Kavak's bet: rebuilding a used-car giant around an agent per customer
Key takeaways from the a16z podcast interview with Kavak Head of AI Ali Masa (August 2026):
An agent per customer, not per task. Kavak spawns a dedicated AI agent with its own virtual machine for each customer, instantiating between 100,000 and 200,000 agents a day. Each remembers years of interaction history, sets a long-term goal to maximise a customer's lifetime value, and wakes for anything from three minutes to three days before setting an alarm for its next task.
Redesigning the company, not bolting on AI. Masa argues most firms are stuck in a "6% trap" — handing staff ChatGPT while leaving their structure intact. Kavak instead rebuilt its APIs and systems around agents and future model capability, shifting from a transactional car-trading business to a relational one managing 10 million customers in its database.
Superhuman sales agents. Rather than customer-support bots, Kavak built sales agents that combine the expertise of 15 separate specialist teams — financing, insurance, trade-in valuation and advisory — into one. NPS and customer-satisfaction scores tripled, and the agents now convert 2.1x better than the company's best human hires.
Evals get equal billing with agents. Kavak spends roughly the same engineer time, tokens and money building evaluations as building the agents themselves, treating evals as the "brakes" that let it move fast. The core metric is business conversion — did the customer buy, get approved, re-engage — not superficial KPIs like call volume.
Agents at the scale of the whole business. Around 96% of customer interactions and 95% of transactions are now handled entirely by agents, including regulated financial products: car loans that once took two months in Mexico are approved in under three minutes, alongside personal loans with deeply personalised pricing and risk.
A pilot "AI CEO". Kavak carved out its Cuernavaca operation and installed an agent as CEO, tasked with doubling monthly profit; after six weeks it delivered a 1.5x increase while micromanaging inventory, financing penetration and daily worker plans via voice notes. Physical, dexterity-heavy roles such as its ~800 Mexican mechanics keep humans, paired with an "El Mike" sidekick that cut warranty rates by around 20-26%.
Retraining the whole workforce. Through an internal "Jedi Academy" programme, everyone from the CEO to mechanics is trained over six weeks to ship production AI agents, producing very flat, senior, cross-functional teams where agents and humans sometimes swap who reports to whom.
Creative destruction favours new builders. Invoking Schumpeter and the 40-year lag between the dynamo and Ford's redesigned factory, Masa contends incumbents will only adopt AI superficially, leaving the deepest gains to founders willing to rebuild a company from scratch around cheap, abundant intelligence.
Read more: a16z Podcast
