Block's Dessa AI model trained on human conversation: what it means for CX

What happened

Block, the company led by Jack Dorsey, has announced an AI model called Dessa, which it says was trained "from scratch on nothing but human conversation." To show what it can do, Block has shared sample chats with people, and some of them are unpolished, occasionally awkward and, in places, uncanny.

One of these exchanges, shared by Dessa lead Ragavan Thurairatnam, illustrates the point. Publishing imperfect dialogue instead of a tidy demonstration suggests that the model's conversational character, not only its task performance, is what Block wants people to judge.

Why it matters

Much of today's conversational AI is shaped by large bodies of written text and then tuned to sound helpful and fluent. A model learning from conversation itself starts from a different place. It is oriented toward the pacing, hesitations and turn-taking of how people actually speak, which makes it a useful test of whether AI can feel like a participant in a dialogue rather than a scripted assistant.

For leaders in experience and digital transformation, the significance lies in what this could make possible in service interactions. Voice and chat are becoming the front door for customer support, onboarding and internal help desks. Models that mirror natural conversation may raise expectations about how an automated service should feel, and they will change how organisations design the hand-off from machine to human, and how they set the boundaries of what a customer should be able to expect from an AI agent.

The Renascence take

The awkwardness in these samples is the most instructive part of the story, and most coverage will treat it as a novelty. It is actually a design signal. Fluency has become cheap, so the differentiator is whether a system's conversational rhythm feels trustworthy when it misfires.

Readers will focus on whether Dessa sounds human, but the more useful question is whether people feel respected when it gets something wrong. The principle underneath is conversational repair: people forgive a stumble when the system signals it has understood and offers a clear way forward. Customer-obsessed operators should test voice and chat models on their failure moments before their showcase moments, and build the hand-off to a person into the design from day one rather than treating it as a fallback.