AI outbound calling: how it works, costs and rules

Most of what is written about AI outbound calling is either a vendor promising to replace your sales floor or a warning that robocalls are illegal. Neither helps you decide whether to run a campaign next week.

This guide is the practical version. It covers what actually happens on a call, which calls are worth automating, what a campaign costs with real per-minute prices, the rules that apply in the US and UK as of October 2026, and the lessons from a real pilot that only show up once calls start going out.

What AI outbound calling is, and what it isn't

Three different technologies get sold under the same name. The difference matters for cost, for the law, and for how the person on the other end feels.

A predictive dialer is a pacing engine for a human team: it dials several numbers per available agent and hands the answered ones over. We compare the two models in detail in predictive dialer vs AI voice agent. Voice broadcasting plays the same recording to everyone. AI outbound calling sits between them: the scale of automation, with a conversation that responds to what the person actually says.

What happens on a single AI outbound call

The quality of a campaign comes down to the first ten seconds of each call and the bookkeeping afterwards. Here is the sequence on Dial Milo, step by step.

1. The call is queued. From a CSV upload, a filtered group in your contact book, or a webhook from your CRM.
2. Local time is checked. Milo calls each person in their own time zone, inside the calling hours you set (10:00–20:00 by default). Outside the window, the call waits rather than being dropped.
3. Limits are checked. Nobody is called twice within your contact cooldown, across all your campaigns, and the daily and total budgets are respected.
4. The phone rings. If nobody answers, there is no minute charge and the call is retried later, for example two tries four hours apart.
5. Voicemail is detected. Carrier detection plus the AI listening for voicemail greetings. On a voicemail Milo hangs up without leaving your brand behind, and the person is retried.
6. A person answers. Milo opens within a second (people hang up on silence), says it is an AI, and confirms it is talking to the right person before naming your brand.
7. The conversation. One reason for calling, questions answered from your script and knowledge sources, an offer presented once with exactly the terms you wrote.
8. The outcome is recorded. Accepted, declined, callback requested with a time, opted out, wrong person or no decision, with a transcript, summary and recording.
9. The next step fires. The outcome goes back to your systems so a confirmation email, SMS or human follow-up happens without anyone copying notes.

If the person rings the number back hours later, which people do, Milo recognises them, thanks them for calling back, confirms who they are and picks the conversation up. Callers it does not recognise get a neutral line that never names your business.

Where AI outbound calling works

The pattern is consistent: short calls, one clear purpose, a person who has a relationship with you, and an outcome that can be written down in a word.

• Following up new leads in minutes. The odds of qualifying a web lead drop steeply in the first half hour, and most companies never get there. See speed to lead.
• Appointment reminders and confirmations. The evidence that reminder calls reduce no-shows is unusually solid. See automated appointment reminder calls.
• Taking repetitive work off a contact centre. Outbound follow-ups and callbacks are the overlooked half of call center automation.
• Customer and player retention. Lapsed customers, failed renewals, registered users who never bought. For online casinos and sportsbooks, see iGaming player reactivation.
• Journey-triggered calls. A CRM event fires, the AI calls, the result comes back. See Customer.io AI phone calls.

And where it is weakest: cold calling strangers. It is legally the riskiest use and commercially the least forgiving, because the person has no reason to stay on the line. If you are considering it, read AI cold calling first, honestly, before you buy any list.

Where it doesn't work

AI outbound calling is the wrong tool when the call is a negotiation, when the person is distressed, when the value of a single conversation runs to thousands, or when you cannot write down what a good outcome looks like. It is also the wrong tool when your list is dirty. An AI dials a stale, unconsented list just as efficiently as a human team would, and it reaches the complaints faster.

What AI outbound calling costs

Two parts make up each call on Dial Milo, and both are published.

• The phone line, priced for the exact number dialled and the country of your caller ID, billed per started minute the way carriers bill. US landlines and mobiles start at 4¢ a minute; a UK mobile from a UK number is 8.5¢; a Dutch mobile 10¢; a German mobile from 9.4¢.
• The AI, 9.57¢ per minute of talk time.

So a connected minute costs roughly 14–20¢ all-in for US, UK and Dutch numbers. Calls that never connect have no minute charge. You can price any number on the rates page and model a whole list with the spend calculator.

Worked example (illustrative figures). Say you call 2,000 US customers about a service renewal, two attempts each.

• 700 people (35%) eventually answer and talk for 2.5 minutes on average. The line bills 3 started minutes: 700 × 3 × 4¢ = $84. The AI bills talk time: 700 × 2.5 × 9.57¢ = $167.48.
• 900 attempts reach voicemail and are hung up. Assume about 5¢ each for the started line minute and a few seconds of AI: $45.
• The rest ring out, at no minute charge.

Total: about $296, or roughly 42¢ per real conversation and 15¢ per customer on the list. Your answer rate and call length will differ, which is why you run a small batch first.

A real reference point. In an October 2026 pilot, an online casino operator (anonymised) ran 61 players through two rounds: 121 dial attempts for a total of $11.53, about 9.5¢ per attempt. Most attempts were unanswered or voicemail, and those cost very little. That is the shape of most outbound campaigns: the bill is driven by the minority who actually talk.

Set a total and daily budget cap on every campaign. The forecast shows what the rest of the list will cost, and email alerts arrive before a cap is reached.

The rules: consent, disclosure, calling hours

This is the part most guides skip, and the part most likely to hurt you. What follows is a summary as of October 2026; the full version, country by country, is in is it legal to make AI calls.

United States. On 8 February 2024 the FCC unanimously ruled that calls made with AI-generated voices are "artificial" under the Telephone Consumer Protection Act. In practice, an AI voice call is held to the same consent standard as a prerecorded call. The FCC's consumer guide summarises it: prior written consent before a prerecorded telemarketing call, oral or written consent before an autodialled or prerecorded call to a mobile, and AI-generated voice calls are illegal unless the consumer has agreed or the caller is exempt. The FTC's Telemarketing Sales Rule also limits telemarketing calls to between 8am and 9pm at the called person's location, without prior consent.

United Kingdom. The ICO's PECR guidance is blunt about automated marketing calls: you need the person's specific consent to that type of call, and general marketing consent, or even consent for live calls, is not enough. You must also say who is calling and display your number.

What that means in practice. Marketing calls with an AI voice need consent you can prove. Informational calls to existing customers (a reminder, an order update, a callback they asked for) sit in a lighter category, but still need a clean list, an opt-out that works, and sensible hours. Dial Milo discloses that it is an AI at the start of the call by default, records "stop calling" as an opt-out, and keeps that person off every future campaign.

Building your first campaign

1. 1

   Pick one call with one outcome

   "Confirm tomorrow's appointment" or "offer lapsed customers their renewal discount." Not three goals in one call. If you cannot name the outcome in a word, the AI cannot record it either.

2. 2

   Clean the list

   Phone numbers in international format, a first name, and only people you have the right to call. Extra columns, such as plan, appointment time or language, become placeholders in the script.

3. 3

   Write a short script

   Opening line, reason for the call, the offer and its exact terms, and what to do with the three most likely questions. Milo presents an offer once, never invents or changes terms, and only gives conditions when asked. See the guide to writing a script.

4. 4

   Set the guard rails

   Calling hours, attempts and retry gap, contact cooldown, maximum calls at once, start and end dates, and total and daily budgets. See calling hours and contact limits.

5. 5

   Hear it yourself

   Use the browser demo or a test call and listen as the person on the list would. Fix anything that takes more than ten seconds to get to the point.

6. 6

   Run 100 to 300 people first

   Read twenty transcripts, check the outcomes, and change the script before you scale. If you have two openings in mind, A/B test up to five scripts side by side.

The full walkthrough of the dashboard is in running an outbound campaign.

Connecting calls to your CRM

The most useful campaigns are not uploaded once; they are fed continuously. Each Dial Milo campaign has a secret "start a call" webhook. Any CRM or journey tool that can send a webhook (Customer.io, Braze, Optimove, HubSpot workflows, an in-house system) can POST one person or up to 500 at a time, with a language, an offer code, a scheduled time and any context the AI should know.

Duplicates and re-fired journeys are ignored, opted-out people are rejected automatically, and the paced dialler still respects calling hours and pauses. With Customer.io the loop closes: after each call the person's attributes update and events such as milo_offer_accepted or milo_opted_out fire, so your journey sends the confirmation by SMS and email, or suppresses the phone channel. The end-to-end setup is in Customer.io AI phone calls.

What to measure

Calls made is a vanity number. These are the ones that tell you whether a campaign works.

Dial Milo also pauses a campaign on its own and emails you if the last 20 calls stop connecting, which usually means networks have started flagging the number. After the calls, an AI campaign report labels every person with what they need next.

Questions to ask any AI outbound calling provider

The demos all sound good. These questions separate a platform you can run a real campaign on from one that only sounds good in a demo.

• What exactly am I billed for? Per connected minute, per attempt, per seat, or a monthly plan? Is the phone line priced for the real number dialled or as a country average? Are unanswered calls free? Ask for the bill on a 1,000-person list at your expected answer rate, as a single number.
• What happens on voicemail? Does it rely on carrier detection alone, or does the AI also listen for greetings? Does it leave a message, and can you stop it doing so?
• Whose local time does it call in? Yours, or each person's? What happens to calls that fall outside the window: dropped or re-queued?
• What happens when someone calls back? Does the number recognise them and continue the conversation, or does it ring an empty line?
• Can it change the terms of an offer? It should not. The AI should present the offer you wrote, exactly, and never invent a discount to save the call.
• How is "stop calling" handled? Recorded as an opt-out, applied across every campaign, and visible to you.
• Can I cap the spend? Total and daily budgets, a forecast for the rest of the list, and an alert before money runs out.
• What do I get after each call? An outcome you can act on, a transcript, a recording, and a way to push the result into your CRM.
• Can I hear it before I pay? If the only way to hear the product is a sales call, assume the experience on your customers' phones will be similar.

A provider that answers all nine plainly is one you can test properly. One that answers with "it depends, let's set up a call" is telling you something too.

Lessons from a real pilot

The October 2026 casino pilot mentioned above was small, which is the point. These issues show up within the first hundred calls of any campaign, and all of them are now handled by default:

• Carrier voicemail detection is not enough. It labelled some voicemail greetings as humans, so the AI must also listen for voicemail and hang up.
• People call back. Players rang the number back hours later. If the callback is not recognised, a warm conversation turns into a confused one.
• Language is not optional. Some players answered in Dutch. Pass a language per person and let the AI switch mid-call when the person does.
• "Yes" in the middle of a sentence is not a restart cue. The AI has to keep listening rather than relaunch the pitch.
• The pause after pickup must be under a second. Silence is the signature of a bad dialler, and people hang up on it.
• Confirm identity before naming the brand. Especially where the brand itself is sensitive information.
• Say how the details will arrive. "You'll get them by SMS and email" stops the "can you repeat the code?" spiral.

Getting started

Price your numbers on the rates page, hear Milo call you in the browser before you pay, then top up from $10. There is no subscription. Run a few hundred people, read the transcripts, and only then decide how big this should be. Create an account when you are ready.