a16zKavak's Playbook for Rebuilding a Company Around AI
CHAPTERS
- 0:00 – 1:03
Why Kavak is betting on “superhuman agents” over knowledge workers
Alejandro Maza Ayala opens with the core thesis behind Kavak’s transformation: investing more in AI tokens than in traditional knowledge labor. He frames the goal as building agents that outperform the best humans across business-critical dimensions.
- •Shift in spend from knowledge workers to tokens/compute
- •Definition of “superhuman agents” as outperforming top humans on key metrics
- •AI as a company-wide operating model, not a side tool
- •Sets the stakes: fear, disruption, and speed of change
- 1:03 – 2:34
From 2013-era machine learning to the ChatGPT/Transformer inflection point
Maza recounts his pre-Transformer ML background and building Oppy Analytics, emphasizing how limited older model families were compared to modern Transformers. The ChatGPT moment clarified that entirely new company designs were now possible.
- •Pre-Transformer ML as a different algorithmic paradigm
- •Oppy Analytics serving Fortune 500 use cases (risk, logistics, forecasting, marketing)
- •“10 years ahead of time” lesson: capability timing matters
- •Transformers/ChatGPT unlock building organizations differently
- 2:34 – 3:11
What Kavak does—and why LatAm forced full-stack vertical integration
Kavak began as a used-car marketplace but had to become more: refurbishing, selling, financing, logistics, and data infrastructure (Carfax-like) due to market gaps in Latin America. This vertical stack becomes a foundation for agents to execute end-to-end journeys.
- •Buy, refurbish, sell, and finance used cars
- •Built fintech + logistics + vehicle history infrastructure internally
- •Vertical integration creates rich data + controllable workflows
- •High-trust, high-ticket transactions amplify AI leverage
- 3:11 – 5:12
Agent-per-customer: long-running memory, strategy, and lifetime value goals
Instead of task bots, Kavak spawns a dedicated agent for each customer session with its own virtual machine and long-term objectives. These agents remember prior interactions, plan strategies over time, and optimize for lifetime value and customer happiness.
- •Per-customer agent instantiation with a dedicated VM
- •Persistent memory of multi-year customer interactions
- •Long-term goal: maximize customer lifetime value (not just resolve a task)
- •Departing from common multi-agent ‘expert workflow’ approaches
- 5:12 – 7:35
Three company-wide bets: redesign the org, train superhuman agents, change success metrics
Maza explains Kavak’s transformation required more than “AI adoption”—it demanded rebuilding APIs for agent use, creating feedback loops for training, and redefining metrics around relationships rather than transactions. The company reoriented toward managing millions of customer relationships via agents.
- •Bet #1: redesign company structure + systems around agents (rebuild APIs)
- •Bet #2: put agents in front of customers to learn via real feedback loops
- •Bet #3: shift from transactional metrics to relational/LTV-based measurement
- •Why it pencils out: high-ticket items + trust + 10M-customer base leverage
- 7:35 – 10:59
Operating at scale with ‘brakes’: evals as the system that enables speed
With most interactions and transactions handled by agents, Kavak treats evaluation as a first-class engineering discipline. Maza argues you can only move fast with robust evals, investing roughly equal effort in evals and agent building.
- •96% of interactions and ~95% of transactions handled by agents
- •100k–200k agents instantiated daily with varying lifetimes
- •Evals as ‘brakes’ that make rapid iteration safe
- •Allocate ~equal time/money/tokens to evals as to agent development
- •Measure real outcomes: conversion, customer happiness, re-engagement
- 10:59 – 13:35
Agents that sell cars: building a ‘mega-expert’ and beating humans by 2.1x
Kavak didn’t focus on generic support bots; it built sales agents capable of navigating a complex, multi-expert car purchase flow. By combining expertise across financing, insurance, trade-in, and recommendations, agents improved NPS and dramatically outperformed human conversion.
- •Sales-first agents (not customer support) for a complex purchase journey
- •Replace 15 specialized human experts with one AI ‘mega-expert’
- •Tripled NPS/customer satisfaction after deploying sales agents
- •Conversion gains: from +50% early to ~2.1x vs humans
- •Compounding learning: one agent’s mistake improves all agents via feedback loops
- 13:35 – 16:07
Regulated fintech workflows: underwriting and loan approvals in under three minutes
Maza details how Kavak’s agents streamline financing in markets where approvals can take months. Deep customer + asset data enables fast underwriting and personalized pricing while maintaining portfolio-level risk constraints and better customer experience.
- •Car loan approval in under three minutes vs months in traditional flows
- •Vertical integration enables creative solutions (e.g., return/swap to cheaper car)
- •Personalized loan terms: interest rate, max amount, risk calibration
- •Agents support long decision cycles (3–4 months) with relationship-driven conversion
- •Evals become crucial due to PII, money movement, and regulatory constraints
- 16:07 – 17:58
The AI CEO experiment: managing a city to +50% profits in six weeks
Kavak piloted an ‘AI CEO’ agent by giving it control over one city’s operations (Cuernavaca). The agent micromanaged metrics, forecasts, and daily execution via direct communication with on-the-ground staff, yielding broad KPI improvements and 1.5x profits.
- •Carved out Cuernavaca as a controlled experiment environment
- •Agent set profit-doubling goal; achieved 1.5x profits in ~6 weeks
- •Data-driven management: forecasts, KPI drilling, daily execution plans
- •Two-way loop with physical workforce via messages and voice-note updates
- •Outcomes improved across satisfaction, inventory, rotation, financing penetration
- 17:58 – 20:01
AI in the physical world: ‘Ratatouille’ sidekicks for mechanics and inspection quality
Where dexterity and sensing still matter, Kavak augments humans rather than replacing them. A mechanic sidekick agent (“El Mike”) guides inspections and repairs, improving speed, cost, and—most importantly—vehicle quality.
- •Physical work remains hardest to automate; humans still essential
- •Mechanics use an agent sidekick to guide inspections and procedures
- •Faster inspections and repairs; reduced costs
- •Quality gains: warranties down ~20–26% after launch
- •Higher customer satisfaction driven by better car quality
- 20:01 – 22:46
The Jedi Academy: retraining everyone to build and collaborate with agents
To address workforce disruption, Kavak created an internal training program that teaches employees—from CEO to mechanics—to ship production-grade agents in six weeks. The program supports a culture where humans collaborate with, build for, and sometimes report to agents.
- •Jobs will change; continuous reskilling is mandatory
- •Internal ‘Jedi Academy’ designed and updated by Maza
- •6-week program culminating in shipping agents to production
- •Participants include executives, engineers, finance staff, and mechanics
- •Cultural contract: adapt and learn, or opt out
- 22:46 – 28:08
Org design for an agent-run company: flat teams, humans as ‘tier-two’ problem solvers
Kavak’s structure becomes flatter and more senior, with cross-functional teams centered on building and supporting agent capabilities. Instead of agents failing and handing off to humans without learning, Kavak closes the loop so human interventions become training data and new skills.
- •Flat, senior, empowered teams spanning AI/engineering/ops/field roles
- •Humans either build agents, work for agents, or handle physical customer moments
- •Avoid ‘handoff and forget’ escalation patterns that break learning loops
- •Agents call for human help via explicit APIs when they hit a wall
- •Human help is structured to generate data and improve future agent performance
- 28:08 – 31:16
Destroying two years of work: moving from multi-agent graphs to one agent per customer
As models improved (e.g., “Opus 4.5”), Kavak concluded complex multi-agent graphs would constrain intelligence rather than unlock it. They scrapped a working system and rebuilt around a per-customer agent running in a VM with memory, tools, evals, and long-term goals—designed to benefit from future model upgrades and recursive improvement.
- •Warning against ‘agentic workflows’ and heavy graph/lattice architectures
- •Tens of thousands of agents ran the business, but paradigm shifted with better models
- •Chose to scrap 2 years of working infrastructure to avoid future constraints
- •New harness: VM + memory + CLI/tools + evals + long-term objective
- •Goal: self-improving organization that compounds value as models advance
- 31:16 – 36:31
Creative destruction and founder advice: don’t merely adopt AI—rebuild around it
Maza argues economic history shows true gains come from redesigning systems around new tech, not superficial adoption—illustrated via electricity and Ford’s factory. He closes with advice to founders: this is a uniquely democratized moment to build, but only if you go deep and design for the trajectory of AI capability.
- •Schumpeter’s creative destruction: new entrants often beat incumbents who ‘adopt’
- •Incumbent CEO constraint: hard to bet the whole company on AI transformation
- •Electricity/Ford analogy: swapping engines yields ~6% vs redesign yields ~3x
- •Token ROI framework: measure value per token, not mere adoption
- •Founder advice: build for the AI future trajectory and commit deeply
