AI voice agent for businesses: guide 2026

Many companies still miss calls after hours, during peak times, or when reception is busy. AI voice agents can now handle those calls with a smooth conversation, rather than the classic “press 1” menu. The question for an executive is no longer whether the technology works, but which calls make sense to automate, how much it costs, and what rules must be followed.

This guide answers those questions using information from official documentation and regulations in force in Mexico and the United States, reviewed in November 2026.

In this article:

• What an AI voice agent is and how it works
• Which calls to automate and which not to
• How much it costs per minute and what else you pay for
• Which rules apply in Mexico and the United States
• How to launch a pilot step by step

# What is an AI voice agent?

It’s a system that handles an entire phone conversation: it listens, understands the intent, decides what to do, and responds by speaking. Unlike a phone system with a menu, the person can speak in their own words, and the agent can look up or record information during the call: check the calendar, look up an order, or create a ticket.

There are two common architectures:

• Chained: One model converts speech to text, a language model decides how to respond, and another model converts the response to speech. It’s easier to audit because each step produces text.
• Speech-to-speech: A single model receives audio and responds with audio in real time. For example, xAI’s voice API offers speech-to-speech agents over WebSocket, with turn detection and tool use, and publishes examples of telephony integrations such as Twilio (xAI, Voice) comes in.

If you’re already familiar with text agents, the principle is the same as the one we explain in when to use an AI agent: value comes from an agent’s ability to take action in your systems, not just chat.

# Why does latency matter so much on a call?

In a conversation between people, there’s very little silence between when one person finishes speaking and the other responds. A study published in PNAS that analyzed ten languages found that the most common response comes between 0 and 200 milliseconds after a question, with an average of 208 milliseconds (Stivers et al., PNAS) comes in.

That’s why an agent that takes two or three seconds to respond feels clumsy, even if its answer is correct, and the person tends to repeat themselves or hang up. When evaluating providers, ask for a demo over a real phone call, not in a browser, and measure how long it takes to respond and whether you can interrupt it.

# Which calls make sense to automate with a voice agent?

The best candidates are frequent calls with answers stored in a system and a clear outcome. These are the use cases that make the most sense to us:

It’s best to leave sensitive complaints, negotiations, cases involving legal or health risks, and any call where the customer asks to speak with someone to a person. A good design always includes a quick handoff to a human, with context about what has already been discussed.

If your customers message more often than they call, start with the written channel; in the AI agent guide for WhatsApp we explain that use case. Many companies end up using the same agent on both channels.

# How much does an AI voice agent cost?

There are four cost components. The voice model is usually the least expensive:

• Voice model: for example, xAI lists its speech-to-speech model at 0.08 dollars per minute, speech-to-text conversion starting at 0.10 dollars per hour, and text-to-speech at 15 dollars per million characters (xAI, pricing). At that price, a 3-minute call costs 0.24 dollars for the model alone.
• Telephony: phone number, incoming and outgoing minutes, and, if applicable, the connection to your switchboard. This is paid separately to the carrier or phone platform.
• Integrations: connecting a calendar, CRM, order system, or ticketing system is what turns the conversation into a completed task.
• Design and operations: script, knowledge base, testing with real calls, monitoring, and monthly adjustments.

Model prices change frequently and are listed in dollars before taxes, so confirm them before preparing a budget. To estimate the full project, see how much it costs to implement AI in a mid-sized company.

# What rules apply to a voice agent in Mexico?

This is general information, not legal advice. Three points matter in nearly every project:

• Personal data: a recorded or transcribed call contains personal data, so the Federal Law on the Protection of Personal Data Held by Private Partiesapplies. At the start of the call, explain that it is being recorded and why, make your privacy notice available, and define how long you keep audio and transcripts.
• Sales calls: Article 18 bis of the Federal Consumer Protection Law prohibits sending advertising to people who have asked not to receive it or are registered with Profeco’s registry (REPEP). Profeco indicates fines of 411.61 to 1,317,141.34 pesos for making advertising calls to a registered number, and the provider is liable even if a third party makes the call (Profeco, REPEP) comes in.
• Transparency: even if there is no specific rule for AI, telling people they’re speaking with a virtual assistant helps prevent complaints and build trust.

# What if your company calls customers in the United States?

In February 2024, the Federal Communications Commission (FCC) confirmed that AI-generated voices, including voice cloning, count as artificial voices under the Telephone Consumer Protection Act (TCPA). In practice, an outgoing call using an AI voice requires the recipient’s prior express consent, except in emergencies or other exempt situations. For sales calls, consent must be in writing (FCC). Incoming calls—when the customer places the call—are a different case, but it’s still a good idea to identify yourself clearly.

# How do you launch a voice agent pilot?

A well-scoped pilot can tell you in a few weeks whether it’s worth scaling:

1. Listen to or review a sample of real calls and group them by reason; choose one or two common, straightforward reasons.
2. Define success before you start: calls resolved without human assistance, wait time, transfers, and satisfaction.
3. Write the script and knowledge base with the person who answers the phone today, and include phrases for when the agent doesn’t know the answer.
4. Connect only the systems needed for that reason, with minimum necessary permissions.
5. Test it with internal calls, then with a percentage of traffic or only after hours.
6. Review transcripts every week and correct responses; analyzing calls with AI and human review can help, as we explain in AI for analyzing sales calls.

Moving from a successful pilot to a stable operation comes with its own challenges; we cover them in from AI pilot to production.

# Frequently asked questions

# Does an AI voice agent replace my receptionist?

Usually not. It works best for handling repetitive calls, peak periods, and after-hours calls, so your team can focus on cases that require judgment. It should always be able to transfer the call to a person.

# How much does an AI voice agent cost per minute?

The voice model can cost a few cents of a dollar per minute; for example, xAI lists 0.08 dollars per minute for voice-to-voice. Phone service, integrations, and operations add to the cost and often make up a larger share of the total.

# Does the voice agent understand Mexican Spanish?

Current models support multiple languages, including Spanish. Before signing up, test it with real calls from your customers, including their vocabulary, proper names, and background noise.

# Can I use a voice agent for sales calls?

Yes, but in Mexico you can’t make advertising calls to numbers listed in REPEP without express consent. In the United States, calls using an AI voice require prior consent, in writing if they’re sales calls.

# Do I have to notify callers that the call is being recorded?

If you record or transcribe calls, you handle personal data. It’s best to disclose this at the start of the call and have your privacy notice available in accordance with the LFPDPPP.

# Conclusion: start with the calls you’re missing today

• An AI voice agent handles frequent, structured calls, with a handoff to a person.
• Latency shapes the experience: test it over a real phone call.
• The voice model costs cents per minute; telephony and integrations cost more.
• In Mexico, protect personal data and comply with REPEP; in the United States, obtain consent under the TCPA.
• Start with one or two call reasons and clear metrics.

If you want to automate your company’s phone support, learn about our AI voice agents and AI call center service.

# About the author

Arturo Sánchez Gándara is CEO of ai¹, an applied AI consultancy based in Guadalajara that helps companies in Mexico and the United States assess, implement, and operate AI solutions, from automation and voice and WhatsApp agents to computer vision.

# Referenced sources

• xAI Docs: Voice APIs
• xAI Docs: Models and pricing
• Stivers et al. (2009), Universals and Cultural Variation in Turn-Taking in Conversation, PNAS
• FCC: The TCPA applies to AI-generated voices (FCC-24-17)
• Chamber of Deputies: Federal Consumer Protection Law
• Profeco: REPEP, penalties
• Chamber of Deputies: LFPDPPP in force