Kavak built copilot tools for 3,500 employees. Nobody used them. So they ripped them out and replaced them with autonomous agents — funnel by funnel, over three years. Today, 90-95% of all customer interactions have an AI agent in the middle. The copilots were the dead end. The agents were the exit.

Key Takeaway: In a new a16z Deep Dives episode, Angela Strange interviews Kavak CEO Carlos Garcia Ottati about the largest AI agent deployment in Latin American e-commerce. The lesson isn’t about cars — it’s about why giving humans better tools doesn’t work when the problem is the humans being in the loop at all. Kavak’s journey from copilots to agents maps exactly to the shift from marketing copilots (Jasper, HubSpot AI) to marketing agents that own the workflow end-to-end.

The Copilot Trap

Carlos describes the moment of truth with unusual clarity: “We made the typical mistake that everybody’s making — we built these co-pilot tools to give them to our teams… We realized very quickly they didn’t adopt them.”

This wasn’t a technology problem. Kavak had spent since 2017 building the data ontology — connecting every system, every customer interaction, every edge case across their car marketplace. By the time ChatGPT arrived in late 2022, they had the infrastructure. They built the copilots. They gave them to thousands of employees.

And then: nothing.

“Even if you had the right information, even if you had the correct highway for people to find out what they needed, they were not going to do it.”

This is the sentence that matters most in the entire interview. It’s not that the copilots were bad. It’s not that the information was wrong. It’s that humans don’t change their behavior just because you give them better tools. They keep doing what they were doing before — the same way most marketers who get access to Claude or Jasper use it to write the same blog posts faster, not to rethink their entire marketing operation.

We’ve written about this dynamic before: Anthropic’s own marketing team saves 100+ hours a month with Claude, but every workflow they automated is content creation — not distribution, not optimization, not strategy. The copilot ceiling isn’t a technology limit. It’s a human behavior limit.

What Kavak Did Instead

Carlos didn’t try to fix the copilots. He replaced them.

“So we went funnel by funnel putting agents in front of critical aspects of the business.”

The key word is funnel by funnel. Not a big bang transformation. Not “we’re replacing everyone with AI.” They picked one process — not customer service (the easy win) but underwriting and post-sale edge cases (the hard problems) — and put an agent in front of it.

The decision framework was brutally simple: “Can this be solved with humans that have blood, or agents that have electricity going through their veins?”

The Orchestrator Pattern

What Kavak built isn’t a single AI doing everything. It’s agents working “symbiotically with humans” — with an orchestrator deciding who handles what based on complexity:

Simple interactions: Agent handles autonomously
Complex edge cases: Agent escalates to human with full context
Novel situations: Human handles, agent learns

This is the pattern that matters. Not copilots that help humans do their job. Agents that do the job, with humans handling the exceptions the agents can’t yet solve.

Angela Strange pushes on this: Kavak had “weeks or maybe even a couple of months where various functions were underperforming the humans.”

Carlos corrects her: “No, we had a year.”

The Year of Pain

This is where the Kavak story gets uncomfortable for anyone selling “AI transformation in 30 days.”

Kavak was growing 100% year-over-year in 2022. In 2023, they went flat. Not because the market crashed (though it had earlier) — because they were rebuilding everything underneath:

Sales dropped because agents weren’t yet at human parity
Customer experience deteriorated because the agents were still learning
Every output KPI went down before it went back up

Carlos is blunt about the approach: “We didn’t have a Plan B. The Plan A was we need to make this work.”

For each funnel, the pattern was identical:

Deploy agent
Watch metrics drop below human baseline
Iterate until agent reaches parity
Agent surpasses human performance (1.5x their best human in the first funnel)
Don’t optimize further — move to the next funnel

The philosophy behind step 5 is revealing: “We don’t build for ChatGPT 4. We build for ChatGPT 7.” They built the architecture knowing the models would improve faster than they could optimize, so they invested in infrastructure over point performance.

After three years of this — funnel by funnel, each one following the same dip-then-rise pattern — Kavak now runs 90-95% of all user interactions with an AI agent in the middle.

Why This Matters for Marketing

Kavak sells cars. We build AI for marketing. But the pattern is identical, and the lesson translates directly.

The Marketing Copilot is the Same Dead End

Every marketing AI tool today is a copilot:

Tool	What It Does	What It Doesn’t Do
Jasper	Writes copy faster	Doesn’t decide what to write, when to publish, or whether it worked
HubSpot Breeze	Generates email drafts	Doesn’t choose the segment, time the send, or learn from results
Claude/ChatGPT	Answers marketing questions	Forgets everything between sessions, has no access to your data
Vibe marketing setups	Executes tasks in sessions	Requires a developer-founder to wire and maintain everything

These tools make marketers faster at tasks they already know how to do. But as Carlos discovered with 3,500 employees: giving people better tools doesn’t change their behavior. The marketer who writes one mediocre blog post a week now writes three mediocre blog posts a week. The startup founder who wasn’t doing marketing still isn’t doing marketing — they’re just procrastinating with a shinier tool.

We covered this in our vibe marketing analysis: the gap between “I can build this” and “this runs while I sleep” is where most AI marketing setups die.

The Agent Model Works the Same Way

Kavak’s architecture maps to what marketing needs:

Kavak	Marketing Equivalent
Ontology (all systems connected since 2017)	Brand context: analytics, CRM, social, email, content — all connected
Orchestrator (routes to agent or human by complexity)	Trust model: auto-publish low-risk content, escalate brand-sensitive decisions
Funnel-by-funnel deployment	Channel-by-channel: start with social scheduling, then email, then content strategy
Agent learning from edge cases	Learning from campaign results: what worked, what didn’t, what to try next
”Build for ChatGPT 7”	Architecture that improves as models improve, not locked to today’s capabilities

The orchestrator is the critical piece. Carlos describes it as “guiding who was going to solve which” — the agent or the human. In marketing, this is the trust model: which decisions can the AI make autonomously, which need human approval, and which should always be human?

We call this the Brand Parent model — an AI that carries persistent concern for your brand and makes decisions on its own within boundaries you set, escalating when it encounters something outside its confidence range.

No Plan B Is the Only Plan

The most counterintuitive lesson from Kavak is the commitment level. They didn’t run copilots and agents in parallel to “see which one wins.” They killed the copilots and went all-in on agents, accepting a year of degraded performance.

“We needed to burn all the ships and just do this for the whole organization.”

For a startup founder choosing between “use AI tools to help me do marketing” and “let AI do my marketing,” this is the decision. The copilot approach feels safer. It’s also the dead end. You’ll spend forever maintaining AI-assisted workflows that still require you to be the marketer.

The agent approach is scarier. The metrics will look worse before they look better. But it’s the only path to marketing that actually runs while you’re building your product.

The Three-Year Moat

There’s one more lesson in the Kavak story that most coverage will miss.

Carlos started building the data ontology in 2017 — five years before ChatGPT existed. By the time the AI tools arrived, Kavak had every system connected, every edge case documented, every interaction logged. They were “really ready from an ontology perspective.”

The companies starting now — the ones saying “we’ll add AI to our workflows” — are five years behind. Not in AI capabilities. In data architecture. In connected systems. In the institutional knowledge that makes agents work.

This is why Kavak’s agent deployment took three years even with the ontology already built. And it’s why the moat for AI marketing isn’t the model — it’s the persistent memory of every campaign, every customer interaction, every A/B test, every market signal that accumulates over time. The same compounding advantage that protects traditional software moats, applied to an AI agent that gets smarter every day.

The copilot has no memory. The agent compounds.

What Happens Next

Carlos has a Netflix analogy for the transition: Blockbuster was the incumbent. DVDs by mail (early digital) was the intermediate step. Streaming was the real shift.

In marketing:

Blockbuster = traditional marketing teams and agencies
DVDs by mail = AI copilots (Jasper, HubSpot Breeze, vibe marketing)
Streaming = autonomous marketing agents with persistent memory and orchestration

We’re in the DVDs-by-mail phase. Everyone is excited about getting content delivered faster. Nobody has built the streaming infrastructure yet — the persistent agents that own the workflow, learn from every interaction, and operate within a trust model that lets them act autonomously.

Kavak proved it works for car sales across Latin America. The same architecture — ontology, orchestrator, funnel-by-funnel deployment, no Plan B — applies to marketing for startups that can’t afford a team but can’t afford to stop growing.

The copilot was the dead end. The agent is the road.

References

a16z. AI Copilots Are a Dead End. Here’s What Actually Works — Kavak CEO. a16z Deep Dives, March 2026.
Kavak. Official website. Latin America’s largest online used car marketplace.
Anthropic. How Anthropic’s Marketing Team Uses Claude. Case study on copilot-based marketing workflows.

This is part of our series on the Brand Parent thesis — why marketing needs persistent AI agents, not just faster tools. Previously: The Copilot Ceiling, Vibe Marketing Won’t Scale, Software Moats in Marketing.