AI agent that makes phone calls: how it works, what to check

A collections company calls 1,200 customers in one day with an AI agent that makes phone calls. The conversations that happen go well, and the pilot looks like a win. But nine out of ten calls went to voicemail, rang out or lasted under six seconds. In Brazil, starting November 12, 2026, that pattern can cost a business 15 days with every one of its lines blocked.

That is why the question is no longer just “does the voice sound human?” It still matters, but the decision to adopt now depends on your list, your operation and telecom rules.

An AI agent that makes phone calls is software that holds a live phone conversation: it transcribes what the person says, decides the next action from a script, rules and CRM data, replies in a synthetic voice and logs the outcome. It takes and places calls with no keypad menu.

How a voice agent works under the hood

The demo feels like magic, but the mechanism is an assembly line. Speech becomes text, a language model reads that text with the customer’s context, a rules layer limits what it may do, actions run in your systems, and the reply comes back as voice. This repeats on every turn.

Two parts tend to be underestimated. The first is interruption: real people cut each other off, and the agent has to stop talking at once and pick up without losing the thread. The second is tools. An agent that only talks is a chatbot with a voice. The value shows up when it checks the calendar, updates the deal in the CRM, issues an invoice and leaves a clean history for the sales rep.

In practice, what separates a good rollout from a bad one is rarely the voice. It is the quality of what the agent can see. If the CRM is stale, the agent will confidently pitch something the customer already bought.

Where an AI calling agent works best

The pattern is simple: a short conversation, a single goal, a verifiable result. The more a call depends on open negotiation or built-up trust, the worse it performs.

Use case	Why it works	Where it breaks
Instant callback to a form lead	The person just asked for contact, so pickup tends to be higher	Forms without clear consent to be called
Qualification and booking	Fixed questions and a clear outcome: meeting booked or lead discarded	Complex sales with several decision makers
Pre-due payment reminder	A short message before the due date, with the payment link right after	Old or disputed debt, or a customer in hardship
Appointment confirmation	Cuts no-shows and frees up the front desk	Someone rescheduling with many constraints
Reactivating an old base	Low cost per attempt	Uncleaned lists that generate lots of short calls

In sales: the agent prepares, it does not close

For high-ticket products, use the agent at the top of the sales funnel: call within minutes, qualify, book. The rep steps in with the conversation already summarized in the CRM. Trying to close a meaningful contract with a synthetic voice usually saves little and burns good leads. The logic is the same as in our piece on AI for sales: remove the repetitive work, not the moment that decides the purchase.

In collections: more care than it seems

Collections is the most cited case because volume is high and the script is repeatable. It is also where mistakes cost the most. An agent that keeps calling a wrong number, reveals a debt to a relative or pressures someone who asked it to stop creates legal and brand trouble. Start with reminders before the due date and simple renegotiation with fixed discount bands. Sensitive cases go to a person.

The rule that changes the math: short calls

This is a Brazilian rule, and worth watching if you call Brazil. Provisional Measure 1,396/2026 and Anatel Decision 100/2026, reported by Teletime in October 2026, define abusive use as a user who, on the same day, exceeds 1,000 calls per line or 10,000 across their numbers, with at least 85% short calls. A short call is one that goes unanswered, reaches voicemail or lasts up to six seconds.

The rules apply from November 12, 2026. The carrier notifies the customer and, if the abuse continues, blocks all of their lines for 15 days, with no new lines during that time. The measure still needs approval from Congress and Anatel is still issuing regulations, so treat this as a moving target and check with your lawyers.

A voice agent makes it easy to fall into this, because dialing a thousand numbers is now cheap. Here is a simulation with 1,200 calls from one number:

• Scenario A, cold list: 12% pick up (144 people) and 30 hang up within six seconds. That is 1,056 + 30 = 1,086 short calls, or 90.5%.
• Scenario B, qualified list: 20% pick up (240 people) and 40 hang up quickly. That is 960 + 40 = 1,000 short calls, or 83.3%.

The same operation lands on opposite sides of the line purely because of list quality. The fix is not to massage the number. It is to call fewer people, better. Prioritize those who asked to be contacted, respect the hours each segment answers and spread volume through the day. The measure also calls for call authentication and for showing the company name and logo, which should help legitimate callers.

What it really costs

Retell, a vendor in this market, lists US$0.07 per minute in an October 2026 article, yet the calculator on the same page defaults to US$0.11. That gap is typical: premium voices, telephony and pricier models tend to land outside the sticker price.

A calculated example, not research: 5,000 calls a month, 20% answered. That is 1,000 conversations of 2.5 minutes (2,500 minutes) and 4,000 attempts of 18 seconds (1,200 minutes), for 3,700 minutes in total. At sticker price that is US$259; at US$0.11, US$407. If 8% of conversations become meetings, you get 80 meetings at a platform cost of US$3.24 to US$5.09 each.

That number is tempting and incomplete. It leaves out integration, script tuning, call reviews and supervision time. And the final metric is not cost per minute. It is cost per meeting held, compared with your current CAC.

What to evaluate before you adopt

1. Human handoff. In a Gartner survey published in August 2026 of 3,566 customers, 87% said it is essential to be able to reach a person when a company uses generative AI. “I want to talk to someone” has to work the first time, and the human should receive the summary.
2. Real language coverage. Test accents, street noise, spelled-out IDs, amounts and proper names. This is where many polished demos fall apart.
3. Interruption. Cut the agent off mid-sentence, change the subject and come back. If it keeps talking, the customer hangs up.
4. Written limits. List what it may not promise: out-of-band discounts, deadlines, contract terms. Log every action.
5. Data and transparency. Have a lawful basis for calling and recording, follow the data protection rules for marketing and say up front that this is a virtual assistant. Hiding it buys a few seconds and costs trust.
6. Cohort measurement. Compare a group handled by the agent with a control group. Without that, you cannot tell whether you gained results or just moved work around.

Start small, as with any AI agent rollout: one use case, one clean list, a few hundred calls reviewed by a person. Only then increase volume.

Frequently asked questions

Can AI make and answer phone calls on its own?

Yes. A voice agent takes and places calls, talks in real time and runs actions in other systems. What still needs careful design are exceptions, complaints and off-script negotiations, which should go to a person.

What is the difference between a voice agent and an IVR?

An IVR follows a fixed menu. A voice agent understands free speech, keeps context and checks systems during the conversation. It solves more cases, but it needs more supervision, because it can fail in ways a menu cannot.

Is it legal to use AI to make calls?

It depends on the country. In Brazil there is no known general ban, but data protection, consumer law and telecom rules apply, including the short-call limits. Before scaling, confirm the lawful basis, recording rules and latest regulations with your lawyers.

Will people notice they are talking to a bot?

In short calls, it can take them a while. Even so, I recommend saying so in the opening line. When someone finds out later, trust in the brand drops and a good conversation turns into a complaint.

How much does an AI calling agent cost?

Some platforms advertise from about US$0.07 per minute, but the real cost rises with telephony, voice, model and implementation. Calculate cost per result, such as a meeting held or an invoice paid, not per minute.

Voice is the channel; the decision comes from data

A voice agent is a great tool if you already know who should get the call, when, and with what offer. If you do not, it just makes mistakes faster and at higher volume. Before you dial, cross your CRM, media and results by list. With Sherlok, you ask in plain language which segment picks up most, which source turns into meetings and where cost per result gets worse, then get the analysis ready to decide who is worth calling.