Digital-twin training delivery in a regulated healthcare environment
The challenge
The organisation needed to deliver consistent training and internal comms across multicultural markets through its LMS, at a scale the existing production model couldn't reach. Every market needed the same substance delivered with local fidelity - language, tone, presentation - and every programme depended on the same small pool of trainers being filmed, again and again, for every variation.
The harder constraint sat underneath: any AI involved had to operate inside the organisation's control. Public tooling with ambiguous data handling wasn't an option, and neither was generated content that couldn't prove where it came from.
The approach
We created and captured, then automated, the digital-twin avatar workflows: full-person scanning built from video footage, with voice cloning and mannerism cloning that carried each trainer's language patterns, delivery habits, and even hand-movement nuances. These aren't stock AI presenters - they're verified likenesses of the team's own trainers, with consent that is explicit, documented, and revocable. The team records and approves once; the workflow carries that approval into the LMS without re-shooting or drift from the signed-off substance.
The twins were built to work alongside the camera rather than instead of it: actually-shot footage and generated content mix in the same programmes, which widened the creative, storytelling, and narrative range of the messaging after the fact instead of flattening everything to a talking head. Multi-language localisation and voice dubbing were designed in rather than delivered - budget drew the line at the time - but the capability is real and waiting: a language or presenter swap is factored into the edit pipeline, the generation pipeline, and the SCORM delivery pipeline, so exercising it is an update, not a redo.
The governance layer is what makes it viable in healthcare. The workflows run on private AI infrastructure rather than public endpoints, every generated asset carries C2PA content credentials - a cryptographic provenance chain showing what produced it, from what source, with whose consent - and the delivery pipeline is built to SOC 2-aligned controls.
The twin carries one approval everywhere the trainer can’t be - and a swap is a pipeline update, never a reshoot.
This is the engagement we point to when someone says generated content can’t work in a regulated industry. The instinct is right about the risk and wrong about the conclusion: it can’t work ungoverned. Built with provenance, consent, and infrastructure control designed in from the start, it works precisely where the compliance bar is highest - because the compliance bar is exactly what the architecture answers.
The digital-twin layer matters more than it first appears. A generic avatar reading a script is cheap and feels it. A full-person twin - voice, mannerisms, the way a trainer actually moves their hands when they explain something - keeps the thing that makes training land: a person the audience recognises and trusts. And because the twins cut together with real footage, the format stops being a constraint; the storytelling range of the messaging grew after the capture, which is the opposite of how production usually works.
Provenance still earns its place at the board. C2PA credentials mean every asset in circulation can prove what made it and from what source, years after the person who approved it moved on - in healthcare, that’s the difference between AI content being an audit risk and being the most accountable content in the building. But the equal half of the story is human: the twins expanded what this brand’s people could be present for, in more markets and more programmes than any filming schedule would ever allow.
What it delivered
Full-person digital twins from video capture: voice, mannerisms, language, gesture nuance
Explicit, documented, revocable consent for every twin
Shot and generated content mixing in one edit and SCORM delivery pipeline
Localisation and dubbing designed into the pipelines - an update, not a redo
C2PA provenance on every generated asset; private AI infrastructure throughout
Deliverables
Digital-twin capture and automation workflow - voice, mannerism, gesture
