How to Rebuild a Company around AI @Alehandromz (Alejandro Maza), Chief Product & AI Officer, Kavak, interviewed by @astrange (Angela Strange) and @gabe_NM (Gabriel Vasquez) (@a16z Podcast) Summary: Kavak, the Latin American used-car marketplace, now runs 96% of customer interactions and 95% of transactions through AI agents, and spawns between 100,000 and 200,000 of them every day. Alejandro Maza rebuilt the APIs, the teams, the metrics and the customer journey around agents. Most companies hand employees a chat tool and keep the org chart. Maza then threw his whole architecture out and started over when a better model made it obsolete. 1. The Six Percent Trap. The technology for Ford's production line existed 40 years before Ford built it. Edison was selling electricity in New York by 1881 with a very efficient dynamo, and factory owners kept their four-story buildings, their shafts and their belts, and swapped the coal engine for an electric one. That bought them about 6% efficiency. The 3x came from demolishing the factory, moving out of the city and redesigning the floor around small distributed motors, and Maza says every CEO faces that same choice right now. 2. An Agent Per Customer. Kavak's design question was how it would build the company in 2035 with GPT-10 level intelligence. Every customer who arrives gets an agent spawned for them alone, with its own virtual machine, memory of every web visit and phone call going back years, and a long-term goal of maximizing that customer's lifetime value. Between 100,000 and 200,000 of these wake up each day, work for three minutes or eight hours or three days, set an alarm for the next task and go back to sleep. Most companies are still wiring up multi-agent expert systems, and Kavak bet on long-running agents with hard goals. 3. One Agent, 15 Experts. Selling a used car in Latin America means being excellent at 15 things at once: 20,000 SKUs, financing, insurance, coverage and quoting the trade-in. The old model routed a buyer through 15 human experts on 15 teams, so Kavak built an agent better than the expert at each piece and fused them into one. Kavak never built a customer service agent, only sellers, and pointed them at the hardest job in the company. Customer satisfaction tripled, and the agent first converted 50% better than the human team, then climbed to 2.1x. 4. Evals Are The Brakes. Kavak spends roughly the same engineer time, tokens and money on evals as on the agents themselves. Maza's rule of thumb is that you press the accelerator in proportion to your brakes, and the companies moving slowly on AI are the ones without them. What gets measured is the business result: did the customer convert, did the loan get approved, will they come back. He watches other companies count calls and minutes-per-call, which does not tell you whether the thing worked. 5. The Three Minute Loan. A car loan in Mexico typically takes two months or more to approve. Kavak approves in under three minutes, because it holds the data on both the customer and the car. Vertical integration also gives it a move a bank does not have: if the customer can no longer pay, they return the car and get a cheaper one with a smaller payment. Agents run the underwriting, the pricing and the servicing end to end, on customers with thin files or none, down to the interest rate, the risk tier and the maximum loan for that individual. 6. The AI CEO. Kavak carved out the city of Cuernavaca and put an agent in charge of it. The goal for the first month was to double profits, and six weeks in it has delivered 50% more. It works the way a very smart and very tireless person would, reading every number and every customer, building the forecast, then micromanaging the daily plan and messaging the physical staff, who send voice notes back on their progress. Customer satisfaction, inventory rotation and financing penetration all improved, in the one job everyone assumed would go last. 7. Ten Million Relationships. Kavak used to count cars bought, cars sold and brake pads ordered. It now counts 10 million customers in a database, most of them with an agent assigned and a mandate to grow that relationship over years. On a book of cars and large personal loans, activating 1% of that base is worth hundreds of millions of dollars. The economics work because a used car is a trust purchase, and trust comes from an agent that remembers a conversation from two years ago. 8. Token Tiers. Maza sorts AI spend into three tiers by whether he can see the return. Tier three, the most valuable, is tokens going into agents doing the actual work of the business, where the ROI of each token is measurable today. Tier two is indirect, like watching developers in the codebase and inferring the value before pushing it to production. Tier one is everyone on Claude Code or ChatGPT or Cowork with no idea what came of it, which is where most companies spending hundreds of millions currently sit. 9. Humans Behind The API. Most production agent systems escalate to a tier-2 human queue and forget about it, so the loop never closes and the agent never learns. At Kavak the agent calls an API asking for help, and a person is on the other side of that call. Draw it on an org chart and it is human teams reporting to an agent. When a mistake gets corrected, the other 200,000 agents have it by the next day. 10. The Jedi Academy. Kavak trains everyone from the CEO to the mechanics to build agents, in a six-week program Maza designed and teaches himself, and graduates ship agents to production. He rewrites the curriculum constantly because the material keeps changing, and there is nowhere to send people to learn it. The message to staff was direct: this is where Kavak is going, you can learn the skills to work in it, or you can leave. Some became AI engineers, and everyone learned to work alongside the technology. 11. Destroy What Works. By December Kavak had tens of thousands of agents in multi-agent graphs running the business profitably. Opus 4.5 shipped and Maza concluded the architecture had become the constraint, because the graph structure was capping a level of intelligence that no longer needed it. He threw out two years of work that was producing growth and profit, and rebuilt on a virtual machine with memory, evals, a CLI and access to every tool and API in the company. His advice to anyone starting now is to skip agentic workflows and graphs entirely. 12. Creative Destruction. Schumpeter's point was that new technology reaches the economy by destroying incumbents. Maza's read is that almost no public-company CEO will stand up and say they are demolishing 40 years of accumulated systems to rebuild as an AI-native company, so they will adopt superficially and take the 6%. That leaves the deep rebuild to companies that do not exist yet. His closing argument to founders is that the most powerful intelligence in the world now costs $20 a month, so draw the trend line out and build for where it lands.