a16zKavak's Playbook for Rebuilding a Company Around AI
At a glance
WHAT IT’S REALLY ABOUT
Kavak rebuilt its org around AI agents per customer
- Kavak shifted from “AI adoption” to a full company redesign so agents could operate directly through rebuilt APIs, tools, and feedback loops.
- The core architectural bet is an agent-per-customer model where 100k–200k dedicated agents are instantiated daily, each with memory, tools, and long-term LTV goals.
- Kavak claims agents now handle ~96% of interactions and ~95% of transactions, with sales-focused agents achieving 2.1x higher conversion and significantly higher NPS.
- For regulated fintech workflows, Kavak uses evaluation systems (“brakes”) and outcome-based metrics (conversion, satisfaction, re-engagement) to safely iterate at speed.
- Organizationally, Kavak retrained the workforce via an internal “Jedi Academy,” flattened teams, and redefined human roles as building skills for agents or executing physical-world tasks with AI sidekicks.
IDEAS WORTH REMEMBERING
5 ideas“Adopting AI” rarely creates step-change value; redesigning the company does.
Maza argues that simply giving teams ChatGPT/Claude keeps the same structure and customer pain points; real gains require rebuilding systems and APIs so agents can act, not just assist.
Agent-per-customer enables long-horizon personalization beyond task automation.
Each customer gets a dedicated agent with persistent memory and a strategy to maximize lifetime value across products over time, turning a transactional business into a relational one.
Evals are the enabling constraint that lets you move fast safely.
Kavak treats evals like “brakes,” investing roughly equal engineering time/tokens/money in evaluation as in building agents, and tying evals to business outcomes rather than superficial activity KPIs.
Sales agents can outperform humans when they unify many expert roles.
Kavak built agents to sell (not just support) by combining financing, insurance, trade-in, and inventory guidance into a single “mega-expert,” reporting 2.1x conversion and tripled satisfaction/NPS improvements.
Fintech automation works when you pair rich data with tight feedback loops.
They cite car-loan approvals in under three minutes and highly personalized pricing/limits, emphasizing trust-building over a multi-month customer decision cycle and continuous learning from real interactions.
WORDS WORTH SAVING
5 quotesI'm investing more today in tokens than in knowledge workers. We could build superhuman agents. This means that by every dimension that matters, our agents would outperform the best human we had ever hired.
— Alejandro Maza Ayala
Every day, between 100 and 200,000 agents get instantiated in a day. They wake up, they work, sometimes for three minutes, sometimes for eight hours, sometimes for three days, and they, like, set an alarm clock for their ner- next task, and to go back to sleep.
— Alejandro Maza Ayala
So, a good rule of thumb here is we spend about the same amount of time, engineer time, tokens, and, and, and money on building the evals than building the agents.
— Alejandro Maza Ayala
We tripled, uh, NPS and customer satisfaction score by putting the, the agent in front of the, of the customer, and it, at first it converted, like, 50% more than our human team, and now it's converting over that, like, 2.1, uh, X more.
— Alejandro Maza Ayala
We decided to, like, destroy everything we had been building for, for two years that was working, that brought us to profitability, that brought us amazing growth, and start over with a harness that we thought would be robust and scalable and, and leverage recursive self-improvement.
— Alejandro Maza Ayala
High quality AI-generated summary created from speaker-labeled transcript.
