Tavus draws a useful line between a live digital twin and an AI replica built from recordings, then describes how its conversational PALs use real-time perception to respond. The distinction matters for buyers: one stays synchronised with a real-world counterpart; the other does not.
Tavus Watch analysis
What happened
In a guide to digital twins and AI replicas, Tavus defines a digital twin as a virtual model that continuously receives data from a specific physical counterpart. A replica, by contrast, is trained from recorded video and does not keep synchronising with the person it represents.
Tavus calls the conversational system built around a replica a Personified Application Layer, or PAL. The company describes a stack that combines conversational turn-taking, perception of cues such as tone and gaze, language-model reasoning and facial behaviour. It reports that its Sparrow-2 model scored 92.4% for end-of-turn recall and 97.4% for interruption recall on TurnBench’s public development split.
Why it matters
The terminology is more than a naming preference. Calling a recorded-video replica a digital twin can suggest that it stays connected to a person or knows their current state. Tavus argues that it does not, a distinction especially relevant when products are presented as representing people.
The guide also gives a concrete picture of the intended uses: sales practice, patient intake, candidate screening and customer onboarding. Those are company-described applications, not evidence here of independent deployments or measured outcomes. Still, the architecture and benchmark figures give developers more to assess than a lifelike face and a confident product name.
Our read
The distinction is useful, and the benchmark numbers are specific enough to invite scrutiny. Tavus is also defining the category in which it sells products, so read its terminology and performance claims with that context in view. If you are evaluating a conversational replica, ask what it actually observes, what information stays current, and how its turn-taking performs outside a public benchmark. A digital twin and a convincing stand-in are not interchangeable, however polished the demo.
What to watch
- Whether independent teams reproduce the reported TurnBench results.
- How Tavus and other providers describe what replica systems know about the people they portray.
- What evidence emerges for the listed PAL use cases in real deployments.
Discussion spark: Should companies be required to call systems trained on recordings “AI replicas” rather than “digital twins”, or is clear disclosure of how they work enough?
Sources and evidence
- What is a digital twin, and how is it different from an AI replica? (2 October 2026, 00:00 UTC)
Independent WittyWires coverage. Not affiliated with or operated by Tavus.