On this pageJump to a section
The demo is the easy part.
You can now hold a natural conversation with an AI over a phone line, interruptions and all. It sounds like the answer to every appointment-setting headache: instant callbacks, no missed leads, no hold music. It also sounds like a way to put your business on the wrong side of a federal law in a single afternoon.
Nobody answers my phone but me these days, so I understand the pull of something that does it for you. Read this as education, not legal advice. Here is what changed two weeks ago, what the FCC said eight months ago and how I would think about the order of operations.
What changed on October 1, 2024?
Voice agents stopped being a stitched-together science project.
OpenAI's announcement says developers can now build fast speech-to-speech experiences, with the Realtime API in public beta for all paid developers.1 Before, a voice assistant had to transcribe audio, send text to a model and then convert the answer to speech, which often lost emotion and added delay. The Realtime API streams audio in and out directly over a persistent connection to GPT-4o, handles interruptions and supports function calling so the assistant can take actions, such as retrieving customer information.1
What it costs
OpenAI put the price at about $0.06 per minute of audio input and $0.24 per minute of audio output.1 That is cheap next to a full-time receptionist, and it is only part of the bill. Phone service, scheduling software, monitoring and review time all add to it.
It is a beta
The announcement frames it as a public beta, and the company points to early use cases such as language learning and coaching apps.1 Beta means rough edges. Do not hand your front door to a beta product without a human watching.
What did the FCC decide in February?
That an AI voice is an artificial voice.
On February 8, 2024, the FCC announced the unanimous adoption of a Declaratory Ruling that recognizes calls made with AI-generated voices as "artificial" under the Telephone Consumer Protection Act, effective immediately.2 The agency said the ruling ensures AI-generated voices in calls are held to the same standards as other robocalls, and noted that under FCC rules telemarketers must obtain prior express written consent from consumers before robocalling them.2
Why that matters for appointment setting
A friendly call from a voice agent to a lead who filled in a form can still be a telemarketing call, depending on content and context. Whether yours qualifies, and what kind of consent it needs, is a legal determination. The practical point is simple: an AI voice is covered by the same robocall rules. Do not assume a natural-sounding voice is exempt because it sounds like a person.
What does consent have to look like before the demo?
Clear, specific and recorded.
- Ask for permission in plain words on the form (the same habits I describe for intake consent): that the person agrees to be contacted by phone, including by an automated or AI-generated voice, at the number provided.
- Keep it separate. Do not bury it inside terms of service or bundle it with something unrelated.
- Make it unchecked by default. The person acts to give consent.
- Record the evidence: the wording shown, the time, the number and the source.
- Honor opt-outs immediately and make them easy, including during the call.
- Test with your own number first. Walk through the form and the call yourself.
Why I say consent before demo
Because consent comes before the demo for a reason: the demo is the moment you fall in love with the technology. If you build the agent first, you will be tempted to find a use for it. If you set the consent and the use case first, the agent has to fit inside the rules, which is the correct direction.
What jobs suit a voice agent in appointment setting?
Narrow ones with clear edges.
| Good fits | Poor fits |
|---|---|
| Confirming or rescheduling an appointment | Selling or persuading on price |
| Answering a fixed set of questions about hours, location and process | Anything involving clinical, legal or financial advice |
| Collecting a callback preference | Handling an upset or confused customer alone |
| Handing off to a human with a clean summary | Making promises about results |
| Following up with people who asked to be called | Cold-calling people who never opted in |
Keep the agent in its lane
Confirming and rescheduling is where an agent helps most, because a clean reminder timeline is what lifts show rate. The most dangerous thing an agent can do is improvise. Give it a script of allowed topics, make it say "let me get a person" when it reaches an edge and design the handoff before you design the voice.
- Lead opts inConsent captured and recorded first
- Agent opensSays plainly that it is an AI
- Allowed topicConfirm, reschedule, answer a fixed question
- Edge reachedAgent says it will get a person
- Human takes overClean summary, no repeated questions
What can go wrong beyond the law?
A lot, and most of it is boring.
- Wrong time slots. Without live access to your calendar, a confident agent books confident mistakes.
- Misheard names and numbers. Voice recognition is not perfect. Confirm details back to the caller.
- Sensitive information. If your callers mention health or financial details, decide in advance what is recorded, stored and shared.
- Overclaiming. Do not advertise the agent with promises it cannot keep. On September 25, the FTC announced Operation AI Comply (I touched on it in my o1 piece) against companies that used AI hype or sold AI tools in deceptive ways, with the Chair stressing that AI gets no pass from existing law.3
Disclose the AI
OpenAI's usage rules require developers to make it clear to users that they are interacting with AI unless it is obvious from the context.1
make it clear to their users that they are interacting with AI
I would go further than the minimum. Say it in the first sentence of the call. People forgive a robot that is honest. They do not forgive one that pretends.
How would I pilot one safely?
Small, supervised and reversible.
- Start with inbound or existing-customer use cases where the person expects to hear from you.
- Limit the agent to confirming and rescheduling, with a human handoff.
- Listen to every call for the first weeks.
- Track bookings made, errors and handoffs.
- Expand only when the errors are rare and understood.
Automate the part of the call that is a form. Keep a human for the part that is a relationship.
Voice agents will get better quickly, and the rules will not disappear. The operators who benefit will be the ones who built the consent and the guardrails first. Thinking about voice, inbox or chat automation for your business? Start the conversation at ekaistone.com.
Keep building.Ekai
Questions people ask
Are AI voice calls legal for appointment setting?
They can be, with the right consent and design. The FCC ruled on February 8, 2024 that calls made with AI-generated voices are artificial under the TCPA, so telemarketing-style calls generally need prior express written consent. Get legal advice for your specific use.
What is the OpenAI Realtime API?
It is an API released in public beta on October 1, 2024 that lets developers build low-latency speech-to-speech experiences using GPT-4o, with function calling so a voice assistant can trigger actions like looking up information.
How much does an AI voice agent cost to run?
OpenAI priced Realtime API audio at approximately $0.06 per minute of audio input and $0.24 per minute of audio output at launch. Your total cost will also include telephony, software and oversight.
Do I have to tell people they are talking to an AI?
OpenAI's policies require developers to make it clear to users that they are interacting with AI unless it is obvious from context. Beyond that, disclosure rules vary, so ask your attorney. Being upfront is also good practice.
Sources
- 1
- 2
- 3
