Digital Humans and AI Avatars
Building synthetic presenters that hold up on screen — with likeness consent documented, disclosure never optional, and an honest conversation about whether your use case needs a face at all.
A digital human is the most conspicuous thing on this list and frequently the least necessary. Before building one we ask what the face is for, because a great many avatar projects would work better as a voice, a well-designed interface, or a real person filmed once.
Digital humans and AI avatars are synthetic on-screen presenters, ranging from a video-generated presenter reading a script to a real-time animated character that responds interactively. They may be entirely invented or based on a real person with that person's documented consent.
Two decisions before anything is built
Does the use case need a face?
A face adds presence, and it also adds cost, uncanny-valley risk, consent obligations and disclosure duties. Training content, product explainers and multilingual announcements often work as well or better with a voice and clear visuals. We will say when the avatar is decoration.
Invented person, or a real one?
An invented presenter avoids consent and revocation questions entirely but has no existing recognition. A real person's likeness carries recognition and brings documented consent, compensation, revocation and the risk that the relationship ends while the avatar remains in circulation.
Likeness consent works exactly like voice consent
Informed, specific and documented; identity verified; prohibited uses named by the person; a revocation process that works operationally; and compensation terms agreed in advance. We do not build an avatar of a real identifiable person without it — including public figures, and including people who are no longer living, absent an estate's documented permission.
Getting it to hold up on screen
Choose the fidelity deliberately
Photorealism invites the uncanny valley, where small imperfections in eyes, micro-expression and lip timing read as unsettling rather than merely imperfect. A deliberately stylised presenter can be more watchable and more honest about what it is, and it is frequently the better choice.
Lip sync across every language you ship in
Sync that works in English and drifts in German or Japanese is the most common quality failure. Every target language is checked by a native speaker, because a viewer notices bad sync long before they can articulate why the video feels wrong.
Get the non-verbal behaviour right
Blink rate, micro-movements, gaze, posture shifts and gesture timing. Their absence is what makes a technically accurate avatar feel dead, and it is where most of the polish budget goes.
Design for real-time separately
An interactive avatar has a latency budget a rendered video does not — response time, generation and animation all inside a conversational rhythm. Where that budget cannot be met, a lower-fidelity presenter that responds promptly beats a beautiful one that pauses.
Decide the personality boundaries
What the avatar will and will not discuss, how it handles hostility, and what it does when it does not know. An avatar carries more implied authority than text, so its refusals need designing. See agent interface design.
Disclose, visibly and durably
Viewers are told they are watching a synthetic presenter, in a way that survives the clip being reposted without context. Beyond the EU AI Act's transparency and marking duties, undisclosed synthetic presenters damage trust badly when discovered. See EU AI Act compliance.
What we decline
- An avatar of a real identifiable person without documented consent, including public figures and deceased people absent an estate's permission.
- Any presenter designed to be mistaken for a real person in a context where the viewer would reasonably believe they were watching a real recording.
- Avatars delivering content the person whose likeness it is has refused, or that falls inside their named prohibited uses.
- Synthetic testimonials, endorsements or customer stories presented as real people.
- Presenters used to give medical, legal or financial advice without the qualified human accountability those domains require.
How the engagement runs
Whether the use case needs a face is settled before anything is built.
Use case and consent
Whether an avatar is warranted; invented or real likeness decided; consent, prohibited uses and revocation documented.
Design and capture
Fidelity chosen, appearance designed or capture completed, voice paired with the presenter.
Build
Rendering or real-time pipeline, lip sync across target languages, non-verbal behaviour and personality boundaries.
Testing
Native-speaker review per language; viewer testing for uncanny-valley response and comprehension.
Handover
Production workflow, disclosure implementation, consent register and revocation process handed over.
What you receive
A presenter that holds up on screen, disclosed as synthetic, with consent that would survive scrutiny.
Use case assessment
Whether a face is warranted, with the honest alternative named where it is not.
Consent framework
Documented likeness permission, prohibited uses, revocation and compensation, where a real person is involved.
The avatar
Designed at a deliberate fidelity, with voice paired and non-verbal behaviour tuned.
Multilingual lip sync
Checked by native speakers in every language you ship.
Personality boundaries
What it discusses, how it refuses, how it handles hostility.
Disclosure implementation
Visible and durable enough to survive the clip being reposted.
Is this the right engagement?
Worth being direct. Digital Humans and AI Avatars is the wrong spend in some situations, and those are listed rather than buried.
Good fit if
- Presenter-led content is needed at a volume filming cannot sustain.
- The same content must be presented in many languages.
- A consistent on-screen presence is wanted across years of content.
- An interactive presenter would genuinely improve a customer experience.
- Likeness consent and revocation need a defensible framework.
Choose something else if
- The content works as voice and visuals, which is often the case.
- You want a presenter viewers would mistake for a real person.
- The likeness belongs to someone who has not consented.
- The requirement is a text or voice assistant. See conversational AI.
Frequently asked questions
Marked up with FAQPage schema so these answers can surface directly in search results and inside AI assistant responses.
Do we need an avatar at all?
Frequently not, and we ask before building. A face adds presence and also cost, uncanny-valley risk, consent obligations and disclosure duties. Training content, explainers and multilingual announcements often work as well with a voice and clear visuals, and we will say when the avatar is decoration.
Can we use a real employee or executive as the avatar?
With their informed, specific, documented consent — covering what it may be used for, for how long, with named prohibited uses and a working revocation process. It is worth considering in advance what happens if they leave the organisation while the avatar is still in circulation.
Do viewers have to be told it is synthetic?
Yes, and we implement disclosure so it survives the clip being reposted without context. Beyond the EU AI Act's transparency and marking duties, an undisclosed synthetic presenter damages trust severely when discovered — and it will be discovered.
How realistic can they be?
Convincing in rendered video, particularly at moderate length. Real-time interaction is harder because the latency budget constrains fidelity. Photorealism is not automatically the goal — a deliberately stylised presenter avoids the uncanny valley and is often more watchable and more honest.
What about multiple languages?
That is one of the strongest cases for an avatar, and lip sync is where it usually fails. Sync that works in English drifts in other languages, so every target language is reviewed by a native speaker — viewers notice poor sync long before they can say why the video feels wrong.
Often paired with this
Most clients combine two or three engagements from the Creative and Media AI pillar. These are the ones that most often run immediately before or after.
AI Voice Cloning and Dubbing
Synthetic voice and dubbing on documented, revocable consent — with verification, watermarking and a refusal list.
Read more →AI Video Generation and Editing
Editing, versioning and localisation at scale, plus short generated sequences with honest limits stated.
Read more →3D and Synthetic Media Generation
AI-assisted 3D assets with production-ready topology, and synthetic training data with domain gap measured.
Read more →Is this the right engagement?
Tell us what you are trying to build. If a different service fits better, or if you do not need us at all, we will say so.