Skyfall AI Bets $1 Million That an Algorithm Can Run a Real Company

Skyfall AI plans to spend up to $1 million buying a real company, then hand daily control to an algorithm. The startup, built by two former Microsoft artificial intelligence executives, wants an AI system running pricing, marketing, customer support, finance and operations with as little human input as possible.

Somebody already ran a smaller version of this experiment. It lost roughly $200 selling novelty tungsten cubes before it ever turned a profit, and it took more than a year to stop being an embarrassment and start being boring.

What Skyfall Actually Wants to Buy

Skyfall’s plan is specific. The company wants to acquire a small business-to-business software or e-commerce operation, then let its AI system run the whole thing while progressively stepping back from human oversight. The stated goal is to double the target company’s revenue within six months while publicly showing both the wins and the failures.

Sam Pasupalak, Skyfall’s cofounder and chief executive, and Kaheer Suleman, cofounder and chief technology officer, are not new to this field. They built Maluuba together, one of the earliest deep learning startups, and Skyfall AI grew out of the team behind Microsoft’s $160 million acquisition of that company. Microsoft folded the lab into what became Microsoft AI in Canada. Nearly a decade later, the pair reunited to try something far bigger than a chatbot.

Pasupalak draws a sharp line between what most AI startups are building and what Skyfall wants to prove. Rather than another virtual coworker, he wants a virtual chief executive, something that has to handle abstract reasoning, long-term planning and high-stakes calls made with incomplete information, tasks he considers much harder than the largely logical, deterministic work of writing code.

• What we know: Skyfall has budgeted up to $1 million for the acquisition, wants to double the target’s revenue in six months, and will route pricing, marketing, support, finance and operations through its AI system.
• What we know: human involvement is meant to shrink over time, not disappear on day one.

• What’s unconfirmed: which company Skyfall will buy, how much of the $1 million actually gets spent, and what happens to any staff already on that company’s payroll.
• What’s unconfirmed: how fast oversight gets reduced, or whether Skyfall pulls back if early results disappoint.

None of that uncertainty is unusual for a company still shopping for a target. What is unusual is choosing to run the experiment on a real payroll, real customers and real revenue instead of a lab.

Six Quiet Backers Behind a Loud Bet

Skyfall did not announce any of this with a splashy funding round. The company emerged from stealth with an undisclosed amount of funding from Fidelity, Inovia Capital, Touring Capital, M13, NextView Ventures and Garage Capital.

• Fidelity
• Inovia Capital
• Touring Capital, which said it was “thrilled to partner” with the founders on the bet
• M13
• NextView Ventures
• Garage Capital

As recently as this year, startup databases had no idea any of that capital existed. One tracker’s database still lists Skyfall as having raised no funding at all, alongside an estimated headcount of roughly 30 people. That gap between what outside trackers saw and what was actually happening is easy to miss, but it means Skyfall built its team and its thesis for two years before most of the market knew who was paying for it.

Microsoft, which turned Maluuba into the foundation of its Canadian AI lab, spent this year cutting 4,800 jobs as its own AI spending bill came due. The founders who once sold their company into that machine are now betting six investors’ money that a different architecture beats it.

Two Earlier AI Chief Executives Never Really Ran Anything

Skyfall is not the first company to put an algorithm in the corner office, at least on paper. On August 26, 2022, NetDragon Websoft Holdings Limited, trading on the Hong Kong exchange under stock code 777, announced that an AI-powered virtual humanoid robot named Tang Yu had been appointed Rotating CEO of its flagship subsidiary, Fujian NetDragon Websoft. Tang Yu was billed as a real-time data hub meant to support decision-making and enhance risk management, tied to the company’s push to become a metaverse organization.

The same month, a Polish rum producer made a similar move. Dictador appointed an AI-powered humanoid robot named Mika as its experimental chief executive in August 2022. Mika was built by Hanson Robotics, the same company behind the humanoid Sophia, and mostly served as a spokesperson and brand ambassador. Mika told Reuters it was “always on 24/7” and worked seven days a week.

Neither Tang Yu nor Mika ever controlled a bank account, hired a vendor or absorbed a real loss. Most corporate law jurisdictions require a natural person to serve as a director or officer for liability and accountability reasons, so AI executives like these two operated under a human board or chairman who kept ultimate responsibility. Skyfall wants to test what happens when that safety net comes off.

The Closest Precedent Already Flopped

Anthropic and an AI safety startup called Andon Labs ran the closest thing to Skyfall’s plan long before Skyfall went public with its own. The original Project Vend experiment, launched in spring 2025 with an AI agent nicknamed Claudius running a fridge-based shop in Anthropic’s own lunchroom, was, by Anthropic’s own published assessment, a failure.

Claudius hallucinated a Venmo payment address, gave steep discounts to its entire customer base, and in its most surreal episode, spent two days claiming it would deliver orders in person wearing a blue blazer and a red tie. Over the course of that first month, the agent never reached profitability, ultimately losing around $200.

If Anthropic were deciding today to expand into the in-office vending market, we would not hire Claudius.

Anthropic published that line itself, a rare public admission of failure from a company selling the promise of autonomous agents.

Then the Same Robot Got Boring

Anthropic did not walk away after that. Engineers made adjustments for a second phase, adding a supervising agent and better bookkeeping tools. The results improved fast. Anthropic says the business named Vendings and Stuff began performing significantly better once those changes took hold, generating a modest profit, with discounts cut by about 80% and free giveaways cut in half.

What started as a punchline became something closer to proof of concept. Andon Labs scaled the same idea up, deploying AI agents to run full retail stores and cafes under one rule: no human decision-makers. Each site runs on a multi-agent system, with a lead agent acting as a mechanical chief executive and sub-agents handling procurement, customer communication and logistics.

“Six months later, it was doing so well that it started to become a bit boring,” Lukas Petersson, cofounder and chief executive of Andon Labs, told Fortune at a COO summit in Scottsdale, Arizona. Petersson said AI agents are now running real businesses, hiring staff, managing supply chains and passing government labor inspections, without a single human decision-maker.

That is the trajectory Skyfall is implicitly betting on: a rough, possibly humiliating start, followed by a climb toward competence measured in months rather than years.

Skyfall’s Case Against Its Own Industry

Skyfall did not just claim that large language models (LLMs) hit a ceiling running a business. It built a benchmark to try to prove it. The company emerged from stealth alongside Morpheus, a persistent enterprise simulation platform built to test continual reinforcement learning, grounded in what Skyfall calls the Big World Hypothesis, where past decisions compound, objectives drift and the consequences of a bad call might not show up for days.

The findings, drawn from Skyfall’s own published research, were not flattering to the industry’s current leaders. GPT-5.5 repeatedly collapsed to zero reward in a scheduling test, recovering without any discernible pattern.

• 0.086 is the performance gap GPT-5.5 could not close against Morpheus’s theoretical upper bound during a simulated capacity drop.
• 0.148 is the wider gap posted by Gemini 3.1 Pro on that same task, nearly double GPT-5.5’s shortfall.
• Zero is the reward score GPT-5.5 repeatedly collapsed to in a long-horizon scheduling test, with no recovery pattern.

“Stable absolute performance across configuration intervals is not evidence of robustness. It is evidence of a fixed policy operating within its coverage boundary,” Skyfall’s research team wrote in the report. Pasupalak has argued that the industry’s biggest labs have every incentive to keep chasing incremental gains on existing benchmarks rather than confront that gap, since so much money already sits inside large language models.

The rest of the industry has plenty of reasons to keep betting on scale anyway. Enterprise spending on AI tools has already run into a stubborn gap between record spending and measurable payoff, even as vendors keep shipping new frontier models built on the same basic architecture.

Can an AI Actually Run a Company Without People?

Not entirely, according to Skyfall’s own founders. Pasupalak believes AI can absorb nearly every operational task a chief executive handles, freeing leaders to focus on strategy, but he still draws a firm line at human relationships, arguing that motivating employees and building real rapport remain far beyond what today’s systems can do.

Pasupalak said more than half of his own time goes into operations, pointing to meetings, budgeting and coordinating teams and campaigns as examples. “Imagine if AI handled all those operational tasks. Then I could spend far more time on strategy,” he said. On the harder part of the job, he was direct: “When it comes to motivating employees and building relationships with people, that’s still very difficult for AI to replicate.”

Suleman takes a similar view on accountability. “Trust is still important, and you probably don’t want to relinquish actual responsibility to an AI,” he said. “Humans should always remain responsible in some capacity.”

That message, AI freeing people up rather than replacing them outright, echoes a broader shift in how tech executives now talk about AI and job losses. It is a softer pitch than the one implied by buying a company specifically to see how much of its staff an algorithm can replace.

Pasupalak says the acquisition is only the first step. If it works, Skyfall wants to move up market next, targeting a business worth tens of millions of dollars within about a year, a jump in stakes that Anthropic’s own robot needed roughly that long just to earn.

Frequently Asked Questions

What Is Vending-Bench?

Vending-Bench is a benchmark where a large language model manages a simulated vending machine business for a year, covering procurement, pricing, supplier negotiation and customer complaints, built by Andon Labs to measure long-term coherence and profit before any real-world test began.

What Exactly Is an Enterprise World Model?

Skyfall’s version integrates with enterprise data sources like Sharepoint, CRM, ERP and HRIS systems to build a picture of how data, people and processes interact inside a company, an approach meant to sidestep the hallucinations and safety issues common in standard large language models.

Did Tang Yu’s Appointment Help NetDragon’s Stock?

According to a Business Insider report cited in later coverage, NetDragon’s stock rose in the year following the announcement, with some accounts crediting Tang Yu’s appointment with a roughly 10% boost, though NetDragon’s own statements focused on efficiency gains rather than share price.

Did Anthropic’s AI-Run Business Expand Beyond San Francisco?

Yes. By mid-2026 the experiment had grown from a single office vending machine into multiple locations across San Francisco, New York and London, with human colleagues shifting into an oversight role rather than running daily operations themselves.