Kavak Rebuilds as an AI-First Company

Kavak is pioneering a strategic shift towards an agent-led architecture, departing from traditional knowledge worker models by deploying superhuman AI agents.

Alejandro Maza Ayala, Kavak's Chief Product & AI Officer, revealed that between 100,000 and 200,000 unique agents are instantiated daily, each operating within its own virtual machine specifically for customer interactions.

The company also implemented a 'Jedi Academy' program to train employees across all levels, from CEOs to mechanics, in building these AI agents, with the goal of surpassing human performance in all key metrics.

Alejandro Maza Ayala reflected on his career path, which began with founding Oppy Analytics in 2013, a company dedicated to machine learning and risk algorithms.

He considered his previous venture ahead of its time, existing a decade before the advent of Transformers, and saw joining Kavak as an opportunity to apply advanced AI to scale a robust business model.

Because the necessary infrastructure was absent in Latin America, Kavak established a vertically integrated model to control the entire lifecycle of purchasing, refurbishing, financing, and selling vehicles.

This extensive vertical integration became necessary due to the underdeveloped infrastructure for these services in Latin America, compelling Kavak to build comprehensive solutions in-house.

Alejandro Maza Ayala indicated that merely providing tools like ChatGPT to existing teams often fails to generate efficiency or solve customer problems.

Kavak chose a more transformative approach by redesigning its entire company, including its APIs and team structures, to be centered around AI agents, a process that initially struggled for a year before yielding success.

Kavak shifted from a transactional business model, where success was measured by the volume of cars bought or sold, to a relational one.

AI agents now manage 10 million customers, tasked with maximizing their lifetime value through personalized financial services, with activating just 1% of this base representing substantial revenue potential.

Building trust is a foundational element in the used-car market, which the relational model aims to solidify for long-term engagement.

Kavak allocates an equal amount of engineering time and financial resources to building evaluations ('evals') as it does to developing AI agents, viewing these evaluations as critical 'brakes' for rapid innovation.

The company disregards superficial KPIs like call volume, instead prioritizing actual conversion metrics to assess agent performance.

The core focus is on whether an agent successfully converts a customer and fosters sufficient satisfaction to encourage future re-engagement.

The sales process in the region typically involves about 15 intricate sub-processes, encompassing financing, insurance, and other services, which would traditionally require 15 different human expert teams.

Kavak's AI agents are designed to adeptly navigate this multi-dimensional sales environment, streamlining the complex process.

A critical feature of this system is its self-improving feedback loop: when one agent learns from a mistake, that knowledge is immediately disseminated to the other 200,000 agents.

This continuous learning mechanism not only drives growth and enhances customer satisfaction but also scales efficiently across the entire workforce.

This AI operates at a 'Fields Medal-level' of intelligence, meticulously analyzing every number and customer data point, and micromanages daily operations by assigning tasks to field workers and requiring voice note updates on progress.

Kavak has implemented an AI sidekick named 'El Mike' to assist its approximately 800 mechanics in Mexico with physical world tasks, particularly vehicle inspections.

Inspired by the collaborative concept in the movie Ratatouille, 'El Mike' guides mechanics with real-time tips during inspections.

Since its launch, this AI-human collaboration has led to a significant 26% decrease in warranty costs.

The six-week curriculum culminates in participants launching AI agents into production.

This academy ensures that all employees are equipped to adapt to and effectively utilize AI-driven systems within the company.

The program reflects Kavak's commitment to embedding AI into every facet of its operations.

Kavak maps its organizational structure as human teams supported by agents, ensuring that workers are directly integrated into workflows rather than simply receiving tasks from generic support queues.

When an AI agent encounters a problem or 'hits a wall,' it calls an API to request human assistance, preventing broken feedback loops and ensuring continuous improvement.

This human intervention is crucial for training agents to perform better in future scenarios, creating a symbiotic relationship between AI and human intelligence.

Alejandro Maza Ayala emphasized that AI adoption initiatives often fail if they are limited to bottom-up hackathons or isolated use cases.

He stated that successful AI transformation requires a top-down strategic approach where leaders define a clear vision for the company's future state within three to five years.

Organizations must operate with a unified strategy, akin to an army, rather than allowing individual tactical experimentation to guide AI implementation.

Core business functions occupy the highest tier of Kavak’s token framework, ensuring that every expenditure produces a measurable return on investment.

Tier 3 tokens, representing the highest value, are those used by agents performing core business functions, allowing for measurable return on investment.

Tier 2 tokens are indirectly measurable, typically used by developers within the codebase, while Tier 1 tokens, the lowest value, are associated with general plug-and-play chatbots that lack clear performance evaluation.

Kavak made a bold decision to abandon two years of successful work on multi-agent systems built with workflows and graphs, opting instead to destroy their existing infrastructure.

The company recognized that the increased intelligence of current AI models made graph-based constraints unnecessary, prompting a pivot to a new virtual machine architecture.

This new architecture is designed to support recursive self-improvement by providing agents with direct access to memory, evaluations, and a command-line interface (CLI).

Maza Ayala cited Joseph Schumpeter’s theory of creative destruction, explaining that innovation often leads to the replacement of old companies by new ones rather than the transformation of incumbents.

He argued that large companies face significant challenges in fully embracing AI because it demands the destruction and rebuilding of decades of legacy systems.

This provides a massive advantage to new AI-native companies, which can build their operations entirely around the strengths of new technology from the ground up.

Kavak draws a parallel between modern AI adoption and the historical shift to electricity, noting that Ford's production technology existed 40 years before widespread adoption.

Initially, companies merely replaced coal engines with electric ones, gaining only a 6% increase in efficiency, before realizing the need to redesign entire factories for electricity to achieve a 3x productivity boost.

Many contemporary companies are repeating this error by using AI superficially instead of fundamentally redesigning their entire organizational structures to fully leverage its potential.

He urged founders to 'build deep' by envisioning a future entirely centered around the AI paradigm.

He suggested that even assuming linear improvements in AI trends would lead to massive value creation across industries.