Can AI Make Outbound Calls for Your Business? Costs and Rules

Yes. Voice AI platforms can dial a list, hold a short scripted conversation, leave voicemails and hand over to a person, for roughly $0.07 to $0.31 a minute on pay-as-you-go plans checked in September 2026. The harder part is the rules: consent, AI disclosure, do-not-call lists and calling hours, which vary by where your customers live.

The use case decides most of the risk. Calls to people who asked to hear from you, about a viewing they booked or a renewal date they're expecting, sit at the easy end. Cold calls to lists that never gave consent sit at the hard end: in many places, calls using automated or artificial voices need prior consent, and cold calls are also where AI callers annoy people most. Start from the easy end.

Follow me on Instagram@sagnikteaches

What an AI calling agent does on the line

An outbound voice agent combines four parts: telephony to place the call, speech recognition to hear the reply, a language model to decide what to say, and a synthetic voice to say it. Around those sit the useful extras: a list or CRM trigger that decides who to call, a knowledge base the agent can draw on, a calendar it can book into, and a transfer to a person.

Connect on LinkedInSagnik Bhattacharya

In practice it handles short, predictable conversations well: confirming an appointment, collecting a yes or no, offering two time slots, taking a message, leaving a voicemail. It's weaker at anything open-ended or emotional: a tenant upset about a repair, a landlord negotiating a fee, someone who has just had bad news. It will occasionally mishear a name or answer a question it should have deferred. Design every call on the assumption that some fraction will need a person, and make that hand-over easy. For the inbound side of the same technology, see what AI can and can't handle when it answers your phone.

Subscribe on YouTube@codingliquids

Published per-minute prices from four platforms

These are the prices on each vendor's own pricing page, checked on 29 September 2026. They change often, so treat them as a snapshot and recheck before you budget.

Two lessons from the table. First, the model is often the smallest part of the per-minute cost; infrastructure, voice and telephony make up most of it. Second, some platforms have moved upmarket. A small business making a few hundred calls a month belongs on pay-as-you-go pricing, not a five-figure contract.

Pricing 300 viewing confirmations a month for a letting agency

Suppose a letting agency, an illustrative one, books about 300 property viewings a month through its website. The day before each viewing, a member of staff phones the applicant to confirm, which stops a lot of wasted trips. Say 210 people answer and talk for about a minute and a half, 60 calls go to voicemail with a 30-second message, and 30 ring out. That's roughly 315 + 30 = 345 billable minutes, assuming the platform bills connected time; check how yours bills ringing and voicemail.

• Retell, budget set-up (platform voice, Claude 4.5 Haiku): $0.055 + $0.015 + $0.025 + $0.015 = $0.11 a minute. 345 minutes is $37.95, plus $1.50 in batch-dial fees: about $39 a month.
• Retell, premium set-up (ElevenLabs voice, GPT 5.6 Terra, branded caller ID): $0.174 a minute is $60.03, plus $1.50 in dial fees and $30 for branded caller ID on 300 calls: about $92 a month.
• Vapi, using its own estimate of $0.082 to $0.129 a minute: about $28 to $45 a month, plus $29 if you take the Core package.
• Bland, Start plan: 345 minutes at $0.14 is $48.30 a month, plus $0.05 a minute for any transferred time.

So the calls themselves cost somewhere between about $30 and $90 a month. Set-up is separate: writing and testing the script, connecting the agency's booking system and CRM, and building the opt-out and callback steps. Expect 10 to 15 hours of someone's time. At the $30 to $50 an hour that Upwork's 2026 rates guide lists for AI developers, a freelancer would charge roughly $300 to $750.

The same arithmetic works for any outbound job. Estimate answered calls times average minutes, add voicemail minutes, multiply by the per-minute rate, and add per-dial and per-call extras. Then add the time a person spends on transfers, callbacks and checking recordings, which is usually the biggest line. For logging each call's outcome automatically, see AI call summaries that log every call in your CRM.

Charges that sit outside the per-minute rate

• Voicemail and unanswered calls. A list with poor answer rates can double the minutes per useful conversation.
• Transfers. Minutes spent connected to your staff may be billed separately, as on Bland's plans.
• Numbers and caller ID. Extra phone numbers, and branded caller ID, which helps stop your calls being labelled as spam ($0.10 a call on Retell).
• Concurrency and retention. Vapi's pay-as-you-go option includes 4 concurrent calls and 14 days of raw data retention; more of either needs a package or add-on.
• Do-not-call screening. Checking marketing lists against a national register can carry its own fee, depending on the country.
• Advice. An hour with a solicitor before any marketing calls is cheap compared with a complaint.
• Staff time. Listening to a sample of calls, handling callbacks, and fixing the script when it gets something wrong.

Calling rules that apply in most places

Automated calling is regulated almost everywhere, and the details differ by country and sometimes by region within a country. What follows is the common ground, not legal advice. Check the rules where the people you're calling live, and take advice before any marketing calls.

1. Consent. Marketing calls that use automated or artificial voices typically need the person's prior consent, often explicit and recorded. At least one national telecoms regulator ruled in February 2024 that AI-generated voices count as "artificial" voices under its robocall law, so the consent rules for recorded-message calls apply to AI callers too. Service calls to people who gave you their number for that purpose are treated more leniently in many places, but not everywhere.
2. Saying it's an AI. If you call people in the EU, the transparency duties in Article 50 of the EU AI Act, in force since 2 August 2026, require that people are told they're interacting with an AI system unless it's obvious. Say it in the first sentence wherever you're calling; people react far worse to finding out halfway through.
3. Who's calling and why. Name the business and the purpose at the start, give a callback number, and display a real number that can be called back. Never hide or fake caller ID.
4. Do-not-call lists. Screen marketing lists against the national do-not-call register where the person lives, keep your own suppression list, and act on "don't call me again" immediately.
5. Calling hours. Many places restrict calls early in the morning, late in the evening, and sometimes on certain days. Use the recipient's local time, not yours.
6. Recording notices. Most platforms record and transcribe by default. Say so at the start; some places require everyone on the call to agree.
7. Silent and abandoned calls. Don't let a dialler place more calls than it can handle; some regulators set limits on calls that connect to silence.
8. Voicemails. A pre-recorded or synthetic voicemail can count as an automated call under the same rules.
9. Personal data. Recordings and transcripts are personal data under data-protection law such as the GDPR: set retention periods, check the platform's data processing terms, and be ready for requests to see or delete them.
10. Cloned voices. Only with written consent from the person whose voice it is, and never to impersonate someone. Staff should also know how voice cloning is used against businesses; spotting deepfake voice and video scams covers that side.

Some sectors add their own rules on top: debt collection, financial services and healthcare especially. If your calls touch money owed, health or regulated products, take specific advice.

Which outbound calls are sensible to hand to AI

Other businesses sort the same way. A campsite confirming late arrival times with guests who booked is at the easy end. A holiday-let manager ringing every past guest to promote winter breaks is at the hard end, and an email with an unsubscribe link does the job with a fraction of the risk. If most of your phone problem is actually incoming calls, setting up an AI phone agent for after-hours calls may be the better first project.

A confirmation-call script with the disclosures built in

Keep the script short, put every disclosure in the opening, and give the caller two ways out at any point. The agency's first version:

Hello, this is an automated assistant calling for [agency] about
your viewing of the property on [street name] tomorrow at [time].
This call is recorded so we can check it.
Can you still make it? Please say yes, no, or change it.

If YES:  Thank you. The agent will meet you outside. Goodbye.
If NO:   Thanks for letting us know. We'll cancel it. Goodbye.
If CHANGE: I can't rebook on this call, but a colleague will call
         you back today. Is there a time that suits you?

At any point: say "person" to speak to someone, or "stop" and we
won't call you again.
Anything not covered: "I don't have that information. A colleague
will text you the answer today."

An illustrative test call showed two problems. The applicant asked, "Is there parking at the flat?", and the agent answered, "Yes, there's a private space," which wasn't in the property details it had been given. The fix was to limit answers to the knowledge base and use the "I don't have that information" line for everything else. The second problem was quieter and worse: when a tester said "stop, don't call me again", the agent replied politely, but nothing wrote the opt-out back to the CRM, so the number stayed on next week's list. That's the kind of failure that turns into a complaint. The fix was a step that sets a do-not-call flag on the contact the moment "stop" is heard, and a weekly check that flagged numbers are really being skipped.

Twenty test calls before a tenant hears it

Before any real applicant is dialled, run about twenty calls to colleagues playing awkward callers: someone who talks over the opening, someone who asks three questions at once, a poor mobile line, a voicemail box, a request for a person, a "stop", and a caller who doesn't speak much English. For each, check four things: did the disclosures come first, did it stay within what it knew, did the opt-out and transfer work end to end, and did the CRM record the right outcome.

Then go live with a small slice, say 30 calls in the first week, with someone listening to every recording. Widen only when the listening turns up nothing new. Keep listening to a sample, perhaps one call in ten, for as long as the agent runs. For how this fits with the agency's other tenant communication, see handling tenant queries with AI.

Decide what you'll track before launch, so the first month answers a question. For the agency, six numbers mattered, shown here with illustrative first-month figures:

• Answer rate: 213 of 300 calls answered (71%).
• Completed calls: 188 of those gave a clear yes, no or change (88%).
• Transfers and callbacks: 18, mostly questions about parking and pets.
• Opt-outs: 2, both correctly flagged in the CRM.
• Complaints: none.
• No-shows at viewings: down from about 14% in the months before to 9%.

The parking and pet questions went straight into the knowledge base for the next month, which is how the transfer rate comes down.

AI or a person on the phone: the monthly sum

Back to the 300 confirmations. Done by a person, each answered call takes about 3 minutes including dialling and logging, and each unanswered one about a minute: 210 × 3 + 90 × 1 = 720 minutes, or 12 hours a month. At $22 an hour that's about $264.

With AI, the calls cost $30 to $90 a month on the platforms above. Add a person's time for callbacks and transfers, say 20 calls at 4 minutes, and listening to one call in ten at 2 minutes each: about 140 minutes, or $51. The total is roughly $80 to $140 a month, a saving of $120 to $180, and a set-up of $300 to $750 pays back in two to six months.

That's a real saving, but not a dramatic one, and it depends on the calls staying short and predictable. If applicants start asking detailed questions, transfers rise, minutes rise and the sum narrows. It's worth running for a month and comparing the no-show rate with the months before; if fewer applicants miss viewings, that's worth more than the staff time saved.

AI outbound calls: questions owners ask

Can an AI caller use my own voice?

Several platforms can clone a voice from recordings, but think carefully before doing it. Use it only with the written consent of the person whose voice it is, never to imply that a real person is on the line, and still tell people at the start that they're speaking to an automated assistant. A clear synthetic voice with an honest introduction usually goes down better than a convincing imitation.

Do I have to tell people the call is recorded?

Assume yes. Most platforms record and transcribe calls by default, and many places require you to say so at the start; some require everyone on the call to agree. A short line in the opening, such as "this call is recorded so we can check it", covers the common requirement. Check the rules where the person you're calling lives.

What should happen when someone asks for a person?

The AI should stop and either transfer the call live or promise a callback within a stated time, then make sure that callback happens. Test it before launch. Transfers can carry their own per-minute charge on some platforms, so include them in your cost estimate, and staff the line at the times the calls go out.

Further reads

• What Does an AI SDR Cost, and What Does It Actually Do? — What AI sales-development tools cost and actually do.
• Can AI Answer Calls and Qualify Leads for an Insurance Agency? — Another sector weighing AI on the phone, with its own rules.
• What to Check in an AI Vendor's Data Processing Agreement — What to check in a voice platform's data terms.
• What to Ask an AI Chatbot Vendor Before You Sign Up — Vendor questions that apply to voice agents too.
• How to Read AI Software Pricing: Seats, Credits, and Usage Fees — Per-minute, platform and credit pricing decoded.
• How Dental Practices Use AI for Recalls and Appointment Reminders — How reminder calls and messages work in a busy practice.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Retell AI, Vapi, Bland AI and Synthflow pricing pages (checked 29 September 2026); EU AI Act Article 50 transparency obligations; a national telecoms regulator's February 2024 declaratory ruling on AI-generated voices; Upwork hourly rates guide (2026).