Outbound and sales calls

AI Outbound Calls: Uses, Limits and Cautions for Teams

Where an AI agent can make outbound calls responsibly, where it should not, and the consent, disclosure, calling-hours and hand-off rules to set first.

Call editorial team · Published · 6 min read

Outbound calling is the part of voice AI that attracts the most attention and the most caution. An AI agent can place calls to a list of people, one at a time, follow a set goal and record the outcome. It never gets tired and it does not hesitate to dial the fortieth number of the day. But a call to a person who did not expect it is a different matter from answering an inbound call that they chose to make.

A person at a desk reviewing a list of planned outbound calls on a monitor
Illustrative image

This guide explains where AI outbound calls make sense, where they do not, what rules to look up before you begin, and how to run a careful first batch. If you are new to the fundamentals, read Outbound Calling Basics first.

Where AI outbound calls fit

The strongest candidates are calls where the purpose is narrow, the answer is short, and a wrong move costs little. Think "confirm" and "collect", not "persuade".

Reasonable uses

  • Appointment confirmationsAsk whether a booked time still works and record yes, no or "call me".
  • Reminder callsRemind a person of a visit or a deadline they already agreed to.
  • Short surveysAsk two or three questions after a service and log the answers.
  • Lead follow-upCall a person who asked to be contacted, to find a good time for a human to ring back.
  • Overdue informationAsk a customer to send a missing document, with a route to a person.

Poor fits

  • Cold sales to strangersUnwanted calls generate complaints and are often restricted.
  • Sensitive topicsDebt, bereavement, health results and complaints need a human.
  • Complicated negotiationsAnything with many exceptions or emotional stakes.
  • Vulnerable callersPeople who may be confused by an automated voice should reach a person quickly.

A useful test: if a thoughtful manager would be uncomfortable hearing the call on a speaker in the office, do not automate it.

Rules to settle before the first call

Outbound calling is regulated more tightly than inbound, and calls placed by an AI voice are often treated more strictly again. Settle these points in writing before you dial.

QuestionWhat to decide
Do we have permission to call this person?Record the basis, such as an existing relationship or a request to be called, and keep evidence.
Are numbers screened?Check your own suppression list and any national registers that apply to your market.
When may we call?Respect the person's local time and the local calling-hour limits, not yours.
Do we say it is an AI?Open with who is calling, on whose behalf and that it is an automated assistant.
What if they say no?End politely at once and add the number to the suppression list.
Who takes over?Define when and how a person joins or calls back.

In some markets, calls with an artificial or pre-recorded voice need prior written consent, and several regulators also expect clear disclosure that the voice is not human. For the US picture see US TCPA Basics for Japanese Businesses, and for the disclosure question see Should You Disclose That an AI Is Answering?. Even where the law is silent, plain disclosure builds trust and ends calls that were going nowhere sooner.

Calling hours and frequency

Every country has its own calling hours and, in some places, limits on how many attempts you may make in a day or week. Set the campaign so that it reads the contact's local time, and do not rely on staff to remember. The same applies to repeated attempts: a person who did not answer on three consecutive days is telling you something.

Designing the call

An AI agent works best with a script that is short, with a single goal and plain language. Because the other person has not asked for the call, the first ten seconds decide whether they stay on the line.

Anatomy of a considerate outbound AI call

  1. IdentifyName the business, say the caller is an automated assistant, and state the reason in one sentence.
  2. Check the momentAsk if now is a good time; offer a callback if not.
  3. One purposeAsk the single question the call exists for.
  4. Handle the common repliesYes, no, not now, who is this, and "stop calling me".
  5. CloseConfirm what was recorded and thank the person.
  6. Hand offOffer a person for anything outside the goal.

Plan for the answers you do not want

Write out the replies you hope not to hear and decide what the agent does with each. "Don't call me again" must trigger an immediate polite exit and a suppression entry. "Are you a robot?" must receive an honest answer. Questions the agent cannot answer should lead to a callback request rather than a guess. For more on limits, see Guardrails for Voice AI Agents.

Voicemail, machines and silence

Many outbound calls reach a voicemail greeting or an automated menu. Decide in advance what the agent does: leave a short, neutral message that names the business and a way to reach it, or hang up. Machine detection is a carrier verdict and can be wrong, so never rely on it to say anything sensitive. Avoid leaving details that would be private if someone else heard them.

Silent calls, where the person says hello and hears nothing, are the quickest way to annoy a recipient and attract complaints. Watch your silent-call rate and pace the campaign so that your agent can start speaking at once.

Running a careful first batch

  1. Pick one narrow use, such as appointment confirmation.
  2. Prepare a clean list: valid numbers, suppression applied, consent noted.
  3. Run a handful of test calls to colleagues' phones and read the transcripts.
  4. Start with a small batch during the contact's daytime hours.
  5. Listen to a sample of calls every day and fix wording, then widen the batch.

What to measure

Looking only at "calls made" tells you nothing. Track the share that were answered, completed the purpose, asked for a person, asked to stop, or ended early. Review the stop requests and complaints personally, because they show where the script or the list is wrong. See Voice AI Testing for how to rehearse before launch.

Go-live checklist

  • Permission basis and evidence are recorded for every number.
  • Suppression lists and applicable registers are applied automatically.
  • Calling hours follow the contact's local time.
  • The agent states who is calling and that it is automated.
  • A polite exit and a hand-off route are tested.
  • A named person reviews calls daily during the first week.

Doing it in Call

Call can place AI agent calls to a list of numbers, one at a time. You create a batch with a goal and instructions (or have the draft written for you), save it, which places no calls, and then press Start. You can pause, resume and stop while it runs, and each call's result appears in call history.

See Building an AI Agent in Call and Running an Outbound Dialer Campaign in Call for the setup steps.

Frequently asked questions

Can an AI agent make cold calls?

Technically yes, but cold calls to people who have not asked to be contacted are restricted in many places and generate complaints. Most teams do better using AI for confirmations, reminders and follow-ups to people they already have a relationship with.

Must the AI say it is an AI?

Rules vary, and some regulators expect disclosure. Even where it is not required, opening honestly about who is calling and that the assistant is automated is good practice and avoids confusion.

What if the person asks to speak to a human?

The agent should offer a transfer or a callback straight away. Define during setup who answers and when, so the answer is not "nobody is available".

Can the agent take payments or confirm a booking?

Not in Call. The agent can take a request, and a person confirms it. Keep payment handling out of automated calls.

How many calls should the first batch have?

A small one, enough to hear a range of replies. Listen to the recordings, correct the script, and widen the list only when the calls match what you intended.

Build an outbound AI batch with calling-hours checks in Call

In the demo console you can try incoming calls in a simulation.

Related guides