AI Transformation: Inside the Kavak Rebuild

An AI transformation is a redesign of how a company works, not a rollout of AI tools inside the company you already have. Adoption gives everyone a chat window and leaves the APIs, the org chart, the metrics and the training exactly where they were. Transformation changes those four things to match what agents can now do, and accepts that some of what already works has to come apart.

Kavak, the Latin American used-car marketplace and Mexico's first unicorn, rebuilt rather than adopted. Agents now handle 96 percent of its customer interactions and 95 percent of its transactions. Alejandro Maza, Kavak's Chief Product and AI Officer, took a16z's Angela Strange and Gabriel Vasquez through the rebuild. One caveat to hold while you read the numbers: a16z led Kavak's $300 million Series F in February 2026, so this is an investor interviewing a portfolio company, and every figure below is Kavak's own reporting as of the August 2026 interview rather than an audited result.

The line a16z opens the episode with compresses the whole thesis: "I'm investing more today in tokens than in knowledge workers."

What follows is the order he did it in, what he measured, and the parts a startup should copy rather than admire. If you are building this way from day one, the companion piece is how to build an AI-native company. This page is the harder case: a business that already works, which you now have to take apart on purpose.

AI adoption is not AI transformation

Maza's diagnosis is specific about where the stall happens. Companies adopt, keep the structure, and then find no efficiency arrives, because customers are still hitting the processes they hit before. The fix, he says, is to "redesign your whole company around the agents and around the future capabilities", which in practice means rebuilding most of your APIs so that agents can use them to get work done.

That is the dividing line. A transformation shows up as a different system diagram, not a different expense line.

What the rebuild bought

Kavak's own numbers from the conversation:

• 96 percent of all customer interactions and 95 percent of all transactions handled by agents, with the human kept for the handover when the customer collects the car.
• Between 100,000 and 200,000 agents instantiated per day, one per customer, each with its own virtual machine.
• Agent sellers that converted 50 percent better than the human team at first, and now convert 2.1 times better.
• NPS and customer satisfaction tripled.
• Car loans approved in under three minutes, against the two months or more Maza says is normal in Mexico.
• An agent acting as CEO of Kavak's business in one Mexican city, six weeks into the run, coming in at 1.5 times profit against a first-month goal of doubling it.

The AI CEO is the detail everyone repeats and the least useful one. The architecture underneath it is the transferable part, and Maza describes it as three bets.

Bet one: one agent per customer, not one agent per task

Kavak did not start from the question of where AI could help. It started from a harder one: what would this company look like if it were designed in 2035, for 2035 levels of model intelligence. Then it built that.

Concretely, when a customer arrives an agent is spawned for that customer with its own virtual machine. It carries years of history, a page they browsed, a call from two years ago. It sets a long-term goal of maximizing that customer's lifetime value and works toward it across products and across months. "they work sometimes for three minutes, sometimes for eight hours, sometimes for three days," Maza says, and then set an alarm for their next task and go back to sleep.

Getting there took demolition. For two years Kavak ran tens of thousands of agents arranged in multi-agent graphs, a system that worked and that Maza credits with carrying the business to profitability. A new frontier model release convinced him the graph itself had become the ceiling on the intelligence inside it, so the team threw out two years of working architecture and rebuilt on one agent per customer, with memory, evals and a command line into every tool and API in the company. His advice to anyone starting now is to skip the stage he spent two years in and not build agentic workflows at all.

Bet two: agents that beat your best hire, and the evals that prove it

Bet two was that agents could beat the best person Kavak had ever hired on conversion, lifetime value and customer experience. The only route there is to put them in front of real customers, collect the data, build the evals, and train on the result.

The brake metaphor is the part worth keeping. "I like to move extremely fast, but in order to move fast, you need to have brakes," Maza says. You press the accelerator because the brakes are good, and the companies crawling on AI are usually the ones without brakes. Kavak spends roughly as much engineer time, token budget and money on the evals as on the agents themselves, which is the opposite of treating them as an afterthought.

What it measures matters as much. "the important thing is did this customer convert? Is it bringing value to the customer and is the customer happy to re-engage with us after a while?" Call counts and average call length are the superficial version, and Maza says that is where he sees most companies break.

The people who sell eval infrastructure describe the same priority, and Kavak's spend ratio is the operational version of it. If you have not built that layer yet, start with LLM evaluation for founders, which is built on the case Ankur Goyal, founder and CEO of Braintrust, makes for putting evals at the centre of an AI product.

Maza's account of why the agents win is structural rather than magical: "they're experts and they're infinitely patient and they know all your history", they plan for the long term, and they never get tired. Then the compounding part, which is the real argument for one shared architecture rather than a fleet of separately tuned bots: "if they make a mistake, they learn it and the next day, not just them, but the other 200,000 agents will have learned from that mistake".

Bet three: change what the company measures

Kavak used to be a transactional business, scored on cars bought, cars sold and brake pads ordered. The transformation turned it into a relational one: roughly 10 million customers in the database, an agent assigned to most of them, each carrying a lifetime-value goal. At Kavak's ticket sizes, Maza's point is that activating even 1 percent of that base is worth hundreds of millions of dollars.

This is the bet most transformations skip, and skipping it is why their AI shows up as cost savings instead of growth. The metric change is what makes the architecture pay for itself: an agent per customer only earns its virtual machine if the company is scored on customer lifetime rather than transaction count. Scoring on lifetime value only works if customers come back, which is a product question before it is an agent question: see product stickiness.

Not every token is worth the same

Maza's token framework is the most directly portable thing in the conversation. He sorts spend into three tiers:

• Tier three, the most valuable: tokens going to agents that perform the organization's actual work, where the return on each token is directly measurable.
• Tier two: tokens you can evaluate indirectly, such as developers working in the codebase, where you can see the output and push it to production.
• Tier one, where most companies sit: people using a chat tool or a coding agent, with no idea what the spend produced.

"It's really about having a very clear vision and then measuring that each token you spend is bringing you those benefits," he says. A company announcing a large token bill has said nothing until it says which tier those tokens were in. The startup-side version of this argument, including when heavy spend is a real edge and when it is burn, is in tokenmaxxing.

Train everyone, or the transformation stops at the org chart

Kavak took the fact that everyone's job would change seriously three years ago, and answered it with a six-week internal program it calls the Jedi Academy. Maza designed it and taught it himself, and keeps rewriting it because the material goes stale in months and there is nowhere external to send people to learn it. The intake is deliberately everyone: the CEO, AI engineers, finance people, mechanics. "we train everyone and after 6 weeks, they launch state-of-the-art agents to production," he says. Not everyone becomes an AI engineer. Everyone learns to work with the systems.

He was also blunt about the deal on offer: this is the direction Kavak is going, here is what changes for each team, you can pick up the skills to work in that reality or you can leave.

Daniel Dines, co-founder and CEO of UiPath, framed the same obligation from the other direction on 20VC: "I've never hidden from my employees that there will be a transformation in the company. But I told them up front, guys, we are not doing anything stupid. We are not just using AI as a pretext to cut a part of the company." Dines published a book on the question, The Work That Remains, in August 2026.

The organization that came out the other side is "very flat teams, very senior teams, super empowered", each team mixing engineering, AI and operations, and each doing one of three things: building the agents, working for the agents, or standing in the physical world in front of the customer.

The escalation design is worth copying at any size. The standard human-in-the-loop pattern kicks a stuck case to tier two support and loses it. "That doesn't really work because you don't close the loops." At Kavak the agent keeps the customer and calls a help API when it hits a wall, and a human answers that call. Because the agent never hands the case off, the resolution becomes training data instead of disappearing into a queue.

The physical work is where Kavak still hires and trains people. Maza puts the mechanic count in Mexico at around 800, kept because dexterity and physical senses are the hard things to replace. They get a sidekick agent on the same architecture, which he compares to the mouse in Ratatouille, that coaches them through an inspection. By his account inspections got faster, repairs got cheaper, warranties came down, and customer satisfaction went up again.

Why it has to come from the top

Asked what he tells the large-company leaders who call him, Maza gave two answers.

First, "it has to be top down". Bottom-up adoption does not produce a company, because the taste and the strategy for deciding what to build cannot be crowdsourced upward. The hackathon that generates use cases and sponsors a few of them is the specific failure he names. The instruction instead is to "be very clear on what the company will look like in three or five years and then start building that", and then to be vertical about pointing everyone at it. His analogy: "an army doesn't really work if everyone comes up with ideas on the strategy and tactics and goes to the battlefield and like does whatever they want". The founder-side version of the same mandate is the AI-first company lessons from Brex, where the Brex CEO treats it as part of the job description. Maza's version is the harder one, because he has to move an org chart and a physical operation that a 40-person startup does not have yet.

Second, measure by tier. A leader who cannot say which tier the company's tokens sit in does not yet know whether the transformation is running, and that is a question only the person who owns the budget can force.

There is a security reading of top-down that Maza does not make and that you should. An agent per customer with a command line into every API in the company is, structurally, a few hundred thousand daily principals holding broad authority and a long memory of your customers. Three decisions make that safe: scoped credentials per agent, the help API as an auditable escalation path, and evals that score customer outcomes rather than tool calls. They are the same three that keep the fleet from becoming your largest insider-risk surface. Those choices cannot be made team by team, which is a second reason the mandate has to sit with leadership rather than with whoever ships fastest.

The part founders should steal

Maza's closing argument matters most to people who do not run Kavak: AI transformation is a startup opportunity precisely because it is so hard for incumbents to do. He reaches for creative destruction, and reads Joseph Schumpeter's term the way founders tend to: innovation reaches the economy less through incumbents adopting the new technology than through new companies built around it displacing the ones that did not. Very few CEOs of large or public companies will stand up and say they are going to dismantle 40 years of accumulated process to rebuild around AI. He calls it the innovator's dilemma at industrial scale, and the book-length version is worth reading on its own in The Innovator's Dilemma.

The story he tells his team is about electricity. The generating stations and the efficient dynamo arrived decades before anyone built a factory that assumed them. The first wave of adopters kept the multi-storey, shaft-and-belt building and swapped the coal-fired steam engine for an electric motor, which bought a few percent. The gains that mattered came from tearing the factory down, rebuilding it on a single floor outside the city, and redesigning the whole layout around small distributed motors. Maza's number for the retrofit is 6 percent efficiency, against roughly 3 times the productivity for the factories that were rebuilt around electricity. He puts today's version of the same split at a 6 to 10 percent improvement for the companies that adopt superficially, against something closer to 10 times for the ones that redesign. "they're not willing to redesign the whole company and they just adopt it superficially."

His last reframe is the one worth taping to a wall. Everyone is obsessed with recursive self-improvement in models, but by his reading economic value has come from organizations rather than individuals for the past 4,000 years, so "what you want to self-improve and to engage in that loop is the organization that can deliver more economic value". Get that loop running and the company compounds with each model release instead of waiting on one.

His advice to first-time founders is short. This is the most exciting time in human history to start something, because the most capable tools in the world are available to anyone for roughly the price of a monthly subscription. "just go for it but go for it deep". Map a straight line of continued model improvement and build for the company that line implies, not the one you could staff today.

What to do this week

1. Write the one-page version of what your company looks like in three years, assuming models keep improving on a straight line. If you cannot write it, you do not have a transformation, you have a tool budget.
2. Take your highest-leverage customer-facing job, the one you would least trust to an agent, and put an agent in front of a small slice of it this week. That slice is where the training data comes from.
3. Audit your API surface for agent usability rather than human usability. If an agent cannot complete a transaction end to end through your own APIs, that is the work that comes before any agent work.
4. Sort last month's token spend into the three tiers. Anything sitting in tier one either moves up a tier or gets cut.
5. Replace one dashboard metric with an outcome metric: did the customer convert, did they come back. Delete a volume metric to make room.
6. Put your escalation loop on paper. When an agent gets stuck, who answers, and how does the resolution get back into the evals. If the answer is a support queue, the loop is open.
7. Name the person accountable for the transformation. If it is not a founder or the CEO, you have the bottom-up version that Maza says does not work.

The course is the structured version of all of this: AI Operating System for Startups covers the agent architecture, the evals and the metrics that tell you the rebuild is working. For the wider map of where AI changes a startup's product, engineering, go-to-market and team, start with AI for startups.

Sources

• Kavak's Playbook for Rebuilding a Company Around AI, the a16z conversation with Alejandro Maza, Angela Strange and Gabriel Vasquez, published 2026-08-10. The interview this article distills.
• Kavak's $300 million Series F led by Andreessen Horowitz (February 2026) and a16z's own investment announcement, for the investor relationship behind the interview.
• Profiles: Alejandro Maza (Chief Product and AI Officer, Kavak), Daniel Dines (co-founder and CEO, UiPath) and Ankur Goyal (founder and CEO, Braintrust).
• Daniel Dines on 20VC, on transforming the workforce at the same time as the company, and his book The Work That Remains (UiPath, August 2026).
• Ankur Goyal on Greylock, on why evals are the core of building an AI product.
• Creative destruction and Joseph Schumpeter on Wikipedia, for the framing Maza applies to AI adoption.