AI Agent Pilot Lessons from Kavak: Why a Small Team Wins

Written by | Nov 20, 2026

Kavak, the Mexico City used-car company, now runs 96% of its customer interactions and 95% of its transactions through AI agents. It reached that point by starting with a small, dedicated team that owned an AI agent pilot on 2% of its customers.

I heard Kavak’s Chief Product & AI Officer, Alejandro Maza, describe that path in the second-day keynote at Vapicon 2026 in San Francisco, in conversation with Vapi CEO Jordan Dearsley. This post covers how that small team ran the pilot and what other engineering leaders can learn from it.

A Builder Pod can staff an agentic AI pilot like this. Builder Pods are AgilityFeat’s small, dedicated teams of senior nearshore engineers in Latin America, set up for proofs of concept, agent-based products and other new builds. Kavak’s pilot, run by its own team in Mexico City, is a useful model for structuring that first project.

Where Kavak started

Kavak buys, reconditions and resells used cars. It became Mexico’s first unicorn in 2020 and reached an $8.7B valuation in 2021, then closed its operations in Colombia and Peru in 2023 as growth slowed. Leadership needed a path to profitability, which gave Maza room to test AI agents. He had joined Kavak when it acquired the ML startup he worked at.

How Kavak ran its AI agent pilot on 2% of customers

Maza began with weekend experiments using voice AI agents in 2023. When he proposed handing customer conversations to agents, Kavak’s engineering team was skeptical.

To help address their concerns and reduce risk, they decided to deploy initially with a small subset of customers. A small team took 2% of Kavak’s customers and committed fully to agents for that group: agents handled the conversations, follow-ups and sales.

They could have decided instead to use AI for a small part of interactions with all customers, but they felt that a better test was to go “all in” with a small part of the customers, and let the rest of the business run as before. This is a good change management approach, since it provided more of a true test of the agentic capabilities, without jeopardizing the whole company if it went poorly.

Expect a dip: six bad weeks, then 2.1x conversion

Maza told the team to expect a J-curve: results would drop before they improved, while the agents and the engineers both learned how to sell cars.

The initial drop was significant, and customer satisfaction in the pilot group was low for the first six weeks. But Kavak kept going, and allowed time for the engineering team to help the agents improve. According to Kavak CEO Carlos García Ottati, Kavak’s AI sellers now convert at 2.1 times the rate of its human sales teams.

AI agent evaluation gets the same budget as the agents

Maza has said Kavak puts as much engineering effort into AI agent evals, the automated tests that score how well agents do their job, as into building the agents. At Vapicon, he explained why: evals build trust in two directions. The business has to trust the agents before it gives them pricing authority. Customers have to trust the agents, because to a customer the agent is Kavak. That trust drives repeat purchases, and repeat purchases drive lifetime value.

How to pilot AI agents: start with a small team

Kavak’s pilot comes down to five decisions any company can make before its first agent goes live. The team comes first, because every other decision depends on the quality of the human team who owns the agentic pilot.

  1. Put a small team in charge of the whole pilot. Kavak gave its pilot to a small team while the rest of engineering kept running the business. Two full-time engineers who understand both the AI stack and your product can build the agents, connect them to your systems and change direction every week. Add specialists, such as voice AI, QA or infrastructure engineers, only when the pilot reaches the work that needs them.
  2. Pick a slice sized to your volume. Kavak used 2% of its customers, which was a large group at Kavak’s scale. Choose a segment big enough to produce conversion data you can trust within a few weeks, and small enough that a bad month will not show up in your quarterly results.
  3. Give agents the whole job for that slice. Kavak’s agents handled the conversations, follow-ups and sales for the pilot group. When agents own the full journey, the results show what the agents can do on their own.
  4. Budget for the dip. Kavak’s lasted six weeks. Agree in writing on how long the pilot can underperform before anyone reviews it, and fund the team for at least that long plus the recovery.
  5. Build evals in the first sprint. Measure the business outcome you care about, such as conversion, from day one. Without evals, nobody can tell whether the dip is ending or the pilot is flattening out.

A Builder Pod is built for this style of engagement: two full-time core engineers in Latin America working US hours, with part-time specialists in areas such as Voice AI, QA and DevOps added as the pilot needs them. For agent pilots, Pod Nova, our Voice AI & Innovation Builders, works on voice AI and agent-based products.

From AI pilot to production: Kavak’s rebuild

With pilot results in hand, Maza took the case to Kavak’s executives, and in 2025 Kavak rebuilt its whole system around agents. Each customer now gets an agent that works to maximize their lifetime value, and each car gets an agent that works to get the best price for it. The two negotiate each sale together, within rules such as a minimum margin.

According to Maza, about 100,000 agents now run each day. García Ottati has said that Kavak’s growth was flat for a year before it grew fourfold. Kavak reached profitability in 2025, and in February 2026 it raised a $300M Series F led by Andreessen Horowitz.

Built by AI engineers in Latin America

Kavak’s agent system was built by its engineering team in Mexico City. I heard that story on the main stage of a San Francisco AI conference, and to me it was a story about Latin American engineering talent as much as voice AI.

It matches what I have seen since I founded AgilityFeat in 2010. I started working with Latin American tech teams that year and quickly saw the chance to bring their talent to a global market. For more than 15 years, we have built engineering teams in Latin America for companies around the world.

The quality of those engineers is also why we created WebRTC.ventures, which has built video and real-time communication applications since before the pandemic. Today engineers at both AgilityFeat and WebRTC.ventures build voice AI products for clients. Alberto Gonzalez, CTO of both companies, attended Vapicon with me, and many of the speakers were describing the kind of work our engineers do every day.

Builder Pods give companies a fast way to put that experience to work. Pod Nova focuses on voice AI and agent-based products. If you are planning an AI agent pilot and need a dedicated AI engineering team to run it, reserve Pod Nova or talk to us about how to scope it.

Further Reading:

About the author

About the author

Arin Sime

Our CEO and Founder, Arin Sime, has been recruiting remote talent long before it became a global trend. With a background as a software developer, IT leader, and agile trainer, he understands firsthand what it takes to build and manage high-performing remote teams. He founded AgilityFeat in the US in 2010 as an agile consultancy and then joined forces with David Alfaro in Latin America to turn it into a software development staff augmentation firm, connecting nearshore developers with US companies. Arin is the host of the Scaling Tech Podcast and WebRTC Live.

Recent Blog Posts