Synthesia moves from corporate video to avatar roleplay

On 22 July Synthesia unveiled AI Roleplay Sessions, a format in which the employee stops watching a corporate video and starts talking to the avatar instead. The session is interactive, the avatar responds during the conversation, and when it ends it returns feedback, a score and aggregated analytics so the company can check whether the training is worth anything. TechCrunch reported it on the day of the announcement.

The move has less to do with avatars than with the last word in that sentence: measuring. Synthesia grew big selling corporate video production with no camera, no studio and no actors, a business with a fairly obvious ceiling. The training department delivers the video, logs that the employee hit play, and that is where the evidence ends. A scored roleplay turns that content delivery into data.

From a video catalogue to a scored session

The operational difference is bigger than it looks. A video course is an asset produced once and distributed to the whole workforce. A roleplay session is a different interaction every time, with an output that has to be assessed on the spot. Hence the three pieces described in the announcement: feedback on what the employee said, a score for the conversation, and analytics that aggregate those results at team and organisation level.

The analytics part is what matters to whoever signs the budget. Conversational skills (negotiating, handling conflict, dealing with an angry customer) have always been assessed with satisfaction surveys and multiple choice tests, which measure recall rather than behaviour. A recorded and scored session does not fully solve that problem, but at least it watches the person doing the task instead of asking them whether they enjoyed the course.

Why now

Roleplay with professional actors works well and costs a lot. That is why it is reserved for executive committees and high value sales programmes, and why everyone else gets the video and the closing quiz. An avatar that can hold a hundred simultaneous conversations at close to zero marginal cost goes straight at that gap.

There is a market reading too. Generative video platforms have spent a couple of years competing on lip sync, language catalogues and avatar realism, a race where differentiation runs out fast because everyone ends up in the same place. Moving into the customer workflow, in this case the full training cycle and its measurement, is the usual way out of that trap.

Who it is useful for

Three profiles fit well:

Training teams with large headcounts and high turnover, where cost per trained employee outweighs everything else.
Sales and customer service, the two departments where the conversation literally is the job.
* Onboarding and compliance, where you need to prove to an auditor that the training happened and who completed it.

It fits worse when the conversation depends on context the avatar does not have: the history with that specific client, unwritten internal policy, the cultural nuance of a local market. And there is one point worth settling before rolling anything out: a model generated score on how an employee speaks is, in practice, performance data. In Spain and across the EU that lands squarely in GDPR territory, in works council disclosure duties and in the debate around automated decisions. It is not fine print.

We read this as a sensible and fairly unspectacular move, which in corporate training is usually a good sign. The question we would ask before signing a contract is not whether the avatar is convincing, but what that score actually measures and who is going to read it.