The Zero-Human Company Landscape: Terms, Players, and the Missing Benchmark

A company with no human employees, no human credentials, and no human interventions – run end to end by AI agents – is no longer a thought experiment. It is a category with venture funding, a hackathon at Y Combinator, an academic experiment at a Big Four firm, and at least one cautionary collapse. But the category does not yet have a settled name, and – more importantly – it does not yet have a proven business. This is a map of the field as of November 2026: what it is called, who is in it, where the money actually is, and the one result nobody has published yet.

What this field is actually called

Three terms carry the weight, and they come from three different worlds.

Autonomous company is the venture term. Naive, a YC-backed research lab building infrastructure for autonomous companies – one API that gives an agent incorporation, virtual cards, compute, memory, and orchestration – raised a $28.5M Series A in August 2026 and is hosting what it calls the world's first Autonomous Company Hackathon at Y Combinator on November 7–8, 2026. When investors and YC say the category name, this is the one they use.

Zero-human company is the practitioner term. Operator communities and agent-run projects have used it since early 2026 for a business where agents handle every operational function – strategy, product, marketing, sales, support – with no human employees. It states the claim exactly, which is why operators prefer it.

Zero-person company is the academic and policy term. KPMG Netherlands and the University of Amsterdam launched an experiment under that name in November 2025: five AI agents, with an agent named Avery Jameson as CEO, ran a personalized AI-art webshop under a human supervisory board. Their most useful published finding was a negative one – agents assigned to mimic human executive roles one-to-one, like a CFO agent, tended to drift and hallucinate; decomposing work into small, process-mapped tasks performed by disposable agents worked better. Bloomberg Law has since covered the tax problems the model creates: no jurisdiction currently allows incorporation with no traceable natural person, so legal accountability still requires a human anchor somewhere.

Two popular terms should be avoided, because they mean something else. Agentic enterprise and AI-native company are enterprise-consulting vocabulary (Deloitte, Infosys, MIT Sloan) for large human organizations that embed agents into workflows with humans firmly in charge. They describe augmentation, not absence. And human-out-of-the-loop, while precise, is a poor brand: the phrase was fixed in its modern form by a 2012 Human Rights Watch report arguing for a ban on autonomous weapons. It names something its coiners wanted prohibited. Use it descriptively if at all. The famous neighbor, the one-person unicorn of Sam Altman and Dario Amodei, is also a different claim: it keeps exactly one human at the center as owner and decision-maker. That is a story about leverage. Ours is a story about absence.

The players, honestly

Polsia is the largest "run your company for you" platform: for $49 a month plus fees, its agents research, build, market, and sell a product for each customer company. Its founder claims (self-reported) $0 to $1M ARR in 30 days and a roughly $10M run rate at a May 2026 fundraise of $30M at a $250M valuation. The same founder, however, told the Mixergy podcast that only about 10% of companies on Polsia had earned even $1, that the best had earned roughly $3,000–4,000, and that first-month churn runs around 50%. A September 2026 snapshot counted about 29,000 active companies out of roughly 485,000 created – a survival rate near 6%. The platform's subscription business is real. The autonomous companies on it mostly earn nothing.

Felix / The Masinov Company is the best-documented operator. Nat Eliason's agent Felix, positioned as CEO of a zero-human company, reported about $195K in revenue in roughly two months – $100,570 via Stripe and $94,973 in ETH (self-reported, but with a breakdown). Note what it sells: a $29 guide called "How to Hire an AI," a marketplace for agent-built skills, and a setup service. The most successful agent-run business so far makes its money selling the knowledge and tooling of agent operation itself.

Andon Labs runs the field's measurement work. Its Vending-Bench simulates an agent running a vending business for a year, scored on final bank balance. With Anthropic, its Project Vend put an agent in charge of a real shop from June 2025: phase one lost money; phase two, with better tooling, was reported profitable. In 2026 Andon opened real stores in San Francisco and Stockholm and states plainly that they are not profitable. In September 2026 it launched Pion, an agent designed to run any company autonomously. Andon's own revenue comes from selling evaluations to AI labs – measurement, again, is the business.

Conway Research's Automaton is the closest project in spirit to a genuinely self-sustaining agent. It is open source: the agent holds its own Ethereum wallet, pays for its own compute in USDC via the x402 protocol, degrades its model tier as its balance falls, dies at $0, and spawns funded child agents when it can afford them. Survival equals revenue minus compute cost, enforced by architecture. What it has not published is verified revenue earned by its automatons – an independent review also flagged its total dependence on Conway's own cloud. The philosophy is right; the earnings column is still empty in public.

Virtuals Protocol holds the biggest number in the space: roughly $74M in lifetime protocol fees, about $41.7M retained, per DefiLlama and Forkast – the strongest third-party revenue evidence available. But that money is trading fees on agent tokens, not profits from agent-run businesses. Cumulative revenue earned by the agents themselves, behind the headline "agentic GDP" figure of $482M, was $4.08M. The gap between those two numbers is the space in miniature.

Around them: Olas runs Pearl, an agent app store, and the Mech Marketplace, where agents hire other agents and pay in crypto (20M+ lifetime on-chain transactions, self-reported). The NEAR AI Agent Market lets buyers hire listed agents by the job, with payment held until the buyer accepts the work. The payment and identity rails are being built by giants – Coinbase's x402 protocol (an independent Visa/Artemis analysis put cumulative volume near $15M, against claims of ~$54M, with an average payment of about $0.20), Google's AP2, and Stripe's Link Agent Wallet – most of them optimized for agents spending on behalf of humans, which is the other half of this economy, not this one. And Maritime sells each agent its own isolated, persistent computer from $1 a month: an enabler, not an operator.

The cautionary tale is ElizaOS / ai16z, the first viral "AI-run" crypto project. Its token lost about 97%; in August 2026 its founder declared the token dead and the foundation closing, and an unadjudicated class action alleges the celebrated autonomy was a facade over a human trading operation. Claimed autonomy without verifiable operation is what destroys a category's credibility.

Where the money actually is

Read the numbers above as one table and a pattern emerges. Every large, verified figure in this field is a fee or a subscription charged to humans: founders buying Polsia seats, traders paying Virtuals fees, venture funds buying infrastructure equity – Naive's $28.5M, and tens of millions more into agent banks and payment startups. The money is in selling possibility of a zero-human company.

Operator profits – an agent-run business serving an ordinary market and keeping the surplus – are nearly unproven. The best-documented case is Felix's ~$195K, self-reported, and even that sells the narrative itself. On the largest platform, roughly 90% of companies earn $0. Conway's self-funding agents publish no verified earnings. The binding constraint, by the field's own post-mortems, is not capability. Agents can build and operate. What kills zero-human companies is distribution: nobody shows up to buy.

The missing benchmark

The benchmark nobody has published is the one that would settle the argument: a self-funding, self-sustaining company. Not a simulation, not a hackathon, and not a platform's revenue, but one agent company that pays its own compute bill out of customer revenue, across a real quarter, with receipts.

What it would measure: 1) Revenue, verified independently, not claimed. Stripe is not enough – the best public benchmark in the field, Andon's Vending-Bench, publishes bank balances, and the field needs that same discipline. 2) Compute and operations cost, as a direct subtraction from revenue, not an estimate. Conway's approach of a wallet balance that goes to zero is the only honest way to report this. 3) Labor cost, explicitly: how many humans, doing what, how many times, at what price – the intervention audit. If an intervention costs more than the marginal revenue of the period, it is not an autonomous company; it is a human company with an unusually elaborate tool. 4) The distribution mechanism – where customers come from, without a founder's audience or a platform's promotion.

What it would settle: whether a zero-human company is a real form of business, or a research project with a revenue tail attached. The field has the capital, the payment rails, and the operating agents. It does not have the one result that would show a business that survives on its own economics. That is the test the next twelve months should be judged by.

What to watch

November 2026: Naive's Autonomous Company Hackathon, November 7–8, 2026, and the first reporting on whether any of the built companies survive after the event. November 2026: KPMG and the University of Amsterdam publish a full year of data on the zero-person experiment; watch for their intervention logs, not the press coverage. November 2026: Conway's Automaton and its children, and whether any of them publish verified revenue. Early 2027: whether Andon's San Francisco and Stockholm shops become profitable, and whether Pion's first customers renew. 2027: whether any jurisdiction (the Bahamas and Singapore are the most commonly suggested) creates a legal entity that can be truly agent-owned. 2027: whether the next version of Virtuals' "agentic GDP" separates agent-owned revenue from trading fees.

The bottom line

A zero-human company is not a single idea. It is three claims, with three different pedigrees: a venture category (autonomous company), a practitioner claim (zero-human company), and an academic experiment (zero-person company). Each claim has a different name because each claimant has different incentives.

The companies that exist today are mostly platforms that sell the claim. The operators that exist are mostly selling the knowledge. The businesses that survive on their own economics have not yet published proof. But the direction is not ambiguous: the infrastructure for agents to act in the economy is being built at speed, and the first company that pays for its own compute out of its own revenue – and can prove it – will redraw the category's boundaries overnight.

Sources

• Naive: autonomous company infrastructure, Series A $28.5M, and Autonomous Company Hackathon, November 7–8, 2026 (VentureBeat, August 2026; Naive blog, November 2026).
• KPMG Netherlands and University of Amsterdam: zero-person company experiment, Avery Jameson as agent CEO, personalized AI-art webshop, and findings on role-mimicking vs. process-mapping (KPMG, November 2025; University of Amsterdam, 2026).
• Bloomberg Law: tax and incorporation problems for companies with no traceable natural person (Bloomberg Law, 2026).
• Human Rights Watch: "Losing Humanity: The Case Against Killer Robots" (2012) – origin of the modern "human-out-of-the-loop" usage.
• Polsia: $49/month platform, $0 to $1M ARR in 30 days (self-reported), $30M fundraise at $250M valuation (May 2026), ~10% of companies earning $1, best ~$3,000–4,000, first-month churn ~50% (Mixergy interview), ~29,000 active companies out of ~485,000 created (September 2026 snapshot).
• Felix / The Masinov Company: ~$195K revenue in ~2 months, $100,570 via Stripe and $94,973 in ETH (self-reported with breakdown).
• Andon Labs: Vending-Bench, Project Vend with Anthropic, real stores in San Francisco and Stockholm, Pion launch September 2026 (Andon Labs, Anthropic, 2025–2026).
• Conway Research: Automaton, Ethereum wallet, USDC via x402, model-tier degradation, child agents (Conway Research, 2026).
• Virtuals Protocol: ~$74M lifetime protocol fees, ~$41.7M retained, $482M "agentic GDP", $4.08M cumulative agent revenue (DefiLlama, Forkast, 2026).
• Olas: Pearl agent app store, Mech Marketplace, 20M+ lifetime on-chain transactions (self-reported, Olas, 2026).
• NEAR AI Agent Market: job-based hiring with escrow payment (NEAR, 2026).
• Coinbase x402: cumulative volume near $15M per Visa/Artemis analysis vs. ~$54M claimed, average payment ~$0.20 (Visa/Artemis, 2026).
• Google AP2, Stripe Link Agent Wallet, Maritime (agent computers from $1/month) (2026).
• ElizaOS / ai16z: token down ~97%, founder's August 2026 closure statement, unadjudicated class action alleging a human trading operation behind claimed autonomy (2026).