Kavak runs a dedicated AI agent for every customer, and the conversion numbers justify the architecture

Carlos García’s description of Kavak’s AI architecture is compact enough to misread as hyperbole. “We have an agent that’s obsessed with each of the customers, like millions of this.” The framing is casual, but the operational claim is specific: not one agent serving all customers, not one agent per task type, but a dedicated agent per individual, instantiated at a scale of millions. That is a structural choice, not a feature toggle, and it implies a particular theory of how feedback loops in car commerce should work.

External reporting published by osmu.app confirms the architecture is real and puts firmer numbers around it. According to that account, Kavak deploys between 100,000 and 200,000 customer-specific AI agents daily. Each agent handles personalized interactions for its assigned customer rather than operating as a shared resource. The daily deployment figure means the system is not a fixed fleet but a continuously refreshed one, matched to whoever is active at a given time.

The performance data attached to that architecture is the part that earns serious attention. The osmu.app report indicates that AI agents handle between 90 and 96 percent of Kavak’s customer interactions, and that those agents convert at 2.1 times the rate of human agents. A conversion multiplier of that size, sustained across a business that operates in one of the higher-friction categories in consumer commerce, used car sales, is not a rounding-error improvement. It is a result that changes the unit economics of the customer acquisition model.

We have an agent that's obsessed with each of the customers like millions of this Carlos García

The architectural choice García describes, one obsessed agent per customer rather than a shared pool, is the part worth unpacking. The standard deployment pattern for AI in customer-facing operations is task-based: a bot handles inbound queries, another manages follow-up, a third escalates to humans when confidence drops. That model treats the customer as a series of discrete interactions. Kavak’s model, as García frames it, treats the customer as a continuous subject with a single agent accumulating context and maintaining orientation toward that one person’s situation. The feedback loop García references is the mechanism that makes that orientation useful: an agent that only handles one customer has every reason to model that customer accurately, because it has no other job.

What the external evidence adds is scale confirmation. García’s claim could plausibly describe a pilot or a directional aspiration. The 100,000-to-200,000 daily deployment figure suggests the system is running at production volume, not proof of concept. That distinction matters when evaluating whether the architecture is replicable or specific to Kavak’s particular mix of data, inventory, and customer base.

The 2.1x conversion figure also raises a question the evidence does not fully answer. It is not clear from the reporting whether the outperformance is attributable to the one-agent-per-customer structure specifically, or to the quality of the underlying models, or to some combination of the two. Personalization at that granularity is one hypothesis for why the numbers are where they are. The counterfactual, the same models in a pooled task-based deployment, has presumably not been published. García’s own framing points toward the architecture as the causal factor, but the external data confirms the outcome rather than isolating the variable.

What is confirmed, and what makes García’s description worth reporting precisely, is that a company operating at Kavak’s scale has committed to a deployment model that most organizations have not attempted. Millions of individual agents, each one calibrated to a single customer, handling the overwhelming majority of interactions, and converting at more than twice the human rate. Whether the industry reads that as a template or an outlier depends on whether the feedback-loop logic García describes turns out to be the actual driver of the result. The evidence so far points in that direction.