Latin American Used Car Platform Kavak Wakes Up 200,000 AI Agents Daily, 96% of Interactions Handled by Agents
Kavak's Chief Product and AI Officer Alejandro Maza revealed in the a16z podcast that the company wakes up up to 200,000 AI agents daily, with 96% of customer interactions completely handled by agents without human intervention, and 95% of transactions completed end-to-end by agents.
Maza emphasized, "I like to move fast, but to be fast, you need brakes," believing that the quality of evaluations directly determines the speed limit at this scale. Many companies choose to slow down due to a lack of effective brakes, while Kavak's approach is to first ask, "How fast can we go?" The answer depends on the quality of the evaluation system.
The company's rule of thumb is that the engineer time, tokens, and funds invested in building agents should be roughly equal to the investment in building the evaluation system, ensuring reliable constraints during high-speed operations.
Kavak has fully restructured its business around agents, with agents responsible for selling cars, approving loans, and guiding technicians. In a city in Mexico, the entire operation is led by AI, resulting in a threefold increase in NPS, doubled sales conversion rates, and loan approvals shortened to under three minutes.
Maza pointed out that the team had to dismantle a two-year-old architecture to restart, shifting to a model where each customer corresponds to a persistent memory agent, and training all staff, including technicians, to be able to launch production-level agents within six weeks.
In market mechanisms, large-scale agent deployment decouples labor costs from processing capabilities, directing capital and efficiency dividends to companies that can simultaneously build high-speed agents and high-quality evaluations, putting pressure on traditional labor-intensive processes and accelerating event-driven operational restructuring.
Investment in evaluations is positioned as a necessary prerequisite for speed rather than a patch applied afterward.
Source: Public Information
ABAB AI Insight
Alejandro Maza has led Kavak's product and data strategy since 2022, driving the company from early machine learning attempts to a complete shift towards agent-centric operational restructuring. Previously, the team had spent two years building a multi-agent framework before deciding to start over.
The capital path is reflected in the equal allocation of engineering resources to agent construction and evaluation systems, motivated by the need for reliable brakes to support rapid expansion, while training all staff (Jedi Academy) to decentralize agent production capabilities to non-engineering positions, reducing reliance on specialized AI talent.
Similar to other companies attempting large-scale agent deployment, they ultimately face constraints from evaluation and reliability bottlenecks, currently transitioning from "equipping employees with AI tools" to "restructuring the company with agents."
Essentially, this involves parallel technological replacement and organizational restructuring, where the mechanism is that when agents handle the vast majority of interactions and transactions, the quality of evaluations becomes the true ceiling for speed and scale.
ABAB News · Cognitive Laws
- You can't accelerate without brakes.
- Investment in evaluations determines the speed limit of agents.
- The larger the scale of agents, the more symmetrical constraint systems are needed.