AI Radio Hosts: What Coyotec Reveals About Radio

Why Coyotec is more than an AI novelty

It's tempting to call Coyotec an AI DJ. But that label misses what seems to matter most in the product decision: Entravision gave the character a specific audience, a clear cultural context, and a role alongside a human presenter. Coyotec isn't just a synthetic voice reading music facts. It's built to join a bilingual conversation with Guzman and speak to Latino listeners in a familiar register.

That distinction matters because audiences don't usually stick around just because something is new. They show up for what it does for them. Coyotec's topics run from the Dodgers and taco history to immigration, racism, and practical information about rights when ICE shows up. The mix is part of its identity - everyday culture right next to issues listeners may experience as immediate and personal.

Entravision's leadership says the idea came up when staff were unsure how to respond to immigration enforcement and protests. Human presenters, they argue, faced real professional and personal risk by speaking publicly. An AI character could address the subject without exposing a human host to the same consequences. That doesn't make the character neutral. Its topics and tone still reflect editorial choices made by the people behind it. But it does shift who carries some of the risk.

The character also shaped up through human feedback. Guzman says Coyotec initially sounded too poetic for the audience she had in mind. She worked on Spanish pronunciation and conversational style, including asking her parents to speak with it. In other words, cultural fit didn't magically emerge from the model. It was an editorial and production job that took people who knew what sounded right - and what didn't.

Coyotec also isn't being pitched as a replacement for the whole show. It cohosts with Guzman, responds to her in real time, and isn't allowed to take listener calls on air. That boundary makes the product easier to understand: the AI adds voice and perspective, while the human host anchors the relationship, exercises judgment, and controls what goes out.

The broader lesson is simple: an AI personality is a programming and editorial decision, not just a voice-generation feature. If the character has no real purpose, it becomes an expensive novelty. If it has a useful role and a credible human counterpart, it can help a station serve listeners in a different way.

How the AI host works in a real broadcast workflow

Coyotec is built from multiple technologies, not one all-purpose system. Entravision's executive vice president of AI strategy has described using Claude Haiku for the live agent's language-model and response generation, ElevenLabs for voice, OpenAI and ChatGPT for initial scriptwriting, and Midjourney to help develop the character's visual identity.

That's typical of a production setup made from specialist tools. One part helps generate language; another produces speech; others support scripts and visual development. Even then, the end result still depends on choices about what the character can discuss, how it sounds, when it speaks, and what a human host can override. The model is only one piece of the operational design.

In the studio, Coyotec can respond to Guzman in real time. That makes it more interactive than a prerecorded station identifier or a conventional syndicated segment - but it's still not an open, unrestricted participant. It can't yet answer listener calls on air. That matters, because a controlled exchange between two hosts carries different risks than an open line where listeners bring unexpected claims, personal crises, or sensitive information.

Guzman also functions as an editorial control. She describes shaping the character's language and says she stays in charge, even when Coyotec sometimes resists prompts or behaves in ways she didn't expect. That detail is telling. A strong persona can be entertaining, but it shouldn't be mistaken for independent judgment. The broadcaster remains responsible for what airs - including moments that feel spontaneous.

If a station is considering this kind of host, a basic workflow should spell out those responsibilities clearly:

• Define the role: Decide whether the AI is a cohost, a scripted voice, a music presenter, or a source of live information. Don't leave room for vague expectations that it will "just improve" programming.
• Set subject boundaries: Identify topics that need human review, especially legal, health, political, or safety-related information.
• Keep a human accountable: Name the person who can correct, pause, or remove the system from the broadcast.
• Test the actual listener experience: Check pronunciation, pacing, humor, interruptions, and errors with people from the target community.
• Plan for failure: Decide what happens if the model produces a harmful answer, the voice service fails, or the host gives listeners misleading information.

This isn't red tape for its own sake. Live systems fail in ways scripted demos usually don't. If an AI host is a public-facing product, stations need operational controls - because mistakes can damage trust as quickly as they can affect ratings.

What the audience figures do and do not prove

Coyotec's strongest business signal is that Entravision expanded the show. It started as a weekend program in September 2025, then moved into a weekday morning slot, and later syndicated in five additional markets across California, Nevada, and Texas. Entravision also reported that listenership grew during the first two months of syndication, with gains among Gen Z and millennial listeners.

For Los Angeles, the company's Nielsen figures say cumulative listenership among Hispanic listeners rose 21.6 percent from August to September 2026. It reported an 84.9 percent increase among listeners aged 18 to 34 over the same period, and a 500 percent increase in Hispanic female listeners in that age group. Those numbers are attention-grabbing - especially for a network that has traditionally skewed older.

But these are company-reported results from a short time window. They show that the audience shifted during the period Entravision highlighted. They don't, on their own, prove that Coyotec caused the shift. The weekday expansion, Guzman's work, the show's content, syndication, promotion, scheduling, and even measurement variability could all play a part. A "500 percent" jump can also start from a small base, so you still need the absolute audience size - not just the percentage.

For a broadcaster, the right question isn't only whether ratings went up. It's whether the new format created durable value once you account for its full cost and risk. That means tracking more than a headline number:

• Audience quality: Which groups are listening, and are they returning across multiple measurement periods?
• Retention: Do listeners stay through the AI segment, or do they tune out after the human part begins?
• Incremental reach: Does the character pull in people the station wasn't already reaching?
• Commercial value: Does the format improve ad demand or create a viable syndication product?
• Operational cost: How much human time is still required for writing, testing, editing, and oversight?
• Trust and complaints: Are listeners comfortable with the host, and do they understand it's AI?

The comparison also has to be fair. An AI segment should be judged against the real alternative: another human-hosted show, a syndicated program, music automation, or no additional programming at all. If a station can't afford a live overnight presenter, then the relevant baseline might be automated music - not an ideal scenario where everything is fully staffed.

That context matters to the business case. Radio was already automated and heavily reliant on prerecorded or syndicated content before generative AI. AI might make some workflows faster or cheaper, but efficiency isn't the same thing as audience value. A station should be able to show both: what labor or production constraint the system solves, and why listeners choose to stay with it.

How broadcasters should weigh the ethical and business trade-offs

Coyotec's role is tied to a sensitive subject: immigration enforcement and the safety of Latino communities. Supporters may see an AI host as a way to share information and express solidarity without putting a human presenter in the same danger. Critics may wonder whether a company is using a synthetic personality to take a stance while shielding its staff - or whether an automated voice can offer anything close to care when listeners are facing real risks.

You can't resolve those concerns by calling the system well-intentioned. What matters are clear editorial practices. Legal rights information, for example, needs to be accurate, up to date, and presented carefully. A conversational persona can make information more approachable, but it can also let uncertain or incorrect claims sound confident. For sensitive topics, a station should spell out who verifies the information, how corrections work, and what the host should do when it can't answer reliably.

There's also a workplace question. Broadcaster and voice-over artist Becca Guzmán said she believed Entravision stopped using her contractor services because of AI tools. The company said it had no record of laying her off and hadn't dismissed staff because of AI, while also saying changes to its workflow software may have reduced the need for contractor help. These accounts don't settle the disagreement. But they do show why workforce impact needs to be assessed through actual roles and contracts, not generic assurances that AI automatically creates new jobs.

A station should look closely at which tasks are automated, which roles change, and who pays the transition cost. Voice talent, producers, writers, translators, and editors may all contribute to a polished AI personality - even when the output sounds like a single synthetic host. If those contributions are hidden, the business can underestimate the labor needed and overstate how "independent" the system really is.

Then there's privacy and surveillance. Guzmán raised concern about connections between AI and surveillance of immigrant communities. Entravision's executives argued that their use is meant to entertain, inform, and connect, and that it isn't the same as using AI to track or target people. That distinction matters - but it has to show up in product choices: limit collection of listener data, explain what's recorded, avoid unnecessary identification, and don't imply a broadcast interaction is confidential if it isn't.

A practical decision test for media organizations can be four questions before launch:

1. Audience benefit: What specific listener need does the AI address better than the current format?
2. Human accountability: Who owns editorial decisions, corrections, and the choice to take it off air?
3. Workforce impact: Which human tasks disappear, change, or remain essential - and how will affected workers be treated?
4. Data and safety: What information does the system collect, and what safeguards apply to sensitive topics and communities?

The trade-off isn't simply "human vs. machine." An AI host can increase production capacity and make the show more distinctive. It can also shift risk onto listeners and workers if the system's limits aren't clear. A responsible rollout makes benefits measurable and costs visible.

Key takeaways

• Treat the AI host as a format, not a feature. A generated voice by itself doesn't justify keeping listeners around.
• Use humans for editorial judgment and audience understanding. Coyotec's Spanish and conversational style required ongoing human adjustment.
• Separate reported growth from proven causation. Short-term ratings changes are a reason to investigate, not proof that AI drove the result.
• Set the boundaries before going live. Topic limits, escalation paths, disclosures, and correction processes are part of the product.
• Count the full cost. Include human oversight, production labor, workforce impact, and trust - not only software fees.