Learn
Live AI agents vs. AI video: which do you need?
Recorded AI video, streamed avatars and live agents rendered in the browser all put a face on screen. They solve different problems. Here is how to tell which one you need.
"AI avatar" has become a catch-all term. It covers a presenter reading a script in a training video, a streamed character on a support page, and an agent that answers questions out loud while it scrolls the page for you. They can look similar in a screenshot. They are very different tools.
This guide separates them into three categories, explains what each is good at, and offers a simple way to decide. We make live agents, so we have a point of view, but recorded AI video is the right answer for plenty of jobs and we will say so.
1. Recorded AI video
AI video generators turn a script into a finished video of a presenter speaking. You write the words, pick a face and a voice, and get a file you can publish anywhere.
What it is good at:
- One-way content: product explainers, onboarding modules, announcements, internal updates.
- Localisation: the same video in many languages without re-shooting.
- Distribution: a video file plays in a learning platform, on social media or in an email, with no live service behind it.
- Polish: you can review and approve every second before anyone sees it.
Where it falls short: it cannot listen. The viewer cannot ask a question, change the topic, or get an answer about their own situation. Every viewer sees the same thing.
2. Streamed cloud-rendered avatars
Streamed video avatars are interactive. A model generates the face on a server and streams it to the viewer as live video, much like a video call.
What it is good at: it can respond in real time, and the look can be very photographic because rendering happens on powerful hardware.
Where it falls short:
- Every second on screen is video being generated and streamed. That tends to make idle time costly, so these avatars often wait behind a "start" button rather than being present on the page.
- Quality depends on the viewer's connection. A weak network means a blurry or stuttering face.
- Start-up can involve a wait while a session is allocated.
3. Live agents rendered in the browser
This is the approach Somebody takes. The conversation runs live, but the face is rendered on the visitor's own device, in the browser, rather than streamed as video.
What it is good at:
- Presence without cost. Because no video is being streamed, an agent can wait on a page without using minutes. Minutes are only metered during an active conversation.
- Crisp on ordinary connections. Once the face has loaded, only audio and small amounts of data travel over the network, so the face stays sharp when the connection is not perfect.
- Doing, not just talking. Because it runs in the page, a website agent can navigate, read the page, scroll, highlight, fill and submit forms, click and search your site.
- Two-way by design. Visitors can interrupt, ask follow-ups and steer. In roleplay and interviews, that back-and-forth is the whole point.
Where it falls short: it relies on the viewer's device to draw the face, so a very old device will have a harder time. And it is not a video file. If you need something to drop into an email or upload to a video platform, this is the wrong tool.
A side-by-side view
- Can the viewer talk back? Recorded video: no. Streamed avatar: yes. Browser-rendered live agent: yes.
- Can it wait on a page? Recorded video: yes, it is just a file. Streamed avatar: usually behind a button, because idle streaming costs money. Live agent: yes, without using minutes.
- Can it take actions on your site? Recorded video: no. Streamed avatar: depends on the product. Live agent: yes, when built for it.
- Can you approve every word in advance? Recorded video: yes. Streamed avatar and live agent: you shape behaviour with instructions, knowledge and guardrails, but each conversation is new.
- Can it play offline or in a learning platform? Recorded video: yes. Live agents need a connection, though Somebody roleplays can be packaged for learning platforms with SCORM 1.2, cmi5 or xAPI export.
How to decide
Start with the direction of the conversation.
If the message only needs to go one way, make a video. If people need to answer back, you need an agent.
Choose recorded AI video when
- The content is the same for everyone.
- You need sign-off on every word, for legal or brand reasons.
- It has to live somewhere without a live service: an email, a social feed, an offline course.
Choose a live agent when
- Visitors have different questions and the right answer depends on who is asking.
- The value is in the exchange: practice, feedback, interviews, qualification.
- You want the agent to act: take someone to the right page, fill the form, build the cart.
- You want it present on the page all day without paying for the time nobody is talking.
Use both
Many teams will. A recorded video can explain a new product in ninety seconds; a live agent beside it can answer the questions the video raised. A training course can open with a video lesson and finish with a roleplay where the learner practises what they just watched.
Questions to ask any vendor
- Is the face generated live for each conversation, or played back?
- Where is it rendered, and what happens on a slow connection?
- Am I billed while the agent is idle on the page, or only during conversations?
- Can it take actions, or only talk?
- Can I use my own AI provider account if I need to?
If you want to see the browser-rendered approach for yourself, talk to our agent, or compare plans on the pricing page. For background on the category, read What is an embodied AI agent?