AI avatars are now a button in the office suite. Decide whose face and voice your company may synthesize
AI avatars for training, marketing and sales video moved into everyday tools like Google Vids. The enterprise questions are consent, disclosure and impersonation, by type of avatar.
Listen to this article · 6 min
AI-generated narration of the full article.
The request usually arrives as a production problem. The compliance team needs its annual training video in twelve languages, the chief financial officer is the one who should present it, and the CFO has no time to record twelve versions. Someone suggests an AI avatar of the CFO. It would take an afternoon.
Two years ago that was a specialist purchase. Today it is a feature in the office suite. Google added AI avatars to Vids, its video tool inside Workspace, in August 2025, with twelve preset presenters and training, onboarding, policy announcements and product demos as the intended uses. In February 2026 they gained Portuguese, Spanish and five more languages. In April, users could create and refine custom avatars. Both of the later updates carry the same line in Google’s notes: “There is no admin control for this feature.” An administrator can switch Vids off for the organization; the avatar features themselves are not a separate setting.
That combination, easy creation and little central control, is why AI avatars have become an enterprise governance topic and not just a marketing one.
Not all avatars carry the same risk
The word covers four different things. They need different approvals.
| Whose likeness | Consent needed | Disclosure | Main risk | |
|---|---|---|---|---|
| Stock presenter | Nobody's, or a licensed actor's | Vendor license terms | Label as AI-generated | Low; brand fit |
| Designed brand character | A fictional face built for the company | Internal brand approval | Label as AI-generated | Looking like a real person by accident |
| Employee likeness | A real employee's face and voice | Written, specific, revocable | Label, and name the person | Use after the person leaves or objects |
| Executive likeness | A leader's face and voice | Written, specific, with a usage register | Label every use | Impersonation and payment fraud |
Stock and designed avatars are a content decision. The questions are brand fit, license terms and labeling.
An employee’s likeness is personal data, and the face and voice used to build it can be biometric data under several regimes. Consent should be written, specific about where and how long the likeness will be used, and revocable. Plan for the day the person leaves the company: a training library full of videos presented by a former employee who withdrew consent is an expensive cleanup.
An executive’s likeness adds a security problem. Every authorized video of the CFO speaking is also training material for someone who wants to fake one. In 2024, the engineering firm Arup confirmed that an employee in Hong Kong had been deceived into transferring about US$25 million after a video call in which the other participants were deepfakes. Companies that publish synthetic executive video should assume their staff will see more of it, real and fake, and change payment verification accordingly.
Disclosure is now a legal requirement in Europe
The EU AI Act’s transparency rules in Article 50, including the duty to disclose deepfakes, apply from 2 August 2026. The Digital Omnibus that delayed other parts of the Act in July did not move them; it only gave systems already on the market until 2 December 2026 to add machine-readable marking. A synthetic video of a real person, shown to people who could take it as genuine, is the case those rules were written for. In the United States, several states, including California and Tennessee, have legislated on digital replicas and voice likeness, and contracts with presenters should say what may be synthesized.
The practical rule is simpler than the law: if a reasonable viewer could believe a real person said it, say it is AI-generated, on screen, at the start. A label in the video description is not enough.
Personalized sales video at scale
Avatars became popular in marketing with a promise of lower customer acquisition cost: one recording, thousands of personalized outreach videos. The economics can work. The enterprise risk is that personalization at scale is also impersonation at scale if a prospect cannot tell the rep did not record it. Disclose, keep a register of every campaign that uses a real person’s likeness, and do not let the avatar make commitments, such as prices, terms or delivery dates, that a human rep would have to approve.
A short policy for AI avatars
- Decide who may create avatars of real people. Stock presenters can be open to everyone; real likenesses should require an owner and a consent record.
- Keep a register of every real-person avatar: whose likeness, consent date and scope, where it is used, and the expiry.
- Label every synthetic presenter on screen, in the viewer’s language.
- Put avatar videos under the same review as the content they deliver. A compliance training video is controlled content whether a person or an avatar presents it; the logic is the one we apply to agent skills as controlled documents.
- Harden payment and approval flows against synthetic executives: call-back on a known number, two people for transfers above a threshold, and no approvals by video alone.
- Check what your office suite allows by default. If avatar creation has no admin control, the policy and the training are your control.
AI avatars are a useful tool for training and communication in many languages, which is exactly why they will spread faster than the rules for using them. The face and voice of a real person are the part that needs a decision before the first video, not after.