{"slug": "the-self-improving-company-kavak-s-ai-playbook", "title": "The Self-Improving Company | Kavak's AI Playbook", "summary": "Kavak's AI agents now handle 96% of customer interactions and 95% of transactions, converting customers at 2.1x the rate of its best human salespeople, according to Chief Product & AI Officer Alejandro Maza Ayala. An AI 'CEO' experiment boosted profits by 50% in one city in its first month, and the company spins up 100,000–200,000 agent instances daily. The Latin American used-car marketplace rebuilt its tech stack around AI agents, with Ayala emphasizing that matching engineering effort on evals to agent-building is key to scaling safely.", "body_md": "# The Self-Improving Company | Kavak's AI Playbook\n\nKavak's AI agents convert customers at 2.1x the rate of its best human salespeople — and an AI \"CEO\" boosted one city's profits by 50% in a single month.\n\nThe a16z Show\n\n# The Self-Improving Company | Kavak's AI Playbook\n\nKavak's AI agents convert customers at 2.1x the rate of its best human salespeople — and an AI \"CEO\" boosted one city's profits by 50% in a single month.\n\nTL;DR\n\nKavak's Chief Product & AI Officer Alejandro Maza Ayala details how the Latin American used-car marketplace tore down its existing tech stack and rebuilt itself entirely around AI agents. Today, 96% of customer interactions and 95% of transactions are handled by agents, with 100,000–200,000 agent instances spun up daily\n[1]\n— Alejandro Maza Ayala\n\"96% interactions handled by agents: Kavak's AI agents handle 96% of all customer interactions with no human involvement.\"\n08:15\n. Kavak's AI sellers convert at 2.1x the rate of its best human team\n[2]\n— Alejandro Maza Ayala\n\"AI sellers convert 2.1x human team: Kavak's AI sales agents now convert customers at more than 2.1 times the rate of the company's best hum…\"\n12:34\n, and an AI \"CEO\" experiment boosted one city's profits by 50% in its first month\n[3]\n— Alejandro Maza Ayala\n\"AI CEO raised profits 50% in 1 month: An AI agent deployed as 'CEO' of Kavak's Cuernavaca city operation increased profits by 50% within it…\"\n16:24\n. The single most actionable takeaway: matching engineering effort on evals to agent-building is what makes this scale safely.\n\nAngela Strange and Gabriel Vasquez join Alejandro Maza Ayala, Chief Product & AI Officer at Kavak, to unpack how the Latin American used-car marketplace rebuilt itself around AI agents, with 96% of customer interactions and 95% of transactions now handled by agents.\n\n-\nThe episode opens in media res, with Alejandro Maza Ayala delivering the thesis that defined Kavak's transformation: the goal was not AI-assisted humans, but superhuman agents that would outperform the best person ever hired by every meaningful measure. Angela Strange immediately surfaces the most ambitious architectural consequence — building one dedicated agent per customer rather than per task — and Alejandro confirms that between 100,000 and 200,000 such agents are instantiated every day. The episode's narrator then distils the founding question that separates Kavak from every other AI story: not 'how do we add AI?' but 'what would we build if we were starting from scratch?' This framing sets up the entire conversation as a case study in organizational reinvention rather than incremental tool adoption.\n\n-\nGabriel Vasquez welcomes Alejandro Maza Ayala as Kavak's head of AI and briefly surfaces his backstory: before Kavak, Alejandro co-founded Oppi Analytics in 2013, a machine learning company serving Fortune 500 clients in risk, logistics, and forecasting. He was building with ML before Transformers existed — a decade ahead of the mainstream. The arrival of Transformers and then the ChatGPT moment made it clear that an entirely new kind of company could be built. He joined Kavak and its CEO Carlos to build exactly that. Angela Strange then asks Alejandro to describe what Kavak actually does: a used-car marketplace that buys, refurbishes, sells, and finances cars — but which also had to build its own fintech, logistics layer, and vehicle-history infrastructure because none of it existed at scale in Latin America.\n\n-\nAngela asks Alejandro to ground the listener in what actually happens when a customer arrives. The answer is striking: a dedicated agent is spawned specifically for that customer, running in its own virtual machine. It remembers years of interaction history — a webpage visit, a call two years ago — formulates a long-term strategy, and relentlessly works to maximise lifetime value across all of Kavak's products over time. Alejandro then explains the three foundational decisions that made this possible. First, Kavak had to resist the instinct to simply hand ChatGPT to employees and instead rebuild its entire API and system layer so agents could actually act. Second, it made the bold bet that agents could become superhuman — not just helpful, but genuinely better than the best human ever hired. Third, it reoriented the company's success metrics from transactional (cars sold, parts ordered) to relational, assigning agents to all 10 million customers with a mandate to maximise lifetime value.\n\n[1] — Alejandro Maza Ayala \"Most companies build task-specific agents. Kavak builds one agent per customer, with its own virtual machine, long-term memory, and a singl…\" 04:00 -\nGabriel asks how Kavak evaluates whether agents are working at the scale of 96–98% of all interactions. Alejandro's answer reframes the question entirely: evals are not a safety tax, they are the accelerator. His analogy is elegant — you'll only floor the gas if you have good brakes. Most companies go slow on AI deployment because they lack rigorous evaluation, not because the models are weak. Kavak inverts this by spending roughly equal engineering time, tokens, and money on building evals as on building agents themselves\n\n[1] — Alejandro Maza Ayala \"Equal eng effort on evals vs agents: Kavak spends roughly equal engineering time, tokens, and money on building evals as it does on buildin…\" 09:05. The measurement focus is ruthlessly commercial: did the customer convert? Did they come back? Not vanity metrics like call duration or number of interactions. This discipline, Alejandro argues, is what separates genuine agentic deployment from theatre. -\nAngela raises one of the most persistent objections to agentic AI: customers won't buy expensive, high-stakes products from a bot. Kavak is empirically disproving this. Alejandro explains that Kavak never built customer service agents — it built sales agents. Buying a car in Latin America involves navigating up to 20,000 SKUs, financing structures, insurance options, and a trade-in valuation — a process that previously required speaking with 15 separate human experts. Kavak's agents consolidated all of that expertise into a single interaction. The early results were already compelling — agents converted at 1.5x the human rate — but the system has since improved to 2.1x\n\n[1] — Alejandro Maza Ayala \"AI sellers convert 2.1x human team: Kavak's AI sales agents now convert customers at more than 2.1 times the rate of the company's best hum…\" 12:34, while also tripling NPS. Angela then probes the even harder challenge: regulated financial services. Kavak approves car loans in under 3 minutes — versus the 2+ months typical in Mexico[2] — Alejandro Maza Ayala \"Car loan approved in under 3 minutes: Kavak approves car loans in under 3 minutes, versus the 2+ months typical in Mexico and other emergin…\" 14:18— by combining deep customer data with vertical integration that allows a repossessed car to be easily exchanged for a cheaper model if the customer falls behind on payments. -\nAlejandro turns to one of the most audacious experiments in the episode. The team asked a simple but alarming question: could an AI do the CEO's job? Rather than debate it theoretically, they ran the experiment. Cuernavaca — a mid-sized Mexican city — was chosen as the test site. An AI agent was given the CEO role, access to all company systems, and a mandate to double monthly profits. It fell short of that goal but still delivered 50% profit growth in its first month\n\n[1] — Alejandro Maza Ayala \"AI CEO raised profits 50% in 1 month: An AI agent deployed as 'CEO' of Kavak's Cuernavaca city operation increased profits by 50% within it…\" 16:24— by making Fields Medal-level analytical moves across inventory rotation, financing penetration, and customer satisfaction simultaneously. It sent daily task plans to every physical worker via message and collected voice-note updates on progress. For roles where physical dexterity remains irreplaceable — the company's approximately 800 mechanics — Kavak took a different approach. It built 'El Mike', an AI sidekick inspired by Ratatouille's Remy that guides mechanics through inspections step by step. Since launch, inspection quality rose, repair speed increased, costs fell, and warranty claims dropped by roughly 26%[2] — Alejandro Maza Ayala \"Warranties down ~26% post-AI rollout: Since deploying AI sidekick agents to Kavak's mechanics, warranty claims fell by approximately 26%, r…\" 17:58. -\nGabriel asks the question that is on every executive's mind: how do you actually change your organization to work this way? Alejandro's answer is the Jedi Academy — a 6-week internal training program he designed and continuously updates. The name is half-joke, but the program is serious. Executives, engineers, finance staff, and mechanics all go through the same curriculum. At the end of 6 weeks, every participant ships a production-ready AI agent\n\n[1] — Alejandro Maza Ayala \"Jedi Academy: 6-week AI training for all staff: Kavak's internal Jedi Academy trains every employee — from CEO to mechanics — to build and …\" 20:00. The content cannot be outsourced to Stanford or any external provider because the field moves faster than any accreditation cycle. The philosophical message delivered to all staff was equally clear: Kavak is going this direction. Employees can train and grow with it, or they can find a better fit elsewhere — but there is no option C where the company stays the same. Alejandro credits this clarity with actually strengthening Kavak's culture rather than creating anxiety. -\nAngela pushes on the structural reality: Kavak had thousands of employees. Agents now do most things. So what does the org actually look like? Alejandro describes it as radically flat — small, senior, cross-functional teams that blend engineering, AI, and operations. Some team members build agents, some work for agents by executing physically, and some serve customers in the real world. Middle management as a coordination layer has largely been replaced by agents as orchestrators\n\n[1] — Alejandro Maza Ayala \"Middle management is disappearing at Kavak. The new org is flat, senior, and hyper-empowered — each team blends engineering, AI, and operat…\" 22:55. Angela then asks about the human-in-the-loop design specifically. Most agentic systems, Alejandro observes, handle failure by escalating to a tier-2 queue and abandoning the case — which breaks the feedback loop and prevents agents from learning. Kavak's design is different: a stuck agent calls an API requesting human help, a human resolves it, and the agent that originally held the case maintains ownership. This closes the loop, generates the training data, and means the system actually improves. -\nAngela acknowledges that Kavak gets inbounded constantly by enterprise leaders who intuitively understand the imperative but can't move. Alejandro offers two unambiguous prescriptions. First: AI transformation cannot be bottom-up. Hackathons and sponsored use-cases produce noise, not compounding value. A leader needs to articulate what the company looks like in 3 to 5 years and give the organization a single direction to march\n\n[1] — Alejandro Maza Ayala \"Most companies measure AI adoption by token spend and call it progress. Kavak built a 3-tier framework: Tier 3 tokens go to autonomous agen…\" 26:30. Second: measure the right things. Most companies have mistaken token adoption for progress. Alejandro's three-tier framework cuts through this. Tier 3 tokens — those powering autonomous agents with measurable per-token ROI — are the only ones that compound. Tier 2, which funds developer tooling with indirect measurability, is useful but not transformational. Tier 1, where employees use consumer AI tools with no accountability, is largely invisible spend. The goal is to move as much spend as possible into Tier 3, where every token can be traced to a business outcome. -\nAngela asks Alejandro to walk through the architectural decision to build one agent per customer rather than per task. He explains how Kavak evolved from task-based workflows to complex multi-agent graphs — tens of thousands of these agents were running at scale in December, profitably. Then Claude Opus 4.5 was released, and Alejandro had a sobering realization: the graph architecture that had taken two years to build was now a constraint, not an enabler. The new models were intelligent enough to pursue complex goals without the scaffolding\n\n[1] — Alejandro Maza Ayala \"When Claude Opus 4.5 arrived, Alejandro realized Kavak's entire multi-agent graph architecture — two years of work, profitable, scaling — w…\" 29:15. So Kavak did something most companies would never consider: it destroyed a working, profitable system and rebuilt from scratch around a simpler, more powerful paradigm — one long-running agent per customer, with its own virtual machine, memory, CLI access to every company tool, and a single long-term goal. Gabriel names it precisely: the self-improving organization. Alejandro expands the concept — for 4,000 years, economic value has been created by organizations, not individuals. The compounding loop that matters is the one that makes the organization smarter every day, not just the model. -\nGabriel surfaces a claim Alejandro made earlier: the biggest AI opportunity may lie with entirely new companies, not incumbents. Alejandro grounds this in Schumpeter's theory of creative destruction — innovation historically destroys old companies through new entrants rather than incumbent adaptation. The reason is structural: a CEO of a large public company faces enormous friction in betting the organization on a complete rebuild. New founders face no such constraint. To make the argument visceral, Alejandro tells a story his team has heard many times: Edison commercialized electricity in the 1880s, meaning Ford's production line could have been built 40 years earlier. Factories that simply replaced coal engines with electric ones captured only 6% of the available efficiency. Those that tore down their multi-story, shaft-and-belt factories and rebuilt flat, distributed facilities around small dynamos captured a 3x improvement in productivity that fueled 20th-century American growth\n\n[1] — Alejandro Maza Ayala \"Factories that replaced coal engines with electric ones in the early 1900s got only 6% efficiency gains. Those that redesigned their entire…\" 32:40. The same pattern played out with the computer, and it is playing out again now. Alejandro's closing advice to founders is simple and urgent: build deep, build for the future capability curve, and recognize that for the first time in history, the most powerful tools in the world are available to anyone for $20 a month. -\nThe episode wraps with a brief outro encouraging listeners to like, subscribe, and follow a16z across platforms including YouTube, Apple Podcasts, Spotify, X, and Substack. A standard legal disclaimer follows, noting that the content is for informational purposes only and does not constitute investment advice, and that a16z and its affiliates may hold positions in companies discussed.\n\n- Agentic architecture\n- A software design where autonomous AI agents — rather than rule-based workflows — perceive context, make decisions, and take actions to achieve long-term goals.\n- Evals (evaluations)\n- Automated test suites used to measure AI agent performance against real business outcomes; in this episode, Kavak treats them as equally important as building the agents themselves.\n- Virtual machine (in agent context)\n- An isolated computing environment spawned for each AI agent instance, giving it dedicated resources, memory, and tool access.\n- Multi-agent system\n- A network of specialized AI agents that each handle a subtask and coordinate to complete a complex goal.\n- Long-running agent\n- An AI agent that persists over hours, days, or weeks — setting goals, pausing, and resuming — rather than completing a single short task.\n- Lifetime value (LTV)\n- The total revenue a business expects to earn from a customer over the entire relationship; Kavak uses this as the primary goal metric for each customer's assigned agent.\n- Recursive self-improvement (RSI)\n- A process where an AI system iteratively improves its own capabilities; Alejandro extends the concept to organizations that continuously improve through AI feedback loops.\n- Fine-tuning\n- The process of further training a pre-existing AI model on domain-specific data to improve its performance for a particular task.\n- CLI (Command Line Interface)\n- A text-based interface for interacting with software; Kavak's agents use CLI access to invoke every API and tool in the company.\n- NPS (Net Promoter Score)\n- A customer loyalty metric measuring how likely customers are to recommend a company; Kavak tripled its NPS after deploying AI sales agents.\n- SKU (Stock Keeping Unit)\n- A unique identifier for a product variant; used here to describe the ~20,000 distinct used-car listings Kavak's agents must navigate.\n- Thin-file customer\n- A borrower with little or no credit history, making standard risk assessment difficult; a key challenge in emerging-market fintech lending.\n- Creative destruction\n- An economic concept coined by Joseph Schumpeter describing how innovation displaces existing industries and companies, creating long-term gains despite short-term disruption.\n- Innovator's dilemma\n- Clayton Christensen's theory that successful incumbents are often disrupted because they optimize for existing customers rather than adopting transformative new technologies.\n- PII (Personally Identifiable Information)\n- Data that can identify a specific individual; mentioned as a key risk area when designing evals for AI systems handling financial services.\n- Fintech\n- Short for financial technology; companies using software to provide financial services; Kavak built its own fintech layer to offer car loans and personal loans.\n- Vertical integration\n- A business strategy where a company controls multiple stages of its supply chain; Kavak built logistics, financing, and inspection capabilities in-house because the infrastructure didn't exist in Latin America.\n- Schumpeter, Joseph\n- 20th-century Austrian economist famous for the theory of creative destruction, cited by Alejandro to argue that AI will create new companies that displace incumbents.\n- Transformers\n- A neural network architecture introduced in 2017 that underpins modern large language models including GPT and Claude; Alejandro contrasts this era with earlier ML approaches.\n- Opus 4.5\n- Refers to Claude Opus 4.5, Anthropic's advanced large language model; Alejandro credits its release as the trigger for Kavak rebuilding its entire agent architecture.\n\nChapter 1 · 00:00\n\n## Cold Open: Superhuman Agents and the Radical Question\n\nThe episode opens in media res, with Alejandro Maza Ayala delivering the thesis that defined Kavak's transformation: the goal was not AI-assisted humans, but superhuman agents that would outperform the best person ever hired by every meaningful measure. Angela Strange immediately surfaces the most ambitious architectural consequence — building one dedicated agent per customer rather than per task — and Alejandro confirms that between 100,000 and 200,000 such agents are instantiated every day. The episode's narrator then distils the founding question that separates Kavak from every other AI story: not 'how do we add AI?' but 'what would we build if we were starting from scratch?' This framing sets up the entire conversation as a case study in organizational reinvention rather than incremental tool adoption.\n\nChapter 3 · 04:00\n\n## The Architecture: One Agent Per Customer\n\nAngela asks Alejandro to ground the listener in what actually happens when a customer arrives. The answer is striking: a dedicated agent is spawned specifically for that customer, running in its own virtual machine. It remembers years of interaction history — a webpage visit, a call two years ago — formulates a long-term strategy, and relentlessly works to maximise lifetime value across all of Kavak's products over time. Alejandro then explains the three foundational decisions that made this possible. First, Kavak had to resist the instinct to simply hand ChatGPT to employees and instead rebuild its entire API and system layer so agents could actually act. Second, it made the bold bet that agents could become superhuman — not just helpful, but genuinely better than the best human ever hired. Third, it reoriented the company's success metrics from transactional (cars sold, parts ordered) to relational, assigning agents to all 10 million customers with a mandate to maximise lifetime value.\n[1]\n— Alejandro Maza Ayala\n\"Most companies build task-specific agents. Kavak builds one agent per customer, with its own virtual machine, long-term memory, and a singl…\"\n04:00\n\n[Agent Per Customer, Not Per Task](/bit/podbit/11798/)\n\nMost companies build task-specific agents. Kavak builds one agent per customer, with its own virtual machine, long-term memory, and a single goal: maximize that customer's lifetime value over years. The agent wakes up, works, sets an alarm, and comes back — it never forgets.\n\n[Why Kavak Blew Up Its Org to Rebuild Around Agents](/bit/podbit/11788/)\n\nSimply giving employees ChatGPT changes nothing. Kavak's first move was to rebuild its APIs, systems, and feedback loops so agents could actually do the work. The second bet: agents could be superhuman — outperforming the best human hire on every metric that mattered.\n\n[10 million customers in Kavak's database](/bit/snapshot/12894/)\n\nKavak has 10 million customers in its database, each with AI agents assigned to maximize their lifetime value.\n\nChapter 4 · 08:10\n\n## Evals: The Brakes That Let You Go Fast\n\nGabriel asks how Kavak evaluates whether agents are working at the scale of 96–98% of all interactions. Alejandro's answer reframes the question entirely: evals are not a safety tax, they are the accelerator. His analogy is elegant — you'll only floor the gas if you have good brakes. Most companies go slow on AI deployment because they lack rigorous evaluation, not because the models are weak. Kavak inverts this by spending roughly equal engineering time, tokens, and money on building evals as on building agents themselves\n[1]\n— Alejandro Maza Ayala\n\"Equal eng effort on evals vs agents: Kavak spends roughly equal engineering time, tokens, and money on building evals as it does on buildin…\"\n09:05\n. The measurement focus is ruthlessly commercial: did the customer convert? Did they come back? Not vanity metrics like call duration or number of interactions. This discipline, Alejandro argues, is what separates genuine agentic deployment from theatre.\n\n[96% of Interactions, 95% of Transactions: Kavak's Agent Scale](/bit/podbit/11787/)\n\n96% of all customer interactions and 95% of all transactions at Kavak are now handled by AI agents. Between 100,000 and 200,000 agent instances wake up every day, each with its own virtual machine, working anywhere from 3 minutes to 3 days before setting an alarm for their next task.\n\n[96% interactions handled by agents](/bit/snapshot/12883/)\n\nKavak's AI agents handle 96% of all customer interactions with no human involvement.\n\n[95% of transactions by agents](/bit/snapshot/12884/)\n\n95% of all Kavak transactions are completed end-to-end by AI agents, not humans.\n\n[100K–200K agents instantiated daily](/bit/snapshot/12885/)\n\nBetween 100,000 and 200,000 unique agent instances are spun up at Kavak every single day, each with its own virtual machine.\n\n[Evals Are the Brakes That Let You Go Fast](/bit/podbit/11789/)\n\nMost companies treat evals as an afterthought. Kavak treats them as the foundation. Spending equal engineering effort on evals and agents is what allows Kavak to deploy at massive scale without blowing up — because better brakes mean you can press the gas harder.\n\n[Equal eng effort on evals vs agents](/bit/snapshot/12886/)\n\nKavak spends roughly equal engineering time, tokens, and money on building evals as it does on building agents themselves.\n\nChapter 5 · 11:00\n\n## AI Sellers Outperform Humans 2.1x — and Tripled NPS\n\nAngela raises one of the most persistent objections to agentic AI: customers won't buy expensive, high-stakes products from a bot. Kavak is empirically disproving this. Alejandro explains that Kavak never built customer service agents — it built sales agents. Buying a car in Latin America involves navigating up to 20,000 SKUs, financing structures, insurance options, and a trade-in valuation — a process that previously required speaking with 15 separate human experts. Kavak's agents consolidated all of that expertise into a single interaction. The early results were already compelling — agents converted at 1.5x the human rate — but the system has since improved to 2.1x\n[1]\n— Alejandro Maza Ayala\n\"AI sellers convert 2.1x human team: Kavak's AI sales agents now convert customers at more than 2.1 times the rate of the company's best hum…\"\n12:34\n, while also tripling NPS. Angela then probes the even harder challenge: regulated financial services. Kavak approves car loans in under 3 minutes — versus the 2+ months typical in Mexico\n[2]\n— Alejandro Maza Ayala\n\"Car loan approved in under 3 minutes: Kavak approves car loans in under 3 minutes, versus the 2+ months typical in Mexico and other emergin…\"\n14:18\n— by combining deep customer data with vertical integration that allows a repossessed car to be easily exchanged for a cheaper model if the customer falls behind on payments.\n\n[AI Sellers Outperform Humans by 2.1x](/bit/podbit/11790/)\n\nKavak's AI sales agents now convert customers at 2.1x the rate of its human team, and also tripled NPS. The agents handle everything from car recommendations to financing to trade-in quotes — a 15-skill job that no single human could match.\n\n[NPS tripled with AI agents](/bit/snapshot/12889/)\n\nKavak's Net Promoter Score and customer satisfaction score tripled after deploying AI agents as the primary customer-facing interface.\n\n[AI sellers convert 2.1x human team](/bit/snapshot/12887/)\n\nKavak's AI sales agents now convert customers at more than 2.1 times the rate of the company's best human salespeople.\n\n[Car loan approved in under 3 minutes](/bit/snapshot/12890/)\n\nKavak approves car loans in under 3 minutes, versus the 2+ months typical in Mexico and other emerging markets.\n\nChapter 6 · 15:55\n\n## The AI CEO Experiment: 50% Profit Growth in One Month\n\nAlejandro turns to one of the most audacious experiments in the episode. The team asked a simple but alarming question: could an AI do the CEO's job? Rather than debate it theoretically, they ran the experiment. Cuernavaca — a mid-sized Mexican city — was chosen as the test site. An AI agent was given the CEO role, access to all company systems, and a mandate to double monthly profits. It fell short of that goal but still delivered 50% profit growth in its first month\n[1]\n— Alejandro Maza Ayala\n\"AI CEO raised profits 50% in 1 month: An AI agent deployed as 'CEO' of Kavak's Cuernavaca city operation increased profits by 50% within it…\"\n16:24\n— by making Fields Medal-level analytical moves across inventory rotation, financing penetration, and customer satisfaction simultaneously. It sent daily task plans to every physical worker via message and collected voice-note updates on progress. For roles where physical dexterity remains irreplaceable — the company's approximately 800 mechanics — Kavak took a different approach. It built 'El Mike', an AI sidekick inspired by Ratatouille's Remy that guides mechanics through inspections step by step. Since launch, inspection quality rose, repair speed increased, costs fell, and warranty claims dropped by roughly 26%\n[2]\n— Alejandro Maza Ayala\n\"Warranties down ~26% post-AI rollout: Since deploying AI sidekick agents to Kavak's mechanics, warranty claims fell by approximately 26%, r…\"\n17:58\n.\n\n[The AI CEO That Raised Profits 50% in One Month](/bit/podbit/11791/)\n\nKavak carved out a city in Mexico, gave an AI agent the CEO role, and watched it raise profits by 50% in its first month. The agent sent daily task plans to every physical worker, monitored KPIs in real time, and micromanaged inventory, financing, and customer satisfaction simultaneously.\n\n[AI CEO raised profits 50% in 1 month](/bit/snapshot/12888/)\n\nAn AI agent deployed as 'CEO' of Kavak's Cuernavaca city operation increased profits by 50% within its first month.\n\n[Mechanics Get a Ratatouille-Style AI Sidekick](/bit/podbit/11799/)\n\nPhysical dexterity is still irreplaceable. So for Kavak's 800 mechanics, the company built 'El Mike' — an AI sidekick inspired by Ratatouille's Remy, guiding mechanics through inspections and repairs. Result: faster inspections, higher quality cars, and a 26% drop in warranty claims.\n\n[Warranties down ~26% post-AI rollout](/bit/snapshot/12891/)\n\nSince deploying AI sidekick agents to Kavak's mechanics, warranty claims fell by approximately 26%, reflecting higher-quality car inspections.\n\nChapter 7 · 19:45\n\n## The Jedi Academy: Training Everyone to Build AI\n\nGabriel asks the question that is on every executive's mind: how do you actually change your organization to work this way? Alejandro's answer is the Jedi Academy — a 6-week internal training program he designed and continuously updates. The name is half-joke, but the program is serious. Executives, engineers, finance staff, and mechanics all go through the same curriculum. At the end of 6 weeks, every participant ships a production-ready AI agent\n[1]\n— Alejandro Maza Ayala\n\"Jedi Academy: 6-week AI training for all staff: Kavak's internal Jedi Academy trains every employee — from CEO to mechanics — to build and …\"\n20:00\n. The content cannot be outsourced to Stanford or any external provider because the field moves faster than any accreditation cycle. The philosophical message delivered to all staff was equally clear: Kavak is going this direction. Employees can train and grow with it, or they can find a better fit elsewhere — but there is no option C where the company stays the same. Alejandro credits this clarity with actually strengthening Kavak's culture rather than creating anxiety.\n\n[The Jedi Academy: Training Every Employee to Build AI Agents](/bit/podbit/11792/)\n\nKavak launched the Jedi Academy, an internal program where every employee — from CEO to mechanics — completes a 6-week training and ships a production-ready AI agent. The message to staff was blunt: train and adapt, or this may not be the right place for you.\n\n[Jedi Academy: 6-week AI training for all staff](/bit/snapshot/12892/)\n\nKavak's internal Jedi Academy trains every employee — from CEO to mechanics — to build and deploy AI agents to production in 6 weeks.\n\nChapter 8 · 22:55\n\n## New Org Structure: Flat, Senior, and Sometimes Humans Work for Agents\n\nAngela pushes on the structural reality: Kavak had thousands of employees. Agents now do most things. So what does the org actually look like? Alejandro describes it as radically flat — small, senior, cross-functional teams that blend engineering, AI, and operations. Some team members build agents, some work for agents by executing physically, and some serve customers in the real world. Middle management as a coordination layer has largely been replaced by agents as orchestrators\n[1]\n— Alejandro Maza Ayala\n\"Middle management is disappearing at Kavak. The new org is flat, senior, and hyper-empowered — each team blends engineering, AI, and operat…\"\n22:55\n. Angela then asks about the human-in-the-loop design specifically. Most agentic systems, Alejandro observes, handle failure by escalating to a tier-2 queue and abandoning the case — which breaks the feedback loop and prevents agents from learning. Kavak's design is different: a stuck agent calls an API requesting human help, a human resolves it, and the agent that originally held the case maintains ownership. This closes the loop, generates the training data, and means the system actually improves.\n\n[Flat, Senior, Empowered: The New Org Structure](/bit/podbit/11793/)\n\nMiddle management is disappearing at Kavak. The new org is flat, senior, and hyper-empowered — each team blends engineering, AI, and operations. Some humans build agents, some work for agents, and some serve the physical world. It's a structure that didn't exist 2 years ago.\n\nChapter 9 · 25:35\n\n## Advice for Leaders: Top-Down Vision and the Three Tiers of Tokens\n\nAngela acknowledges that Kavak gets inbounded constantly by enterprise leaders who intuitively understand the imperative but can't move. Alejandro offers two unambiguous prescriptions. First: AI transformation cannot be bottom-up. Hackathons and sponsored use-cases produce noise, not compounding value. A leader needs to articulate what the company looks like in 3 to 5 years and give the organization a single direction to march\n[1]\n— Alejandro Maza Ayala\n\"Most companies measure AI adoption by token spend and call it progress. Kavak built a 3-tier framework: Tier 3 tokens go to autonomous agen…\"\n26:30\n. Second: measure the right things. Most companies have mistaken token adoption for progress. Alejandro's three-tier framework cuts through this. Tier 3 tokens — those powering autonomous agents with measurable per-token ROI — are the only ones that compound. Tier 2, which funds developer tooling with indirect measurability, is useful but not transformational. Tier 1, where employees use consumer AI tools with no accountability, is largely invisible spend. The goal is to move as much spend as possible into Tier 3, where every token can be traced to a business outcome.\n\n[Token Tiers: How to Measure Real AI ROI](/bit/podbit/11794/)\n\nMost companies measure AI adoption by token spend and call it progress. Kavak built a 3-tier framework: Tier 3 tokens go to autonomous agents with measurable ROI per token; Tier 2 to dev tooling with indirect measurability; Tier 1 is employees using ChatGPT with no accountability. Only Tier 3 compounds.\n\nChapter 10 · 29:15\n\n## Destroying Two Years of Work: The New Architecture\n\nAngela asks Alejandro to walk through the architectural decision to build one agent per customer rather than per task. He explains how Kavak evolved from task-based workflows to complex multi-agent graphs — tens of thousands of these agents were running at scale in December, profitably. Then Claude Opus 4.5 was released, and Alejandro had a sobering realization: the graph architecture that had taken two years to build was now a constraint, not an enabler. The new models were intelligent enough to pursue complex goals without the scaffolding\n[1]\n— Alejandro Maza Ayala\n\"When Claude Opus 4.5 arrived, Alejandro realized Kavak's entire multi-agent graph architecture — two years of work, profitable, scaling — w…\"\n29:15\n. So Kavak did something most companies would never consider: it destroyed a working, profitable system and rebuilt from scratch around a simpler, more powerful paradigm — one long-running agent per customer, with its own virtual machine, memory, CLI access to every company tool, and a single long-term goal. Gabriel names it precisely: the self-improving organization. Alejandro expands the concept — for 4,000 years, economic value has been created by organizations, not individuals. The compounding loop that matters is the one that makes the organization smarter every day, not just the model.\n\n[Destroying Two Years of Working Tech to Start Over](/bit/podbit/11795/)\n\nWhen Claude Opus 4.5 arrived, Alejandro realized Kavak's entire multi-agent graph architecture — two years of work, profitable, scaling — was the wrong paradigm for the new level of intelligence. They tore it down and rebuilt around long-running single agents with virtual machines, memory, and CLI access to every company API.\n\n[The Self-Improving Organization](/bit/podbit/11796/)\n\nEveryone is obsessing over model-level recursive self-improvement. The real prize is organizational RSI. For 4,000 years, economic value has been created by organizations, not individuals — so the compounding loop that matters is the one that makes your company smarter every day, not just your AI model.\n\n[The Electricity Parallel: 6% vs 3x](/bit/podbit/11797/)\n\nFactories that replaced coal engines with electric ones in the early 1900s got only 6% efficiency gains. Those that redesigned their entire factory around electricity got a 3x productivity improvement. Swapping AI tools into your existing structure will give you 6%. Rebuilding around AI will give you 10x.\n\nChapter 11 · 32:45\n\n## Creative Destruction, the Electricity Parallel, and the Case for Founders\n\nGabriel surfaces a claim Alejandro made earlier: the biggest AI opportunity may lie with entirely new companies, not incumbents. Alejandro grounds this in Schumpeter's theory of creative destruction — innovation historically destroys old companies through new entrants rather than incumbent adaptation. The reason is structural: a CEO of a large public company faces enormous friction in betting the organization on a complete rebuild. New founders face no such constraint. To make the argument visceral, Alejandro tells a story his team has heard many times: Edison commercialized electricity in the 1880s, meaning Ford's production line could have been built 40 years earlier. Factories that simply replaced coal engines with electric ones captured only 6% of the available efficiency. Those that tore down their multi-story, shaft-and-belt factories and rebuilt flat, distributed facilities around small dynamos captured a 3x improvement in productivity that fueled 20th-century American growth\n[1]\n— Alejandro Maza Ayala\n\"Factories that replaced coal engines with electric ones in the early 1900s got only 6% efficiency gains. Those that redesigned their entire…\"\n32:40\n. The same pattern played out with the computer, and it is playing out again now. Alejandro's closing advice to founders is simple and urgent: build deep, build for the future capability curve, and recognize that for the first time in history, the most powerful tools in the world are available to anyone for $20 a month.\n\n[Electricity parallel: 6% vs 3x productivity](/bit/snapshot/12893/)\n\nSuperficial adoption of electricity gave factories only 6% efficiency gains; fully redesigning factories around it delivered a 3x productivity improvement that powered 20th-century US growth.\n\n[Advice for Founders: Build for 2035, Not 2025](/bit/podbit/11800/)\n\nThe right way to build now is to project a linear trend for AI improvement and design your company for that future state — then build backwards. This is the first time in history that the most powerful tools available are accessible to any founder for $20 a month.\n\nNo indexed bits in this chapter.\n\n## Show stoppers\n\n## Snapshots ()\n\n## Key Quotes ()\n\nSign in to keep viewing\n\nCreate a free account to keep exploring this episode's insights, snapshots, quotes and claims.", "url": "https://wpnews.pro/news/the-self-improving-company-kavak-s-ai-playbook", "canonical_source": "https://vuci.ai/the-a16z-show/episode/how-kavak-rebuilt-itself-around-ai-agents-alejandro-maza-ayala/", "published_at": "2026-08-10 10:00:00+00:00", "updated_at": "2026-08-15 20:11:39.420549+00:00", "lang": "en", "topics": ["artificial-intelligence", "ai-agents", "ai-products"], "entities": ["Kavak", "Alejandro Maza Ayala", "Angela Strange", "Gabriel Vasquez", "Oppi Analytics", "Carlos"], "alternates": {"html": "https://wpnews.pro/news/the-self-improving-company-kavak-s-ai-playbook", "markdown": "https://wpnews.pro/news/the-self-improving-company-kavak-s-ai-playbook.md", "text": "https://wpnews.pro/news/the-self-improving-company-kavak-s-ai-playbook.txt", "jsonld": "https://wpnews.pro/news/the-self-improving-company-kavak-s-ai-playbook.jsonld"}}