What are the key takeaways from this a16z Podcast episode?

Key insights include: **Agent-per-customer architecture:** Rather than building task-based agentic workflows, assign one persistent agent per customer with its own virtual machine, long-term memory of all past interactions, and a single goal — maximizing lifetime value. Kavak instantiates 100,000–200,000 such agents daily, each running anywhere from three minutes to three days before setting a next-task alarm.; **Evals as the accelerator, not the brake:** Allocate equal engineering time, tokens, and budget to building evals as to building the agents themselves. Measure only outcomes that matter — did the customer convert, did they reengage — not proxy metrics like call duration. This discipline is what allows Kavak to run agents across 96% of all interactions safely.; **Token ROI tiering framework:** Categorize AI spend into three tiers: Tier 3 tokens go to production agents with measurable per-token ROI; Tier 2 tokens fund developer tooling with indirect measurable value; Tier 1 tokens are untracked employee usage with no feedback loop. Shift budget aggressively toward Tier 3 to generate compounding organizational improvement rather than diffuse adoption.

What did Alejandro Ayala discuss on a16z Podcast?

Alejandro Maza Ayala, Chief Product and AI Officer at Kavak, details how the Latin American used car platform rebuilt its entire organization around AI agents — deploying 100,000–200,000 agents daily, achieving 96% agent-handled interactions, and converting customers at 2.1x the rate of human sales teams. Key topics include: **Agent-per-customer architecture:** Rather than building task-based agentic workflows, assign one persistent agent per customer with its own virtual machine, long-term memory of all past interactions, and a single goal — maximizing lifetime value. Kavak instantiates 100,000–200,000 such agents daily, each running anywhere from three minutes to three days before setting a next-task alarm.; **Evals as the accelerator, not the brake:** Allocate equal engineering time, tokens, and budget to building evals as to building the agents themselves. Measure only outcomes that matter — did the customer convert, did they reengage — not proxy metrics like call duration. This discipline is what allows Kavak to run agents across 96% of all interactions safely..

How long is this episode of a16z Podcast?

This episode is 37 minutes long. SignalCast provides an AI-generated summary so you can get the key insights in about 3 minutes.

a16z Podcast

The Self-Improving Company | Kavak's AI Playbook

August 10, 2026

37 min episode · 2 min read

Alejandro Ayala

Episode

37 min

Read time

2 min

Topics

Startups, Fundraising & VC, Leadership

AI-Generated Summary

Published Aug 10, 2026

Key Takeaways

✓Agent-per-customer architecture: Rather than building task-based agentic workflows, assign one persistent agent per customer with its own virtual machine, long-term memory of all past interactions, and a single goal — maximizing lifetime value. Kavak instantiates 100,000–200,000 such agents daily, each running anywhere from three minutes to three days before setting a next-task alarm.
✓Evals as the accelerator, not the brake: Allocate equal engineering time, tokens, and budget to building evals as to building the agents themselves. Measure only outcomes that matter — did the customer convert, did they reengage — not proxy metrics like call duration. This discipline is what allows Kavak to run agents across 96% of all interactions safely.
✓Token ROI tiering framework: Categorize AI spend into three tiers: Tier 3 tokens go to production agents with measurable per-token ROI; Tier 2 tokens fund developer tooling with indirect measurable value; Tier 1 tokens are untracked employee usage with no feedback loop. Shift budget aggressively toward Tier 3 to generate compounding organizational improvement rather than diffuse adoption.
✓Top-down transformation with mandatory reskilling: AI adoption fails when it is bottom-up or hackathon-driven. Kavak's Jedi Academy trains every employee — mechanics to executives — in a six-week program that ends with each participant deploying a production AI agent. The program requires continuous curriculum updates because the technology changes faster than any external institution can track.
✓Destroy working systems to capture architectural step-changes: When Claude Opus 4.5 released, Kavak dismantled two years of functioning multi-agent graph architecture that had already driven profitability, replacing it with flat virtual-machine agents that leverage raw model intelligence directly. Constraining a more capable model inside an older harness limits returns; rebuilding unlocks recursive self-improvement at the organizational level.

What It Covers

Alejandro Maza Ayala, Chief Product and AI Officer at Kavak, details how the Latin American used car platform rebuilt its entire organization around AI agents — deploying 100,000–200,000 agents daily, achieving 96% agent-handled interactions, and converting customers at 2.1x the rate of human sales teams.

Key Questions Answered

•Agent-per-customer architecture: Rather than building task-based agentic workflows, assign one persistent agent per customer with its own virtual machine, long-term memory of all past interactions, and a single goal — maximizing lifetime value. Kavak instantiates 100,000–200,000 such agents daily, each running anywhere from three minutes to three days before setting a next-task alarm.
•Evals as the accelerator, not the brake: Allocate equal engineering time, tokens, and budget to building evals as to building the agents themselves. Measure only outcomes that matter — did the customer convert, did they reengage — not proxy metrics like call duration. This discipline is what allows Kavak to run agents across 96% of all interactions safely.
•Token ROI tiering framework: Categorize AI spend into three tiers: Tier 3 tokens go to production agents with measurable per-token ROI; Tier 2 tokens fund developer tooling with indirect measurable value; Tier 1 tokens are untracked employee usage with no feedback loop. Shift budget aggressively toward Tier 3 to generate compounding organizational improvement rather than diffuse adoption.
•Top-down transformation with mandatory reskilling: AI adoption fails when it is bottom-up or hackathon-driven. Kavak's Jedi Academy trains every employee — mechanics to executives — in a six-week program that ends with each participant deploying a production AI agent. The program requires continuous curriculum updates because the technology changes faster than any external institution can track.
•Destroy working systems to capture architectural step-changes: When Claude Opus 4.5 released, Kavak dismantled two years of functioning multi-agent graph architecture that had already driven profitability, replacing it with flat virtual-machine agents that leverage raw model intelligence directly. Constraining a more capable model inside an older harness limits returns; rebuilding unlocks recursive self-improvement at the organizational level.

Notable Moment

Kavak carved out an entire Mexican city, Cuernavaca, and placed an AI agent in the CEO role as a live experiment. Within six weeks, the agent increased city-level profits by 50% — sending daily task plans to physical workers and requesting voice-note progress updates in return.

Know someone who'd find this useful?

Episode Transcript

I'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. The most ambitious companies listening to this will decide to follow suit, which is you decided to build an agent per customer. Yes. Every day, between a hundred and two hundred thousand agents get instantiated specifically for this customer with its own virtual machine. There's a lot of people worried about how the organizations of the future are gonna look like and the role that humans are gonna play. If you haven't faced fear before you haven't felt it, then you haven't tried AI. We launched a program inside Quebec that's called the Jedi Academy. From the CEO to, like, AI engineers to mechanics, we train everyone. And after six weeks, they launch state of the art agents to production. What advice do you have to future founders or first time founders that might be listening? What works right now is Most companies are asking how to add AI to the organization. Kavaca asked a much more radical question. What would we build if we were starting the company from scratch with AI? Angela Strange and Gabriel Vasquez sit down with Kavac's chief product and AI officer Alejandro Ayala to unpack what happened when the company bet on rebuilding itself around agents. Today, hundreds of thousands of agents can be instantiated each day, handling everything from selling and financing cars to maintaining long term customer relationships. They discuss why Kovac tore down an agent architecture that was already working to start again, how Evals became the foundation for moving faster, and what happens when agents don't just work for humans, but humans sometimes work for agents. Welcome back to the ACC podcast. Today, we have Ale Masa, the head of AI at Quebec. We're gonna discuss today the transformation that Ale led within Quebec to turn into an AI native company. Thank you, Ale, for being with us today. Thanks for having me. Before starting at Quebec, you were running a company called Oppy Analytics. That's right. And you were very much into AI before Chargegbt. Yes. Wanna tell us a little bit about that journey? Yes. Yes. Of course. Well, we called it machine learning back then. It was a different family of algorithms. And we founded a company with this very, like, ambitious vision there that new machine learning models would be so powerful that they could solve any complex problem. This was pre transformers. Right? This was like 2013. So we started building the company that way, and I think we were like ten years ahead of time, but we built a great company. We served fourteen, five hundred companies around like risk algorithms, logistics, forecasting, marketing. But really the power of what transformers and then the chat deputy moment when he arrived make things like very clearly that we could now build …

Get the full transcript (6,229 words) + summary by email — free

One-time email with the complete transcript and AI summary of this episode. No account needed.

One email, no spam. We’ll also show you what SignalCast does.

Browse all a16z Podcast transcripts →