Can AI Agents Make Phone Calls? Voice Agents and the Rules

Yes, AI agents can make phone calls. A voice agent listens, decides what to say and speaks through a real phone line, and it can both answer calls and place them. Google Search already calls local businesses on its users’ behalf to ask about prices and availability. In the US the harder question is legal, not technical. In February 2024 the FCC confirmed that an AI-generated voice counts as an “artificial or prerecorded voice” under the Telephone Consumer Protection Act, so an outbound call that uses one needs the called party’s prior express consent, prior express written consent if it is telemarketing, and must identify who is calling. The Do Not Call registry, calling hours and call-recording laws apply as they would to any call. This is a plain-language summary, not legal advice.

How a voice agent makes a call

• A phone connection. The agent reaches the phone network through a telephony provider or a SIP trunk, which turns a phone number into an audio stream a program can use. OpenAI’s guide to connecting phone calls over SIP↗ (opens in a new tab) describes pointing a SIP trunking provider at its Realtime API, which fires a webhook for each incoming call so your code can accept it and configure the session.
• Listening and speaking. Either a chain of speech recognition, a language model and speech synthesis, or a model that takes speech in and gives speech out directly. The second is usually quicker to respond, which matters on a phone where a pause sounds like a dropped call.
• Tools. Mid-call, the agent looks things up and acts: checks a calendar, reads an order, books a slot. The same care applies as for any agent with tools, plus the fact that the other person hears every mistake in real time.
• A way out. A transfer to a person when the caller asks for one, when the agent is unsure, or when the call turns out to be about something it should not handle.

Answering calls vs placing them

An agent that answers your own business line, as a receptionist that books appointments or routes callers, is the gentler case. The person chose to call you. The FCC’s AI ruling is about calls a business initiates, so it does not stop you from answering your phone with an agent, though recording laws, honesty about what the caller is talking to, and a quick route to a person still matter.

An agent that places calls is where the rules bite, because the person on the other end did not ask to hear from it. Almost everything below is about outbound calls.

The FCC ruling on AI voices, in plain words

The FCC’s declaratory ruling of February 8, 2024↗ (opens in a new tab), adopted six days earlier, answers the question most voice-agent vendors had hoped was open. Its main points:

• The TCPA’s restrictions on “artificial or prerecorded voice” cover current AI technologies that generate human voices, including voice cloning, because a person is not speaking them.
• Callers need the prior express consent of the called party before making such a call, unless there is an emergency purpose or an exemption applies. If the call carries an advertisement or telemarketing, the rules require prior express written consent.
• There is no carve-out for technology that claims to be as good as a live agent. The ruling says the TCPA “does not allow for any carve out of technologies that purport to provide the equivalent of a live agent.”
• Artificial voice messages must identify the business or person responsible at the beginning of the message and give a phone number, and telemarketing messages must offer a way to opt out.
• In the statement accompanying the ruling, the FCC’s chairwoman noted that it also lets State Attorneys General go after the people behind these calls and seek damages.

Among its examples of harm, the ruling cites news coverage of the fake Biden robocall sent to New Hampshire voters in January 2024. Its logic, though, covers ordinary business calls just as much as scams.

The other rules that still apply

• Do Not Call and calling hours. The FCC’s telemarketing rules in 47 CFR 64.1200↗ (opens in a new tab) bar telephone solicitations to a residential number before 8 a.m. or after 9 p.m. local time at the called party’s location, and to numbers on the national Do Not Call registry.
• Disclosure that the voice is AI. In August 2024 the FCC proposed further rules↗ (opens in a new tab), including a requirement that callers using an AI-generated voice disclose it at the beginning of each call. That was a proposal; check the docket, CG Docket No. 23-362, for whether it has become a rule. Saying it anyway costs nothing.
• Recording. Voice agents usually record and transcribe. Federal law, in 18 U.S.C. 2511↗ (opens in a new tab), allows recording when one party to the call consents, but some states require every party’s consent, so say at the start that the call is recorded.
• State law generally. States have their own telemarketing and consumer protection laws. Google’s own help page for its business-calling feature says it works in every US state except five, without saying why.

An example already in use: calling businesses for you

Google Search now offers to call local businesses in the US on a user’s behalf to check availability and pricing, then reports back. Google’s help page for businesses↗ (opens in a new tab) that receive these calls says they are made to book appointments, check wait times and confirm prices, that Google never asks for payment details or passwords, and that a business can opt out by telling the caller “Please remove my business from your list” or through its Business Profile settings.

That design is a useful template for any outbound agent: a narrow purpose, a clear sponsor, no sensitive data requested, and an easy way to say no.

Calls to hand an agent, and calls to keep

• Good fits: answering your own line after hours, booking and rescheduling for people who called you, reminders to customers who agreed to receive them, and calls to businesses to ask about hours, stock or prices.
• Handle with care: payment reminders and anything about money owed, where the consent and identification rules above apply in full and a wrong word is costly.
• Keep for people: first sales calls to anyone who has not given written consent, anything medical, legal or financial beyond logistics, complaints, and any call where the other person is upset, confused or vulnerable.
• Never: a cloned voice of a real person without their permission, or an agent that claims to be human when asked.

For sales teams, the usual answer is that the agent prepares the list and the talking points and a person dials, as AI sales agents describes. Support lines have their own escalation rules, covered in AI agents for customer service.

Consent is a record, not a feeling

Before any agent places a call, you should be able to show, for that number, when and how the person agreed to be called this way, and that they have not withdrawn it. That record belongs in the system that holds your customer data, with an opt-out that takes effect immediately, not in a spreadsheet the agent keeps for itself. If you cannot produce it, the call is made by a person or not at all.

Where the follow-ups go

Calls produce work: a callback someone promised, a price the business quoted, a complaint that needs a decision. fenbs does not place or record calls. With it connected over MCP, a voice agent’s summary can become a task: the outcome and what was promised in the note, a priority from 1 to 10, and the kind (feature, enhancement or bug) when the call reported a problem with your product. A person moves it through To Do, Next Up, In Progress and Completed, and History records who changed what. Keep phone numbers and personal details out of the note; link to the record in your CRM instead.

A standing rule such as “assistants never place outbound calls” belongs on the Decisions and rules page, recorded by a person, so every connected assistant reads it before it starts. fenbs has no due dates, so a callback time goes in the title.