Tavus's Griffin Can Hold a Video Conversation. The Business Case Is Still Unproven.

On October 1, Tavus introduced Griffin, a video model that keeps listening while it talks, reads visual cues, and generates a lifelike face and voice in real time. For a small B2B marketing team, that points toward a new kind of product demo: a visitor could interrupt, show the agent a problem, and ask a question instead of clicking through a prerecorded walkthrough.

The catch is that Griffin isn’t available to customers yet, and the launch materials don’t report a conversion lift, qualified-lead rate, or cost per resolved question. The technology is moving toward the sales funnel faster than the business case.

Tavus described both the promise and the risk in its announcement:

The same properties that make Human Interaction Models powerful interfaces for natural communications between human and machine allow them to deceive a human into believing it is not AI.

The company says Griffin-Lite is limited to a research preview for selected testers while it works on disclosure and safety features.

A face that listens while it speaks

Most voice agents wait for a person to finish, turn speech into text, generate a reply, then pass it to a voice or avatar system. Griffin is designed to process video and audio continuously. It can react while a person is speaking, wait through a pause, or respond to something shown on camera. Those behaviors matter in a product walkthrough where a buyer may point at a confusing screen or interrupt with a pricing question.

The clearest outside measurement comes from NVIDIA’s VideoFDB benchmark. On its generation track, Griffin-Lite scored 3.83 out of 5, close to the human reference score of 3.92 and above the next listed system at 2.80. On the perception track, it scored 3.73 against a 4.20 human reference. NVIDIA evaluated 237 annotated clips from real video calls and used a language-model judge for its rubric scores. The benchmark is a useful comparison, but it doesn’t measure whether a buyer trusts an avatar, books a meeting, or purchases a product.

Tavus also reports that 26 of 54 participants thought Griffin-Lite was a real person after a one-minute call. Participants were told to expect another person, then learned they had spoken with AI after the survey. The sample was small and company-run, and The Rundown notes that the result says little about longer calls. Infobae points out that the setup also differs from the classic Turing test: participants weren’t told an AI might be on the call or asked to identify one. The 48% figure shows how many didn’t suspect during a brief conversation, not how well people can detect AI when they know to look for it. It’s evidence about a convincing interaction, not evidence of market demand or sales impact.

More than one company is building the format

Griffin is a preview, but the broader move is already visible in products aimed at business use. On October 1, Synthesia launched Sessions for role-play and customer interviews, and its Interactive Avatar API lets enterprise teams add a real-time avatar to an agent they build. The company positions the format for sales practice, support, onboarding, and research. Earlier this year, D-ID introduced Agentic Videos, which lets viewers ask questions about a video and receive answers tied to its script and connected knowledge.

That distinction matters to marketers. A generated spokesperson reads a prepared message, like the AI-produced video in Cuisinart’s recent ad test. An interactive avatar can become another interface for a product page, demo, or support flow. It may help when buyers need to explore details in their own order, though chatbots already talk some buyers out of purchases. It also adds a new surface where an answer can be wrong, disclosure can be missed, or an avatar can imply a human is present.

Can Timağur, founder of Allbound OS, tested Griffin and called it a research preview rather than a product launch:

I wouldn’t call this a product launch. It feels more like a research preview of a new category.

That is the right read on the model’s availability. Synthesia’s offers are further along, though its API is currently for enterprise customers and requires a team to connect the avatar to its own AI and conversation logic.

Pete Brooks, global head of total talent and performance at LQRA, described one organizational use in Synthesia’s launch post:

The irony of AI is that it’s shining an even brighter light on the need for enhanced human interaction, putting a premium on the human skills we’ve all neglected for so long.

His comment is about practice and training, not measured customer outcomes. That distinction is the gap marketers should keep in view. Even in a real customer-service deployment, the TSA’s narrower AI agent depends on clear data and escalation paths, not a humanlike face.

Test the conversation, not the face

There is a reason to be cautious beyond a single product’s safety plan. ITPro reported that a Sinch-commissioned survey of more than 2,500 industry leaders found 74% had shut down or rolled back AI agents used for customer communications, citing governance failures. The survey was commissioned by a communications vendor, so the number should be treated as a reported survey result, not a universal failure rate. Still, it shows that adding a more human-looking interface doesn’t solve the harder work of keeping answers accurate, auditable, and on-brand.

For a founder-marketer, the useful test isn’t whether the avatar looks convincing. Start with one narrow, high-intent use case, such as answering questions about a complex product demo. Make the AI disclosure visible before the conversation begins, restrict the answers to approved product information, and compare qualified meetings or resolved questions against the current experience. If the avatar adds cost without improving one of those outcomes, its realism is decoration.

Tavus has shown a stronger benchmark result than other systems NVIDIA evaluated, and Synthesia and D-ID show that interactive video is not confined to one launch. Neither fact proves that an AI face sells better. Until a vendor shows business outcomes and reliable disclosure in a customer-facing deployment, keep the idea in a controlled test, not at the center of the funnel.