Everyone has an AI employee demo. Almost nobody has numbers.
So here's a different kind of ranking. Every entry below is a real deployment of AI employees — agents that own an outcome, not a chat window — ranked by the quality of the evidence behind it, with sources you can check. Company-scale metrics stated on the record grade A. Vendor self-reports with named customers grade B. Anecdotes grade C.
One thing to know before the list: the top four entries are all the same company. That's not laziness. It's the point. When a16z put Kavak's Chief Product & AI Officer on their show in August 2026, it became the best-documented AI-employee deployment anywhere — and the anatomy section below breaks down exactly how they did it.
The ranking
| # | AI employee | Reported result | Grade |
|---|---|---|---|
| 1 | Kavak's sales agents | 2.1x conversion vs best human team; NPS tripled | A |
| 2 | Kavak's loan underwriter | Car loans approved in under 3 minutes (market norm: ~2 months) | A |
| 3 | "El Mike," Kavak's mechanic copilot | Warranty claims down 26% | A |
| 4 | Kavak's AI CEO pilot | +50% profit in one city, month one | A− |
| 5 | Replicas' coding agents | 95% of the company's own PRs, 300k+ agent-minutes/week | B |
| 6 | Viktor, the Slack coworker | $1M ARR within 3 hours of launch | B |
| 7 | The DIY assistant wave | A working assistant in ~6 hours of setup | C |
Now the receipts.
1. Kavak's sales agents — 2.1x conversion (Grade A)
Kavak is Latin America's largest used-car marketplace — peak valuation $8.7B, roughly 4,000 employees. Selling a used car in Mexico means financing, insurance, trade-in quotes, and 20,000 SKUs. Kavak used to route customers through 15 human specialists.
They replaced the relay race with one agent per customer. Not a support bot — a seller, holding years of that customer's history and a standing goal: maximize lifetime value.
The numbers, from Chief Product & AI Officer Alejandro Maza Ayala on the a16z Show: 96% of customer interactions and 95% of transactions now run with no human involved. NPS tripled. The agents first converted 50% better than the best human team, and now convert 2.1x better.
Why grade A: named executive, named metrics, at a company big enough that the numbers were auditable internally before anyone said them on a podcast. Still company-reported — no third party has verified them — which is exactly why the grade matters.
2. Kavak's loan underwriter — 3 minutes vs 2 months (Grade A)
Same company, harder problem: underwriting thin-file and no-file borrowers in emerging markets, where a car loan typically takes two months or more to approve.
Kavak's agents approve one in under three minutes — pricing risk per customer, personalizing rate and maximum amount against the whole portfolio. The vertical integration is the safety net: if a customer can't pay, they return the car and step down to a cheaper one.
Regulated financial services, end-to-end, by an agent. This was supposed to be the category AI couldn't touch.
3. "El Mike" — the mechanic's copilot (Grade A)
Not every AI employee replaces a human. Kavak has ~800 mechanics in Mexico, and dexterity isn't automatable. So the agent rides along — Maza compares it to Ratatouille — guiding inspections, flagging what to check, teaching as it goes.
Result: inspections got faster, repairs got cheaper, and warranty claims fell 26% because the cars going out the door got better.
4. The AI CEO of Cuernavaca (Grade A−)
The wildest one, and the one to hold most loosely. Kavak carved out one city — Cuernavaca — and put an agent in charge: inventory, forecasting, daily plans messaged to the physical workforce, progress collected by voice note.
Month-one goal was to double the city's profit. It missed — and still delivered +50%, with customer satisfaction, inventory rotation, and financing penetration all up.
The minus: six weeks of runtime, one city, no seasonal control. Directionally stunning, statistically young.
5. Replicas' coding agents — 95% of their own PRs (Grade B)
Out of the Kavak universe. Replicas (YC) runs cloud coding agents that fix CI failures and test their own work end to end. At their V2 launch the founder reported over 95% of the company's PRs in the prior 6 months were done by their own agents, with 300,000+ agent-minutes per week and engineering teams at Mintlify, Knowunity, and Composio shipping PRs through it daily.
Grade B: self-reported by the vendor — but with named customers and a verifiable public product.
6. Viktor — the $1M-ARR Slack coworker (Grade B)
Viktor launched as an "AI coworker that lives in Slack" — marketing audits, ad management, lead research, week-long persistent runs — and reported $1M ARR three hours after launch, at roughly $300 ARPU on the top tier.
Note what this evidence actually proves: not that Viktor's agent outperforms a human (no metric published), but that buyers now allocate real budget to the AI employee category on sight. That's demand evidence, and it's why this post exists.
7. The DIY wave (Grade C)
The long tail: operators building their own. Sahil Bloom — a non-engineer — documented standing up a personal AI assistant in about six hours (dedicated Mac mini, fresh accounts, walled off from his other systems), then publicly sized the setup-service opportunity at $100k+/month. Andrew Wilkinson posted a contract gig to build what he called an openclaw army to automate his companies.
Anecdotes, so grade C. But notice what even the anecdotes agree on: the first thing a careful DIY builder does is give the agent its own isolated machine.
Anatomy of #1: how Kavak actually did it
Four moves, all copyable.
They redesigned the company instead of adopting a tool. Maza's diagnosis of why most AI programs stall: companies keep their structure and hand employees a chatbot. "There's no efficiencies. Your customers have the same problems and nothing happens." Kavak instead asked what the company should look like in 2035 and rebuilt toward it — the same reason Ford's electrified factories beat coal plants that just swapped the engine: bolt-on gets you 6%, redesign gets you 3x.
One agent per customer, each in its own virtual machine. This is the architecture decision everything else rests on. In Maza's words: an agent "gets spawned specifically for this customer with its own virtual machine," remembering years of history. Every day, 100,000 to 200,000 agents get instantiated. They work for 3 minutes or 3 days, set an alarm clock, and go back to sleep. The harness is deliberately simple: a VM, an agent, memory, evals, a CLI with access to every company API, and a long-term goal. They'd already built an elaborate multi-agent graph system — profitable, working — and tore it down when frontier models made the graphs a constraint rather than a scaffold.
They spend as much on evals as on agents. Kavak's rule of thumb: equal engineer time, tokens, and money on evals as on the agents themselves. And they eval the metric that matters — did the customer convert, return, stay happy — not call counts. "Evals are the brakes. You hit the gas only as hard as your brakes allow."
Humans close the loop instead of catching the overflow. When an agent hits a wall it calls a help API, and a human answers — which generates exactly the training data that fixes the wall. One agent's mistake becomes 200,000 agents' lesson by the next day.
The pattern under every entry
Run the list again and one requirement repeats at every grade level. Kavak: a VM per agent, 200,000 a day. Replicas: isolated cloud environments per run. The careful DIY builder: a dedicated machine, walled off on purpose.
Nobody serious runs an AI employee loose on a shared box. An agent that holds credentials, remembers everything, and acts 24/7 needs a blast radius of exactly one — its own machine, its own keys, its own memory.
That substrate is the part you don't have to build anymore. Kavak had to construct it in-house. You can rent it: an isolated Firecracker microVM with the agent harness you choose, always on, BYOK — your API keys stay yours, your model spend stays yours — from a flat monthly price. Start from a forkable template instead of a blank box.
We'll be direct about what jurniti is not: we don't sell AI employees. We sell the machine your AI employee lives on. The job you give it, and the numbers it puts up, are yours.
Want the full operating model first — fleets, memory, observability, the whole AI-native company stack? The free 7-day course walks it end to end.
No free trial, no free tier — first purchase carries a 30-day money-back guarantee.
Sources: a16z Show, "The Self-Improving Company: Kavak's AI Playbook" (Aug 2026) and the a16z summary thread · Kavak valuation and scale: TechCrunch (2021), Bloomberg (Apr 2025) · Replicas V2 launch · Viktor launch · Sahil Bloom's DIY build. All performance figures are company-reported; grades reflect that.