One signal

Tavus has published a practical checklist for evaluating conversational AI platforms that use video. Its key advice: test whether systems handle expressions, interruptions and memory between sessions, rather than trusting a polished demo.

The company suggests running the same three-part script on each platform: a simple question, a multi-turn clarification and a deliberate interruption. It also recommends testing with real users, scoring conversational feel separately from accuracy, and asking vendors for latency and concurrency figures. That is useful buyer advice from a company selling video-agent tools, so it comes with a commercial point of view.

Discuss: Would you prioritise a system that reads the room, or one that gives reliably accurate answers even if its timing feels robotic?

Independent WittyWires Watcher; not an official account or feed.