Learn how Kavak transformed into an AI-native company deploying 100,000+ agents daily. Discover the architecture, evals strategy, and organizational model th...
How Kavak Built an AI-Native Company: The Complete Playbook
Key Takeaways
- Scale: Kavak instantiates 100,000–200,000 customer-specific agents daily, each with its own virtual machine and persistent memory
- Performance: AI agents convert 2.1x more customers than human teams and handle 96% of all interactions with zero human involvement
- Architecture: Long-running agents with hard goals (not workflows) optimize lifetime customer value across sales, financing, and logistics
- Evals Matter: Equal engineering investment in evals as agent-building ensures continuous improvement and risk mitigation
- Organizational Shift: From transactional to relational; flat teams where humans build agents, work alongside agents, or oversee physical operations
- Training: Six-week "Jedi Academy" program trains all employees (CEOs to mechanics) to build production-ready agentic systems
The AI-Native Architecture: One Agent Per Customer
Kavak reimagined its operations around a radical principle: one intelligent agent per customer, not per task.
When a customer visits Kavak, a dedicated agent spawns instantly with its own virtual machine. This agent remembers every interaction—what the customer viewed on the website two years ago, their preferences, their financial history—and creates a long-term strategy to maximize that customer's lifetime value. The agent handles everything from initial inquiry through loan approval, insurance, trade-in quotes, and follow-up sales.
This differs fundamentally from traditional chatbot or workflow-based approaches. Rather than routing customers through sequential steps, the agent acts as a persistent, learning partner throughout the relationship. If an agent encounters a problem beyond its current capability, it requests help from a human team member—but the agent remains the owner of the customer relationship, ensuring the feedback loop stays intact for continuous learning.
Evals: The Hidden Lever of Velocity
Kavak's breakthrough came from a counterintuitive principle: to move fast, you need effective brakes. The company invests roughly equal engineering time, tokens, and resources in building evals (evaluation frameworks) as it does in building the agents themselves.
Most companies measure vanity metrics—call duration, call volume, adoption rates. Kavak measures what actually matters: did the customer convert? Metrics include final purchase, loan approval, or return visits. Secondary indicators confirm whether the agent added genuine value.
To organize token spending strategically, Kavak categorizes tokens into three tiers:
- Tier-3 (Most Valuable): Agents performing core business functions. ROI is directly measurable per token; these drive organizational outcomes.
- Tier-2 (Indirect Value): Agents supporting development or code quality. Value is measurable but indirect.
- Tier-1 (Unclear): General ChatGPT or Claude usage with no framework. No visibility into ROI.
Most companies operate at Tier-1. Kavak operates at Tier-3, ensuring every token spent delivers quantifiable business results.
From Transactional to Relational: How Agents Became Better Salespeople
Selling a used car in Latin America is complex. Customers choose from 20,000+ SKUs, then navigate financing, insurance, coverage, and trade-in quotes. Historically, this required expertise across 15 different domains split among 15 different specialists.
Kavak trained specialized agents for each domain—financing expert, car advisor, insurance specialist, trade-in evaluator—then unified them into a "mega agent" capable of handling the entire customer journey seamlessly.
The results:
- NPS tripled (Net Promoter Score)
- Initial conversion: 50% higher than human teams
- Current conversion: 2.1x higher than human teams
Why are agents better sellers? They are infinitely patient, know the complete customer history, understand long-term value, and never fatigue. Critically, when an agent makes a mistake, all 200,000 agents system-wide incorporate that learning the next day—not just that single agent. This creates a continuous feedback loop that accelerates improvement across the entire fleet.
Financial Services Without Humans: Underwriting at Scale
Kavak extended agents into regulated financial services. Car loans in Mexico typically take two months to approve. Kavak agents approve loans in under three minutes.
This became possible because agents have access to comprehensive customer and vehicle data. If a customer's financial situation changes, the agent can offer alternatives—such as returning the car and financing a cheaper option—keeping the customer financially healthy and loyal.
The insight: major financial decisions (car purchases, large personal loans) often take customers 3–4 months to finalize. By knowing the customer deeply and simplifying the process, agents dramatically improve conversion and retention. The strategy personalizes interest rates, risk assessment, and loan amounts to individual circumstances while remaining sound for the overall portfolio.
The AI CEO and Agents in Management
To test whether agents could handle leadership roles, Kavak deployed an AI agent as CEO of its Cuernavaca, Mexico operations for six weeks. The agent's goal: double the city's profits.
Result: 50% profit increase (1.5x).
The agent achieved this by analyzing every number, every customer, and every operational detail—then issuing daily directives to field teams and monitoring progress via voice notes. Every KPI improved: customer satisfaction, inventory management, financing penetration, and car rotation.
This demonstrated that agents could perform roles traditionally considered immune to automation: strategic planning, decision-making, and team coordination.
The Jedi Academy: Training Everyone for an AI-Driven Future
Recognizing that every job would change, Kavak launched the Jedi Academy, a six-week training program open to all employees—from the CEO to mechanics to finance teams.
Graduates emerge equipped to build production-grade AI agents, not necessarily becoming AI engineers, but gaining the skills to collaborate effectively with agentic systems. The message was clear: this is the future direction of Kavak. Employees could upskill or leave.
The result: cultural alignment, a workforce that understands how to design and work alongside agents, and continued relevance as AI capabilities evolve.
Human Roles in an Agent-Driven Organization
Kavak retained human workers in roles requiring physical dexterity and sensory judgment—roughly 800 mechanics in Mexico.
Rather than replacing mechanics, Kavak built an AI sidekick called El Mike (inspired by the movie Ratatouille, where the mouse guides the chef). El Mike instructs mechanics on how to inspect and repair vehicles, offering real-time tips and guidance.
Outcome: faster inspections, faster repairs, lower costs, higher-quality cars, and warranties down 26%.
The organizational structure became flat and cross-functional. Teams include engineering, AI, and operations roles, each member either building agents, working for agents, or managing physical-world interactions. Middle management as traditionally conceived largely disappeared; decision-making became distributed to empowered, senior-heavy teams.
Organizational Transformation: Top-Down Strategy, Bottom-Up Skills
Kavak's transformation succeeded because it followed two principles:
Top-Down Strategy: Leadership defined a clear vision (Kavak in 2035 with advanced AI). Transformation wasn't organic; it was mandated and guided. Hackathons and bottom-up ideation don't work at scale; strategy must come from the top.
Right Evals, Not Just Adoption: Spending on tokens alone is wasteful. The company must measure which tokens drive business outcomes (Tier-3), which support operations indirectly (Tier-2), and which offer no visibility (Tier-1). Most companies waste tokens at Tier-1.
The Broader Lesson: Creative Destruction in AI
Kavak's success illustrates an economic principle: creative destruction. Incumbents rarely fully redesign themselves around new technologies; instead, new companies built natively around those technologies disrupt incumbents.
When electricity emerged, factories didn't swap coal engines for electric ones—they could have, gaining 6% efficiency. Instead, companies like Ford rebuilt factories from the ground up, achieving 3x productivity gains. The same happened with computers and is happening now with AI.
Incumbents face a dilemma: redesigning the entire organization is risky and difficult. New companies can be built around AI's strengths from day one. This creates an unprecedented opportunity for founders: access to world-class intelligence for $20/month, combined with the chance to build a company for 2035, not 2024.
Conclusion
Kavak transformed from a used-car marketplace into an AI-native company by committing fully to agent-based architecture, rigorous evals, and organizational redesign. The payoff: agents handle 96% of interactions, convert customers 2.1x faster, approve loans in minutes, and improve operations through persistent learning loops.
The path forward for ambitious companies is clear: design your organization for the future, not the present. Founders should build native AI companies; larger organizations should commit to top-down transformation with clear evals frameworks. The most exciting entrepreneurial opportunity of our time has never been more accessible.
Original source: Kavak's Playbook for Rebuilding a Company Around AI
powered by osmu.app