What is interactive AI avatar software, and how should you evaluate it for enterprise use?
Taipei · 12 November 2026 · by the Perxona team
Interactive AI avatar software puts a character with a face, a voice and a body in front of a customer and lets it hold a live conversation. It listens or reads, decides what to say with a language model and the knowledge you gave it, speaks the answer in a synthetic voice, and moves its face and body while it talks, all in real time. It is not the same as AI video tools, which render a clip in advance, and not the same as a text chatbot, which has no face. To evaluate it for enterprise use, check eight things: what the character can be, whose language model it uses and how it learns your knowledge, how fast it answers, how it moves, which languages and voices it speaks, where it runs, how it is secured and integrated, and what a month costs at your real volume. The checklist below gives the question to ask for each one and what a good answer looks like.
Disclosure first: Perxona is one of the platforms in this category, built by XRSPACE in Taipei. This post describes the category and the checklist before it describes Perxona, and where it mentions Perxona it says so. Every Perxona fact comes from perxona.ai and the Connect Kit handbook as read on 12 November 2026.
TWhat interactive AI avatar software is, and what it is not
Three kinds of software get called "AI avatar", and buyers mix them up. The difference is when the content is made and whether a face is involved.
Kind | What it produces | When the content is made | Typical use |
|---|---|---|---|
AI video avatar | A video clip of a presenter reading a script | Before the viewer arrives | Training videos, marketing clips, localized announcements |
Text chatbot | A typed reply in a chat window, sometimes read aloud | Live | Support widgets, FAQ bots |
Interactive AI avatar | A character that answers live, out loud, with a face and body that move while it speaks | Live, inside the conversation | Reception, concierge, exhibition guide, sales assistant, coach |
Only the third kind is interactive AI avatar software. The test is simple: can a visitor ask something the script did not anticipate and get a spoken answer from a face, now?
The six parts inside every interactive AI avatar: character, ears, brain, voice, motion, channel
The six parts inside every interactive AI avatar
Every interactive avatar, from any vendor, is built from the same six parts. Knowing them makes the evaluation faster, because each one becomes a line in the checklist.
The character. What the visitor sees: a photoreal clone of a real person, a stock human, or a designed character such as a 3D figure, a cartoon or a mascot the brand already owns.
The ears. Speech recognition, a text box, or both.
The brain. A language model, plus the knowledge you give it (documents, FAQs, product data) and the rules you set for its role.
The voice. Text-to-speech in one or more languages, standard or custom.
The motion. Lip sync, facial expression and body language driven by what is being said. This is the part that separates an avatar from a voice bot with a picture on it.
The channel. Where it appears: a website widget, a tablet or kiosk, an exhibition screen, a native app.
Vendors differ mainly in which parts they own and which they let you swap. That is the first thing to find out, and the checklist asks it part by part.
The enterprise evaluation checklist
Eight criteria. For each: the question to ask the vendor, and what a good answer looks like. Run every shortlisted platform through the same table, with the same questions, and the decision usually makes itself.
# | Criterion | Ask the vendor | What a good answer looks like |
|---|---|---|---|
1 | Character | Can the avatar be our character, or only a human face? Can we bring a model we already own? | A plain statement of what the avatar can be, and a documented way to import your own. For 3D characters the common file format is VRM. |
2 | Brain and knowledge | Which language model runs the conversation? Can we bring our own? How is our knowledge loaded, and how much fits? | Your choice of model on the plan you can afford, a knowledge limit stated in megabytes or documents, and a way to update it without a developer. |
3 | Response time | What is the measured time from the end of the visitor's question to the first spoken word, on our network? | A number from a live demo on your own connection, not from a marketing page. Anything that would feel like a pause on a phone call is too slow for a reception desk. |
4 | Motion and expression | What drives lip sync, facial expression and gesture? Can our application control the emotion? | Motion generated from the content of each reply rather than a loop, and an API or markup to set emotion line by line. |
5 | Voice and languages | Which text-to-speech providers, which languages, and can we bring our own voice? | Named providers, a language list you can read, and a bring-your-own-audio route if your brand already has a voice. |
6 | Channel and browsers | Where does it run, and on which browsers and devices, without plugins? | Website embed plus the kiosk, tablet or event screen you actually need, on current Chrome, Safari and Edge, desktop and mobile. |
7 | Security and integration | What is the key model? What reaches the browser? Is there an API, an SDK and documentation we can read before signing? | A split between a secret server-side key and a publishable browser key with domain restrictions, public documentation, and a sample you can run. Then your own questions: where data is processed, how long it is kept, a data processing agreement, single sign-on. |
8 | Cost at volume | What is the billing unit, what does one conversation cost in it, and what happens when we run out? | A unit you can model (minutes, credits or interactions), a published overage price, and a clear failure mode: metered calls stop, nothing silently degrades. |
Two more items are not features, but they decide most enterprise deals:
Time to first deployment. Ask how long until a working avatar is on one of your pages with your knowledge in it. Days is good. Weeks means you are buying a services project.
Operations after launch. Who watches the conversations, tunes the answers and reports on usage? Ask whether that is you, the vendor, or an extra plan.
The eight questions to put to every shortlisted vendor
How to run the evaluation in two weeks
Week one. Write down ten real questions your front-line staff hear every day, plus three they would escalate to a manager. Put the same thirteen questions to every vendor's demo, on your network, on a phone. Score the eight criteria from what you see, not from the slide deck.
Week two. Deploy the best one on a single page or a single screen with the real knowledge base. Count conversations, note the questions it missed, and have the staff who would live with it say whether they would keep it. Their answer is the evaluation.
Where Perxona sits in this category
Perxona is a real-time 3D AI avatar platform: a character with a face, a voice and a body that talks to your customers on your website, built by XRSPACE.
Against the checklist, as stated on perxona.ai and in the Connect Kit handbook on 12 November 2026:
Character. 3D characters, including realistic humans and cartoon characters. Bring your own avatar in VRM 1.0 through the Connect Kit's VRM uploader. A custom 3D character built from your brand mascot is offered through sales.
Brain and knowledge. A no-code builder with knowledge management and control by prompt. Developers can use any language model through the Connect Kit; the Express sample accepts any OpenAI-compatible endpoint. Knowledge and media: 5 MB on the free plan, 100 MB on the US$25 plan, enterprise amounts through sales.
Response time. Not published. Ask for a live demo on your network, which is the right test for any vendor on your list.
Motion and expression. Behavior AI drives the avatar's motion and expression. In the Connect Kit, motion cues travel inline in the reply text, so body language follows what is being said.
Voice and languages. Azure and Google text-to-speech, standard voices on self-serve plans, expressive voices through sales, and bring-your-own audio in the Connect Kit. Perxona does not publish a language count, so ask for the list you need.
Channel and browsers. Publish to a website on every plan. The handbook states that all features run on Chrome, Safari and Edge, desktop and mobile, with no platform workarounds.
Security and integration. Two keys: a secret Connect API key that stays on your server and a publishable key for the browser with domain restrictions. A public handbook at connect.perxona.ai, an open-source sample on GitHub, and conversation history on every plan. Data processing location, retention and data processing agreement terms are not on the public pages. Ask sales, as you should with every vendor.
Cost at volume. Credits. The free plan has 20 credits and one active avatar. The US25 a month plan has1 50credits. Extra credits are US0.095 each. Enterprise credits go through sales. When credits run out, metered API calls return an error rather than degrading.
Time to first deployment. The handbook's claim is a talking 3D avatar on screen in under 15 minutes with the Connect Kit. The no-code path for a single page is in the same range.
Operations after launch. An Enterprise Management option lists 10 to 30 monthly service hours of monitoring and tuning, plus monthly data reports.
Where Perxona is the wrong choice: if the brief is a photoreal clone of a named executive, or an avatar that lives inside a game engine, other vendors fit better. We listed them in our post on HeyGen LiveAvatar alternatives.
Frequently asked questions
What is the difference between an interactive AI avatar and an AI video avatar?
An AI video avatar is rendered in advance from a script and cannot answer a question. An interactive AI avatar answers live: a language model chooses the words and the face and body move as it speaks.
Does an interactive AI avatar replace a chatbot?
It usually sits on top of one. The language model and the knowledge base do the same job as in a chatbot. The avatar adds a voice, a face and body language, which is what makes it usable on a reception desk or an exhibition floor, where nobody wants to type.
How long does an enterprise deployment take?
A single-page or single-screen pilot with your own FAQ can be live in a day on self-serve platforms; Perxona's handbook says under 15 minutes for a first talking avatar with the Connect Kit. A multi-site rollout with custom characters, single sign-on and a data processing agreement is a project of weeks, and the checklist above is how to scope it.
Next step
Run the checklist. Copy the table into your evaluation sheet and send the "ask the vendor" column to every shortlisted platform.
Try the category. Perxona's free plan gives one active avatar and 20 credits, and the open-source Connect Kit on GitHub (XRSPACE-Inc/perxona-connect-kit) runs in minutes.
Talk to us. If you are evaluating for a reception, a venue or an event, send us the ten questions your staff hear most and we will show you the answers, spoken by a character.
© 2025-2026 XRSPACE CO., LTD. All rights reserved.
