Blog

AI Avatar on Your Website: What We Tested and the Fine Print

We tested an AI avatar that talks by video on your website. What works, what the contract hides, and the questions to ask before you sign up.

November 22, 2026 · Agência Primeira Página

AI Avatar on Your Website: What We Tested and the Fine Print

This week, our agency tested an artificial intelligence avatar platform: a realistic video face that talks by voice with website visitors, answers questions, and shows information on screen. The demo is impressive. But before putting something like this on a client's website, we read the entire technical documentation and the commercial use agreement. In this article, we explain what works, what doesn't, and the fine print that changes the decision.

What a conversational AI avatar is

The market uses the term "AI avatar" for three different things, and it's worth separating them before going further:

  • Recorded avatar video: you write a script and the tool generates a video of a digital person speaking. It's great for product videos and training, but it doesn't talk with anyone.
  • Text chatbot: a written conversation, on the website or on WhatsApp, which we already covered in detail in the post on how much an AI chatbot costs.
  • Conversational avatar: a live video face that listens to the visitor through the microphone, understands, answers out loud, and moves its face as it speaks. This is what this article is about.

Under the hood, the conversational avatar combines four components: a speech recognizer, which turns the visitor's speech into text; a language model, which decides what to answer; a synthetic voice; and a video generator, which animates the face in sync with the speech. All of this happens in under a second, which is why it feels like a conversation rather than question and answer.

What we tested

The platform we tested was Nuance Live, from Nuance Labs, which in November 2026 was in a preview phase, open to developers. We created an account, talked with the avatar in the platform's own test environment, and read the documentation and terms of use. Here is what we found:

  • The avatar stays on your website. The visitor is not sent to another address. The video window is designed by whoever builds the site, with the brand's colors and logo. The connection to the platform's servers happens behind the scenes, like an embedded video.
  • There are few faces to choose from. The documentation lists three ready-made personas, with an adjustment for how much the face moves. There is no way to create your own face or voice.
  • It can do things, not just talk. You can give the avatar functions: show a card on screen with a price, look up information in a system, register a contact request.
  • It requires programming. It's not a code snippet to paste. The access key cannot be exposed on the page, so the site needs a small server-side component that authorizes each conversation with a one-time password.

The fine print

The demo shows what the technology does. The documentation and the contract show what you can do with it. These were the points that changed our assessment.

1. The language

The documentation is direct: the avatar speaks only English, and the speech recognizer understands only English. Asking it to speak another language is refused by the platform itself. For a Portuguese-language website, that ends the conversation. For the English version of a site, it works.

This is the first question to ask any avatar vendor, and the answer needs to be tested, not read in the sales material. Speaking Portuguese and understanding Portuguese with accents, slang, and background noise are different things.

2. "Training" doesn't mean uploading documents

Many people imagine they will upload the catalog, the price list, and the entire website, and the avatar will "learn." On the platform we tested, there is no document upload. What exists is an instructions field, with up to roughly 65,000 characters, which the avatar reads at the start of every conversation: who it is, its tone, what it can and cannot say, the services, the prices, and the frequently asked questions.

A lot fits in a text of that size, but not an entire blog or a large catalog. For that, you need to give it a search function that looks up the data in real time. And the text doesn't learn on its own: when a price changes, someone has to change the text.

3. It records the visitor's face and voice

The contract defines as conversation data any image, face, voice, and expression captured during the interaction, and states that all interactions may be recorded, logged, and used by the company. If the visitor turns on the camera, their face goes to the platform's servers.

In Brazil, the General Data Protection Law (LGPD) treats biometric data linked to a person as sensitive personal data, which requires a stronger legal basis, usually specific and prominent consent. Another detail: the contract lists the data protection laws of the United States, the United Kingdom, the European Union, and Canada as applicable. The LGPD does not appear. Whoever puts the avatar on the website bears full responsibility for the notice, the consent, and the record of that consent, which the contract requires you to keep and hand over to the platform on request.

The same contract prohibits use by residents of some US states and of the Canadian province of Quebec, places with specific rules on biometrics and privacy. It's a sign that the company itself knows the terrain is sensitive.

4. You must disclose that it's an AI

The contract requires that the visitor know they are talking to an artificial intelligence, not a person, and prohibits making the avatar claim to be human. In Europe, this is already law: since August 2026, Article 50 of the European AI Act requires systems that talk with people to disclose that they are AI, as we explained in the post on the European AI law and chatbots. In Brazil, the bill for the AI legal framework goes in the same direction.

The contract also prohibits making the avatar impersonate a real person, even with that person's permission. The idea of an avatar with the name and manner of the company owner, for example, is not allowed.

5. Minors

Use involving anyone under 18 is prohibited. For a school, a hobby course, or a toy store, that alone rules out the tool.

6. It can be shut down tomorrow

During the preview phase, the contract allows the company to change, limit, suspend, or terminate the service at any time, without notice and without liability. There is no guarantee that it will work. For a test, that's fine. For a company's main customer service channel, it's a risk that needs a plan B.

7. Agencies can't resell

The contract prohibits reselling, sublicensing, or making the service available to third parties outside the application of whoever signed up for it, and prohibits sharing the access key. In practice, an agency cannot have one account and serve multiple clients with it. Each company needs its own account, or written authorization from the platform.

8. Cost and abuse

The conversation consumes credits from start to finish, rounded up, with limits on simultaneous and hourly conversations depending on the plan. This has two consequences. First: without a safeguard, anyone can open conversations back to back and burn the company's credits. Second: the avatar needs a clear purpose, because a long conversation with no goal costs money and doesn't sell.

Avatar or text chatbot: when each is worth it

A video avatar draws attention like nothing else on a website. But attention is not the same as results. From our experience with chatbots, an avatar makes sense when:

  • Presence is the product. Reception at an event kiosk, a demo at a trade show booth, a welcome in a virtual tour, where voice conversation is part of the experience.
  • The audience speaks the avatar's language. Today, in practice, that means websites and events for a foreign audience.
  • The conversation is short and has a goal. Presenting a product, answering three questions, and leading to contact.

The text chatbot remains the best choice when customer service is high-volume, when the customer prefers to write, when the conversation happens on WhatsApp, and when the cost per interaction matters. It also keeps a written history, which makes it easier to check what was promised to the customer.

Ten questions to ask before hiring an avatar vendor

  1. Does it speak and understand Brazilian Portuguese? Can I test it with accents and noise?
  2. How do I add the company's information: document upload, an instructions text, or search in my systems?
  3. Does the conversation stay on my website, or is the customer taken to another address?
  4. Does it record the visitor's face and voice? For how long and for what purpose?
  5. Is the data used to train the platform's models?
  6. Does the contract mention the LGPD? Who is responsible for consent?
  7. How is the visitor told that they are talking to an AI?
  8. How much does each minute of conversation cost, and how do I prevent someone from burning my credits?
  9. Does the service come with a guarantee of operation, or can it be suspended without notice?
  10. Can I use my own face and voice, or only the platform's?

If the vendor can't answer half of these questions, the technology may well be good, but the project is not yet ready to go live.

How we would do it on a website

If the answers are good, the path we follow in AI projects is always the same: start small and measure. First, define the avatar's goal, for example getting the visitor to request a quote. Then, write the instructions with the services, the prices, and the customers' real questions. Next, build the window on the website with the AI notice, the consent request before turning on the camera and microphone, and the usage safeguard. Finally, launch it on a single page and compare contact requests before and after.

If the goal is day-to-day customer service, the best start is still a well-trained AI chatbot. And if the idea is to bring AI to other parts of the company, our team handles AI implementation for businesses, from diagnosis to a working system.

Frequently asked questions

What is a conversational AI avatar?

It’s a digital face on video that talks live with the visitor: it listens through the microphone, understands the question, answers by voice and moves its face in sync with the speech. It’s different from a pre-recorded avatar video, which doesn’t hold a conversation, and from a text chatbot.

Can I put an AI avatar on my website?

Yes, but it usually requires programming. The avatar window sits on your site, with your own visual identity, and the site needs a small server-side component that stores the access key and authorizes each conversation. It’s not usually a ready-made snippet you can just copy and paste.

Can I train the avatar with my company’s information?

On the platform we tested, there’s no document upload. Your information goes into a block of instructions, up to roughly 65,000 characters, which the avatar reads at the start of each conversation. For large catalogs, you need to give it a function that looks up the data in your systems.

Does the AI avatar speak Portuguese?

It depends on the platform. The one we tested, in November 2026, only spoke and understood English. Before signing up for any avatar, test it in Brazilian Portuguese, with accents and background noise, and don’t rely on the sales material alone.

Does an AI avatar need the visitor’s consent?

If it captures face or voice, yes. In Brazil, biometric data tied to a person is sensitive data under the LGPD, which normally requires specific, prominent consent. You also need to make it clear that the conversation is with an artificial intelligence.

Is an AI avatar better than a text chatbot?

An avatar is worth it when presence is part of the experience, such as event kiosks, trade show booths and demos, with short, goal-oriented conversations. For high-volume support, WhatsApp and a low cost per interaction, the text chatbot is still the better choice.