Kavak
Confidence 0.70 · 1 source · last confirmed 2026-09-16
Latin American used-car marketplace, founded in Mexico. Its stated business is buying, refurbishing, selling and financing used cars — but because the supporting infrastructure did not exist regionally, it built the adjacent businesses itself: “we also had to build a fintech and a logistics company and a Carfax, and basically all the infrastructure for this to work didn’t exist in LatAm. So we had to build everything vertically.”
Why the wiki holds it: Kavak is the corpus’s first end-to-end account of a company that rebuilt itself around agents and reports operating figures. Almost everything else here on enterprise adoption is survey data, vendor testimony, or framework. That makes this page valuable and also makes its single-source status worth stating plainly — see Open questions.
Appears in this wiki via
- 2026-08-10-maza-a16z-kavak-rebuilding-a-company-around-ai — Alejandro Maza Ayala, Chief Product & AI Officer, on the a16z Podcast (10 Aug 2026), interviewed by Angela Strange and Gabriel Vasquez. The full account of the transformation: three bets (redesign the company, build superhuman agents, change the metrics), the agent-per-customer architecture, the eval-resourcing rule, the AI-CEO experiment in Cuernavaca, the Jedi Academy, the token tiering, and the Schumpeter/Ford argument for why incumbents under-realise.
What Kavak reports
All figures are self-reported by one executive on a venture-capital podcast, unaudited, and without stated definitions or baselines. Recorded here as claims:
| Measure | Reported |
|---|---|
| Customer interactions handled by agents | 96% |
| Transactions handled by agents | 95% |
| Agents instantiated per day | 100,000–200,000, each with its own VM |
| AI seller vs human team conversion | +50% initially, 2.1× at time of recording |
| NPS / customer satisfaction | tripled |
| Car-loan approval | under 3 minutes |
| AI “CEO” of Cuernavaca, first month | 1.5× profit against a 2× target |
| Warranty claims after mechanic sidekick | down ~26% |
| Customers in database | 10 million |
| Mechanics in Mexico | ~800 |
The architecture, in brief
One long-running agent per customer, not per task — a virtual machine with memory, evals, a CLI, access to every internal tool and API, and a long-horizon goal of maximising that customer’s lifetime value. Agents sleep and wake on their own schedule, running from minutes to days.
This replaced a working multi-agent system of “tens of thousands” of function-decomposed agents, discarded after the Opus 4.5 release on the view that decomposition now constrains rather than enables a sufficiently capable model. Kavak spends “about the same amount of engineer time, tokens, and money” on evals as on agents.
Humans remain where the physical world does — ~800 mechanics, who work with a sidekick agent — and the escalation path is inverted: agents call an API for help and a human answers it, so “if you map this out in an org chart, it’s really human teams that have an agent.”
Why it matters to the wiki’s open questions
Kavak sits directly on the corpus’s largest unresolved question. Covello states that “enterprises collectively are not making or saving money on their AI implementations” and names the missing piece as a data-management plus model-orchestration layer. Kavak reports the opposite outcome and describes building something close to that layer. Whether Kavak is the exception that proves the aggregate, the leading edge of it, or a case whose numbers would not survive audit is not settled by anything the wiki currently holds. See micro-productivity-trap and enterprise-ai-adoption.
Open questions
- Single source, and an interested one. Everything on this page comes from one podcast episode produced by a venture firm. No filing, report, customer account, competitor view, or independent measurement corroborates any figure. A second source is the highest-value addition this page could get.
- Financial position undisclosed. The only profitability claim in the episode is that the prior architecture “brought us to profitability” — made about the system Kavak then discarded. Current economics are not discussed, and Kavak’s funding history and valuation are outside what the wiki has ingested.
- The 2.1× conversion figure has no stated baseline. Against which human cohort, over what window, with what selection into agent-handled leads? Unstated, and the firm controls the funnel.
- The AI-CEO result is one city over six weeks with no control. Whether it survives a longer window or generalises beyond Cuernavaca is unknown.
- Headcount effects are alluded to but not quantified. The interviewer notes Kavak “had to downsize dramatically”; Maza does not give figures, and the episode does not return to it. Relevant to ai-employment-effects.
Related pages
- Concepts: agent-harness, multi-agent-failure-modes, agent-fleet-management, agent-oversight-and-delegation, micro-productivity-trap, enterprise-ai-adoption, automation-vs-augmentation.
- Entities: a16z (publisher of the only source), Anthropic (Opus 4.5 triggered the re-architecture).