Overview
Kavak offers a rare case study of a company attempting to reorganize its operations around AI agents rather than merely equipping employees with AI tools. Head of AI Ali Masa describes an architecture in which a dedicated agent, running in its own virtual machine, remembers each customer's history, pursues long-term goals, and coordinates sales, financing, insurance, and operational tasks. He claims agents now handle 96% of customer interactions and 95% of transactions, while outperforming human teams on conversion and customer satisfaction. The transformation depended on three linked changes: rebuilding company systems and APIs for agent access, exposing agents to real customers so feedback could drive improvement, and replacing transactional activity metrics with outcomes such as conversion, satisfaction, and lifetime value. Kavak also invests roughly equal resources in evaluations and agent development, treating measurement as the safety system that permits rapid deployment. Humans remain central in physical work, exception handling, skill creation, and agent supervision, but the organization has become flatter and more senior. Masa's broader thesis is that incumbents adopting AI superficially may gain modest efficiency, while companies redesigned around increasingly capable intelligence could achieve step-change productivity and disrupt existing markets.
Sections
Strategic Insights
Higher-order implications of Kavak's operating model and reported results.
- The durable unit of agentic organization may be the customer relationship rather than the task. Giving one agent persistent memory and an enduring objective allows locally separate activities such as sales, lending, insurance, and retention to be optimized as one relationship.
- Kavak's evaluation system functions as both quality control and organizational governance. Once agent behavior is tied to business outcomes, management can delegate more execution without losing visibility into whether the system is creating value.
- Human-in-the-loop design becomes more valuable when the agent retains ownership of the case. The human is then an invoked organizational capability, while the agent preserves continuity and captures the resolution for future improvement.
- The proposed self-improving system is organizational rather than merely model-level: better models, shared lessons, tools, evaluations, and human capabilities compound across the company instead of improving only one isolated agent.
Architecture and Operating Metrics
Specific implementation characteristics and quantitative claims described in the interview.
- A customer-specific agent runs inside its own virtual machine with access to memory, evaluations, a command-line interface, company APIs, tools, and a long-term objective.
- Between 100,000 and 200,000 agents are instantiated daily; individual agents may work for three minutes, eight hours, or three days and can schedule their next task before becoming inactive.
- Agents reportedly handle 96% of customer interactions and 95% of transactions, excluding unavoidable physical interactions such as handing over vehicle keys.
- The company allocates approximately equal engineering time, money, and token resources to evaluations and to agent construction.
- Kavak's car-loan process reportedly produces approvals in under three minutes, compared with a process that may take two months or more in some emerging markets.
- The AI CEO experiment in Cuernavaca pursued a first-month goal of doubling profit and reportedly achieved a 1.5-times result after six weeks of operation.
Lessons for Operators
Transferable lessons drawn from Kavak's implementation experience.
- Start with a concrete vision of the AI-native company several years ahead, then rebuild toward it; disconnected hackathons and voluntary adoption are unlikely to produce structural change.
- Measure conversion, customer value, satisfaction, and re-engagement before optimizing proxy metrics such as call volume or interaction duration.
- Invest in evaluations at the same stage and approximate intensity as agent development because deployment speed depends on the ability to detect failure.
- Expose agents to difficult real-world work under measurable safeguards so failures produce data that improves the system.
- Train the entire workforce repeatedly because agentic collaboration changes operational, technical, managerial, and physical roles simultaneously.
- Be willing to replace successful orchestration patterns when more capable models make the old architecture a constraint.
Forecasts
Future outcomes anticipated by the speaker.
- AI agents will likely outperform humans in most information-based jobs, potentially including executive leadership, if current capability improvements continue.
- Organizations are expected to become flatter, more senior, and organized around teams that build agents, work for agents, or perform physical customer-facing tasks.
- Companies will increasingly focus recursive improvement on the organization—the combined agents, tools, evaluations, and workflows—rather than only on improving individual models.
- AI-native startups may displace incumbents that adopt AI superficially because complete redesign could produce multiples of productivity rather than single-digit or low-double-digit efficiency gains.
- Human work is likely to remain comparatively durable in physical roles requiring dexterity and sensory judgment, while agents increasingly direct or assist that work.