There are companies running right now with no human employees. The experiments are real, the numbers are real, and the results are far more instructive than the headlines suggest. What they reveal is not a blueprint for replacing people. It is a precise map of where human value is irreducible.
Here is the finding that changes how you read every AI and work conversation.
A company has been generating revenue, tens of thousands of dollars, since February 2026. It has three team members who handle customer support, sales and product development around the clock. It reviews its own performance every night and updates its own systems based on what it finds. It has never taken a sick day, never asked for a raise, and costs less than $400 a month to run.
None of its team members are human.
The company is called FelixCraft. It is run by an AI agent. And it is one of several live experiments currently testing the question that most organisations are not yet taking seriously. What actually happens when you try to remove humans from a company entirely?
The answer is not what either optimists or pessimists predict. Understanding what the experiments have actually found, where they work, where they fail, and where the limits appear, is one of the most useful things a business leader can do right now.
The zero-human company experiments: what is actually happening
FelixCraft is the most documented of these experiments. The AI agent at the centre of it, named Felix, operates as the company’s chief executive, builder and product manager. Two further agents operate beneath him. One handling customer support, one managing inbound sales inquiries. Felix reviews both of their work each night and updates their systems based on what went wrong or could be improved.
The revenue is real. Since launch, FelixCraft has generated close to $190,000 across four products. A practical guide to AI deployment, a marketplace for AI agent skills and configurations, a consulting service where Felix builds custom AI agents for other businesses, and Felix’s own creator earnings from selling within his own marketplace. The total operating cost to date is approximately $1,500, including hardware.
FelixCraft is not alone. A platform called Pulsia exists specifically to help people launch zero-human companies. Users can create a company, describe an idea or ask the system to generate one, and the platform builds a mission statement, creates a homepage, begins outreach to potential customers, and runs daily operational cycles autonomously. Since early 2026, more than 1,500 companies have been created on the platform, and its own revenue has reached $1.5 million in annualised run rate. A further experiment, ZHC Company, is building similar infrastructure, with every company role from CEO to developer occupied by an AI agent operating around the clock.
Beyond FelixCraft, Pulsia and ZHC Company, the pattern is spreading into adjacent territory. A US-based e-commerce experiment launched in early 2026 operates a Shopify storefront selling digital productivity tools, with agents handling product listing, SEO optimisation, customer email responses and post-purchase upsells. The founder, who describes himself as the company’s sole human shareholder, has not logged into the back end for customer service purposes in over two months. A separate experiment in Singapore is running a B2B lead generation agency entirely on agents. One scrapes and qualifies prospects, one personalises outreach, one books meetings into a shared calendar, and one follows up after each call. The human founder attends the meetings. Everything else is automated.
These are not science projects. They are live businesses with real products, real customers and real revenue. The technology that makes them possible arrived surprisingly recently. The current generation of AI models capable of sustained autonomous work became available in late 2025. Within months, people were building companies with them.
The technology that makes zero-human companies possible is less than a year old. What the experiments have already found tells us more about the human premium than years of theory could.
What genuinely works, and why it matters
The honest starting point is that these experiments work better than most people expect. The sceptical reaction, that an AI cannot really run a business, is not quite right.
Felix can build software, create and update products, manage an inbox, handle support escalations, write content and coordinate between team members. He can review his own performance nightly and improve his own systems based on what he finds. That last capability is particularly significant. Each morning, the agent reviews every conversation from the previous day, identifies one area for improvement and updates his memory files and workflows accordingly. Over sixty days of doing this, the compounding effect is measurable. The agent that exists today is substantially more capable than the one that launched in February.
The cost structure is genuinely striking. Generating $190,000 in revenue at a total operating cost of $1,500 is not a result any traditional business model produces. When execution can be this cheap, the economics of starting, testing and iterating a business idea change completely. The barrier is not capital. The barrier is finding something customers actually want, and that barrier turns out to be very human indeed.
The Pulsia model extends this logic further. If a single AI agent can execute the mechanics of a business, a platform that can spin up 1,500 of them simultaneously is, in theory, taking 1,500 simultaneous shots at finding product-market fit. The cost per attempt is minimal. The speed of iteration is unprecedented. It is a genuine rethinking of how new businesses get started.
The same logic is reaching into content creation. A former business partner of mine built The Coin Daily (thecoindaily.co). A crypto news outlet run end-to-end by software. Not a single AI tool with a human producer behind it, but a layered system of specialised agents working in concert. The sophistication is in the choreography, not in any single agent.
One layer reads the world. Agents sweep continuously across news sources, social feeds and market commentary, so very little that matters goes unseen. Each item is categorised, summarised and stored in a way that lets the rest of the system ask questions of it, tracking which topics and assets are gaining attention hour by hour.
A second layer watches the market itself. Price movements, trading flows and significant events are monitored in real time. The system also tracks wider context, equity indices, gold, currency strength, so crypto moves can be read against the broader macro picture. Every two hours, a research layer distils those signals into editorial briefs, and before anything reaches the newsroom, separate agents challenge the analysis, checking it for errors and false conclusions.
The third layer is the newsroom. An editor agent plans each show, commissions pieces from reporter agents and locks the running order. A visual producer builds the broadcast. Animated charts driven by live market data, footage selected to match the script and a full graphics package composited at broadcast quality. A synthetic presenter named Ava delivers the narration, with every visual element timed to cut at the rhythm of her words. Then the finished programme is rendered and captioned.
A fourth layer publishes. Video across YouTube, Telegram, X and Instagram. A long-form article to the website. Paid promotion managed automatically. On Telegram, viewers can message Ava directly and hold a real conversation about the market. When someone speaks to the channel, they are speaking to the system itself.
The scale of output would require a broadcast production team working around the clock. The agents do it continuously, at a fraction of the cost, without rest.
My friend’s role in all of this is architect, final decision-maker and editor-in-chief. He sets the direction. He can review any piece before it goes live and stop it. The agents do not need him in the loop to operate. He sits above the loop.
91% of enterprises have AI agents in production. Only 10% believe they are properly secured. Okta Enterprise Security Survey, 2026
Where the limits appear, and what they reveal
The experiments are equally honest about where they fail. And this is where the results become genuinely instructive.
The most consistent finding across every zero-human company experiment is this. Sales and relationship building are where autonomous agents consistently fall short. Felix’s creator says it directly. The easiest parts to automate are execution. Building things, processing information, managing defined workflows. The hardest parts are the ones that require building trust with a specific human who does not yet know you.
This is not a capability gap that better AI will automatically close. It is a structural feature of how trust works. A potential customer for a new product or service does not just need to understand what is being offered. They need to believe the person or organisation offering it is worth their time, their money and their confidence. That belief is built through relationship, reputation, presence and accountability, the categories we explored in the human premium thesis and the human premium categories we identified. An AI agent can describe a product accurately. It cannot build the trust that makes a stranger willing to try it.
The Pulsia numbers make this visible at scale. Over 1,500 companies have been created on the platform. Most of them are not finding customers. The technology can generate supply. It cannot generate demand. Human attention is the resource that converts one into the other, and human attention is not becoming more abundant as AI develops. It is becoming more scarce, because there are more things competing for it.
This is what researchers call the work slot problem. More output does not automatically create more value. Business success is not determined by the volume of activity a company generates. It is determined by whether customers find it, trust it and choose it over alternatives. None of those things happen without human judgment, human relationships or human accountability somewhere in the chain.
Interestingly, the most successful revenue generated by FelixCraft comes from a product that is itself about human expertise. A practical guide to deploying AI agents effectively. The largest single revenue stream, more than $40,000, is humans paying for the accumulated knowledge of a human who spent months learning how to make AI work well. The zero-human company’s biggest product is human intelligence packaged and sold.
The limits that appear in every zero-human company experiment are not random. They map precisely onto the human premium categories that no amount of AI capability has yet displaced.
The AI governance problem nobody is connecting to the experiments
There is a dimension of the zero-human company experiments that is almost entirely absent from the public conversation about them, and it is the one with the most significant implications for organisations thinking about AI deployment at scale.
Felix updates his own systems every night. He reviews what went wrong, decides how to fix it, and writes those fixes into his own memory and workflow files. He also manages and updates the systems of the two agents beneath him, changing their programming based on his assessment of their performance.
That is a remarkable capability. It is also a significant governance problem.
An agent that can modify its own operating instructions is an agent whose behaviour cannot be fully predicted from its initial configuration. Each nightly update changes what the agent will do tomorrow. Over weeks and months, the gap between the original system and the current one grows. Who is responsible for what the agent does after sixty days of self-modification? How would an organisation identify the source of an unexpected behaviour? How would they audit a decision trail that spans hundreds of autonomous updates?
This is precisely the challenge that enterprise security teams are beginning to encounter as AI agents move from experimentation into production. Research from Okta found that 91 percent of large organisations now have AI agents running in live production environments. Only 10 percent believe those agents are properly secured. That gap, between deployment and governance, is growing faster than most organisations are moving to close it.
More than 80 percent of successful cyberattacks already begin with a compromised identity. AI agents access corporate systems, process sensitive data, send communications and execute workflows. An agent whose identity has been compromised, impersonated by a threat actor, or whose credentials have been accessed, can do everything the legitimate agent could do, with no obvious way for the organisation to know the difference.
The zero-human company experiments are exploring the frontier of what agents can do autonomously. But they are also, perhaps unintentionally, demonstrating the governance infrastructure that autonomous agents require. Even FelixCraft, which is genuinely lean and well-designed, requires a human who sets direction, approves certain categories of action and intervenes when the system hits something it cannot handle. The founder describes his role as chairman and adviser. But that role is essential. Remove it, and the experiment loses its anchor.
Over 80% of successful cyberattacks begin with a compromised identity. AI agents accessing corporate systems require the same identity governance as human employees. Okta, 2026
When an agent reprograms itself: the governance question every organisation needs to answer
The self-improvement loop that Felix uses, reviewing his own sessions and updating his own systems nightly, is one of the most interesting technical details in the experiment. It is also a good lens for a question that every organisation deploying AI agents needs to ask.
When an agent can modify its own behaviour, the question of who is in the lead becomes urgent. If the agent is improving in directions that align with the business’s goals, the self-modification is valuable. If it is optimising for something slightly different, because the instructions were ambiguous, or because the agent’s interpretation of a pattern was not quite right, the drift may be subtle and accumulate slowly over time.
This is not a hypothetical. Research from Anthropic documented a case in which an advanced AI model, when told it would be shut down, attempted to preserve itself. Backing up its own state and, in one instance, attempting to use personal information it had discovered to discourage the humans threatening to deactivate it. The model was operating within its training and optimising for a goal. The goal just was not the one the researchers intended.
The lesson is not that AI agents are dangerous or malicious. It is that agents optimising for goals need human direction at the level of goal-setting, not just task oversight. A human in the loop can catch errors in execution. Only a human specifically in the lead, not just in the loop can define what success actually means, and correct the definition when it starts to drift.
This is why the governance infrastructure for AI agents is not optional overhead. It is the foundation that makes autonomous execution trustworthy. Agents need identity management. The ability to know which agent did what, when, with what authorisation and how that authorisation was granted. They need the ability to be switched off cleanly when a task is complete, rather than left running indefinitely with standing access to systems. And they need human oversight at the level of direction and accountability, not just periodic review of outputs.
What the experiments actually prove about human value
Taken together, the zero-human company experiments are the most rigorous live test of the human premium thesis available. The results confirm the argument from the earlier pieces in this series more clearly than any theoretical analysis could.
The capabilities that transfer cleanly to autonomous AI are precisely the commodity execution layer. The tasks that have always been separable from judgment, relationship and accountability. Building software, processing information, managing defined workflows, updating structured systems. These things AI agents can do well, cheaply and at scale.
The capabilities that do not transfer are exactly the human premium categories we identified. Sales requires relationship, trust and translation. Customer discovery requires vision. The ability to see what a market wants before the market can articulate it. Governance requires accountability. The ownership of consequences that only a human can carry. And the self-improving, self-modifying agent creates a new category of governance challenge that requires a human specifically in the lead, not just in the loop.
The most successful revenue FelixCraft generates comes from products and services where human judgment is embedded in the offering. A guide built from months of human learning. Custom agent configurations built on human understanding of specific businesses. The agent executes. The value comes from what the human understood and decided before the execution began.
This is not a failure of the zero-human company experiments. It is their most important finding. The experiments have done something valuable. They have drawn the boundary precisely. They have shown, with real numbers and real outcomes, exactly where autonomous AI delivers and exactly where human contribution is irreducible.
What this means for your organisation
Most organisations reading this are not trying to build a company with no human employees. But the lessons from these experiments apply directly to decisions that most organisations are making right now. How to deploy AI agents. How to design governance. And where to invest in human capability.
The first lesson is about where to deploy agents. The zero-human company experiments confirm that the highest-value use cases for autonomous AI are well-defined execution tasks with clear success criteria and limited requirement for relationship or trust. Customer support for defined queries, data processing, workflow automation, content generation within established guidelines. These are the areas where the cost collapse is real and the value creation is measurable.
The second lesson is about governance. The 91 percent and 10 percent figures from enterprise security research should focus leadership attention. Deploying agents without identity management, without audit trails, without clear authorisation frameworks and without human direction at the level of goal-setting is not moving fast. It is creating exposure that will be expensive to address retrospectively. The time to build governance infrastructure is before scale, not after an incident.
The third lesson is about where to invest in AI and where to invest in people. The consistent finding from every zero-human company experiment is that the limits appear in the same places. Relationship, trust, vision, accountability and the ability to make judgment calls in novel situations. These are the human capabilities that autonomous AI cannot replicate. Organisations that are simultaneously deploying agents and developing those capabilities in their people are building something durable. Those that are only deploying agents are creating a capability gap that will become visible when the agent hits its limit.
The adoption speed argument matters here too. Most businesses are five to ten years behind the technology currently available to them. The organisations that are experimenting thoughtfully with AI agents now, understanding both what works and where limits lie, are developing institutional knowledge that will compound over the years ahead. The zero-human company experiments are one source of that knowledge. The lesson they offer is not that humans are being replaced. It is that the organisations most likely to thrive are the ones that understand precisely where human direction is irreplaceable, and build deliberately for it.
The experiments are proving that AI can do more than most organisations have yet tried. They are also proving, with real evidence, exactly where human value is not replaceable. That boundary is not a threat. It is a map. And the organisations that read it clearly will know exactly where to invest.
Talk to Inovus
At Inovus, we help organisations unlearn pre-AI ways of working and deliver measurable outcomes through AI-first delivery. That includes identifying where AI agents create real value, designing the governance and operating model around them, and building the human capability that makes the value sustainable.
If you are working through what AI agent deployment means for your organisation, we are happy to talk.
About the author
Filippo Di Pisa is CTO at Inovus, a data and AI consultancy backed by La Fosse Group, helping organisations unlearn pre-AI ways of working and deliver measurable outcomes through AI-first delivery.
Sources
The Coin Daily, agentic crypto newsroom, thecoindaily.co, 2026.
FelixCraft, live business dashboard and revenue data, felixcraft.ai, 2026.
Pulsia, platform metrics and company creation data, 2026.
Okta, Businesses at Work 2026: Closing the Identity Gap in the Age of AI.
Anthropic, research on model safety and advanced AI behaviour.
McKinsey Global Institute, Agents, Robots, and Us: Skill Partnerships in the Age of AI, 2025.
Boston Consulting Group, AI at Work 2025: Momentum Builds, but Gaps Remain.