hangup result ends with hangup cause 9100, which separates machines from no-answers in your data.
• AI voice agents. Audio streams carry call audio to your agent, and the AI voice agent guide shows the integrations.
• India compliance. See calling regulations for KYC and media anchoring, UCC and NDNC for scrubbing and complaint timelines, 140 and 160 series acquisition for commercial numbers, and DLT registration. Current rates are on the Vobiz pricing page.

Campaigns and test calls ring real people and are billed to your account. Run them with your own Auth ID and Auth Token, pilot on a small list you have consent to call, and confirm numbers before you launch.

​

Frequently asked questions

​

Further reading on Vobiz

• How an outbound campaign manager works
• Number utilization guide
• Automated outbound calling
• Scaling outbound calling and AMD
• What is answering machine detection?
• CPS and concurrency visualizer

​

Sources

Indian regulators

• TRAI, Telecom Commercial Communications Customer Preference Regulations, 2018, consolidated text, 21 May 2026. TRAI notes that the gazetted text prevails.
• TRAI, Second Amendment Regulations, 2025, February 2025.
• TRAI, Press release 11/2025 on the amendments, February 2025.
• TRAI, Press release 91/2026 on the 1600 and 140 series, 10 July 2026.
• TRAI, Third Amendment Regulations, 2026, with explanatory memorandum, 18 September 2026.
• TRAI, Press release 119/2026 on the Third Amendment, 18 September 2026.

Other regulators and courts

• US FTC, 16 CFR 310.4, Telemarketing Sales Rule, via Cornell LII.
• US FCC, 47 CFR 64.1200, via Cornell LII.
• US Supreme Court, Facebook, Inc. v. Duguid, April 2021.
• Ofcom, Persistent Misuse policy statement, December 2016, applying from 1 March 2017.
• Ofcom, Tackling abandoned and silent calls, October 2010.

Standards

• IETF, RFC 3261: SIP: Session Initiation Protocol, June 2002.