Kavak's AI Sales Agent Converts at 2.1x the Rate of Human Reps

Source: a16z | Published: 2026-08-10T14:30:04Z

Kavak's AI sales agent now converts at 2.1x the rate of its human team, handling financing, vehicle selection, insurance, and delivery end-to-end within a single agent.

Alejandro Maza, known as Ali, is Kavak's head of AI. Before this, he founded a machine learning company — in 2013, before pretrained models existed. He describes that experience as "ten years too early." He no longer feels early.

One Agent per Customer

Kavak is Latin America's largest used car platform — buying, refurbishing, selling, and financing, all under one roof. That business logic shapes its AI architecture: not a unified customer service bot, but a separate agent instantiated for each individual customer, running on its own virtual machine.

The agent remembers every interaction the customer has had with Kavak over the years — a call from two years ago, the model pages they browsed. Its job isn't to answer questions. It's to develop a long-term strategy: with the explicit goal of maximizing that customer's lifetime value, proactively planning the next move.

The critical design choice here is "long-term objective" over "workflow." Ali's view is that most companies have turned agents into process automation — stringing together existing steps with AI — which is just wrapping an old architecture in a new layer. Kavak's bet is that a continuously running agent with a clear long-term goal outperforms any carefully designed workflow.

The Agent Outsells the Humans

Kavak never set out to use AI for customer service. They went straight to sales.

Selling a used car in Latin America requires coordinating at least 15 steps: vehicle recommendations, loan approval, insurance, trade-in valuation, delivery logistics. In 2020 and 2021, Kavak's approach was to route customers through 15 different specialists, each handling one piece of the puzzle.

When AI came online, Kavak's approach was to first beat the corresponding human expert at each individual task, then consolidate all those capabilities into a single agent — a composite expert fluent in financing, vehicle selection, insurance, and trade-ins all at once.

The result: agent conversion rates beat the human sales team by 50%. That number is now 2.1x. NPS and customer satisfaction have tripled.

Evals Take as Long to Build as the Agents

This is what Ali emphasizes most — and where he thinks most companies get it wrong.

His go-to analogy is brakes: how fast a car can go depends on how good its brakes are. Most companies deploy AI slowly not because the models aren't good enough, but because they lack a strong evaluation framework. Kavak's principle: spend as much engineering time, compute, and money building evals as building the agents themselves.

They track one core metric: did the customer convert? Not call duration. Not response speed. Did this customer ultimately buy a car, get a loan, or sell their car to Kavak? Other metrics can inform, but they don't drive decisions.

For financial products, Kavak can complete an auto loan approval in as little as three minutes — a process that takes two months through traditional channels in Mexico. The speed comes from data integration, but what makes that speed reliable is a continuously refined evaluation system.

They Made an AI the CEO of a City

Six weeks ago, Kavak ran an experiment in the city of Cuernavaca, Mexico: they placed an agent in a management role, gave it CEO-level authority, and let it independently run that city's entire business operation.

The first month's target was to double profit. The agent didn't hit it — but it achieved 1.5x. It managed inventory turnover, financing penetration rates, and forecast accuracy. It sent direct messages to workers on the floor, laying out their daily assignments and asking them to report back via voice messages.

Customer satisfaction rose. Inventory quality improved. Every KPI moved in the right direction.

Ali's read on the experiment is blunt: the roles considered "last to be touched by AI" — strategic management positions requiring judgment — may be displaced far sooner than anyone expects. What's left for humans, increasingly, is work that requires physical presence and manual operation.

Engineers Get Retrained, Then Work Alongside Agents

To navigate this shift, Kavak launched an internal program called the Jedi Academy. From the CEO to AI engineers to frontline mechanics, everyone participates. The training runs six weeks, and at the end, every participant must deploy a real AI agent to production.

Ali designed the curriculum himself — and keeps updating it, because the pace of AI change makes any fixed syllabus obsolete quickly. His message to the team is direct: this is the direction Kavak is going. You can choose to keep up, or you can choose to leave. The company isn't going to wait.

Mechanics have their own copilot system, internally called El Mike — like the rat in Ratatouille, hiding under the chef's hat and guiding every move. It doesn't replace the mechanic's hands. It tells them how to run diagnostics, offers recommendations, and logs every step. The result: faster inspections, higher repair quality, and warranty complaints down roughly 20–26%.

When Opus 4.5 Arrived, They Threw Out Two Years of Work

This is the most counterintuitive detail in the whole conversation.

Before Opus 4.5 launched, Kavak had already built a multi-agent system that was running reliably and driving company profitability — tens of thousands of agents operating across a complex orchestration graph. That system had powered their scale-up. It had driven real growth.

But when Ali saw Opus 4.5, he reached a conclusion: the architecture itself had become the constraint. The new generation of models no longer needed such elaborate orchestration — the complexity was actually limiting what the model could do.

They decided to tear it down and rebuild. The new architecture's core: one agent, one virtual machine, memory, an evaluation system, and command-line access to every company API and tool. No intricate orchestration graph. Just a long-term goal and enough tools.

The logic: if a more capable model arrives every few months, there's no point engineering an architecture around any particular model. Better to build a general framework that automatically benefits from whatever comes next.

Most Companies Just Swapped the Engine. Kavak Demolished the Building.

An analogy Ali keeps returning to is worth recounting in full.

In 1879 and 1881, the key technologies for electric motors had already been invented. Edison began commercializing electricity in New York and London; his generators were extraordinarily efficient. Technically, Ford's assembly line could have been built 40 years earlier.

But factory owners at the time kept their four-story buildings intact and simply swapped the coal engine for an electric one. The efficiency gain was roughly 6%.

The right move would have been to tear down the building, find a flat piece of land on the outskirts of town, and redesign the entire production process around small electric motors. That approach yielded a threefold increase in productivity — the gains that underpinned America's entire industrial growth through the twentieth century.

The computer era repeated the same story. The AI era is repeating it again. Most companies today are handing employees access to ChatGPT, or automating their existing workflows with AI — a 6% improvement, not a 10x one.

Kavak is doing the demolish-and-rebuild version. Which is also why Ali believes new companies have a structural advantage over incumbents: an established CEO has to convince a board to tear down something it spent 40 years building. That almost never happens in practice. A startup can just design for what 2035 looks like from day one.