Labelled AI presenters for explainers and localisation

The transparency duty in Article 50 of the EU AI Act has applied since 2 August 2026: synthetic presenters and cloned voices shown to people in the EU have to be disclosed as artificial. So everything here ships labelled, on screen, in the caption text and in the file's provenance metadata. That is the starting condition of the work, not an upsell, and it is worth asking any other supplier what they hand over instead. Where an avatar earns its place: Information you have to repeat. Patch notes and release videos for a live game, product explainers that change every sprint, onboarding and internal comms, publisher and franchise updates, the same ninety seconds in eight languages. The economics only work where the script changes weekly and the face is not required to mean anything. Where it does not: Endorsement. An avatar cannot want your product, and viewers read that faster than the tooling improves. If the brief is a customer saying they love the thing, that is a different service on this profile and it is done with real people or not at all. How it is made: HeyGen and Synthesia for stock presenters, and for anything recurring a custom avatar built from a three to five minute consented capture of a real person on your team. Someone who will still be there in a year, or under a contract that survives them leaving. The rights document names an expiry and a kill clause: if they leave, the avatar is retired on a stated date and the source recordings are deleted rather than archived, and that is written before the capture, not after. Script craft: Lipsync and prosody degrade above roughly 150 words per minute, and on acronyms, figures and long German compounds. So the German script is written first here and the English derived from it, which is the reverse of the usual order. German is where the model fails first, and writing for the failure mode costs less than discovering it in the render queue. Numbers are spelled out, compounds are broken, sentences carry one clause. Version lock: Avatar model versions change without notice, and a re-render in November will not match the batch you shipped in August. Same face, different mouth, and the mismatch is visible when the videos sit next to each other in a playlist. I pin versions, keep the project files and the exact prompts, and tell you when a forced upgrade means the library needs a full re-run rather than a patch. What I refuse: An avatar with an invented name presented as a real customer. An avatar of anyone who has not signed, whether a public figure, a deceased person, or your own former spokesperson whose contract has lapsed. Unlabelled delivery, including for clients confident their market will not notice. Voice cloning without a paid licence with an end date. Not included: No product handling. Avatar hands cannot hold your product and composited attempts look exactly like composited attempts. No media buying. No translation QA beyond German and English without a native reviewer, whom you pay directly and who holds the right to reject a render. Who this is not for: Anyone whose interest in the format is skipping the disclosure. And brands who want an avatar because it looks cheap: below roughly twenty videos a quarter, a real person on camera is both cheaper and better, and I will tell you that on the first call rather than after the invoice. How it runs: We agree what the format is carrying and what it is not, in writing. Then capture and rights, then a pilot batch of three videos in German and English so you can see the failure modes on your own script before committing a library to them.