Most of the AI industry is quietly building the same thing: a narrow agent that drafts the email, summarises the call, closes the ticket. The pitch is that companies will automate one job at a time until humans supervise the work rather than do it. Skyfall AI thinks that whole approach is aiming too low. Instead of selling another sales assistant, the startup plans to spend as much as $1 million buying a small, real B2B software or e-commerce company and installing AI in the one seat everyone else has left alone: the chief executive's. The AI will run pricing, marketing, support, finance and operations, with a stated goal of doubling revenue in six months while human involvement steadily shrinks. And it will do it in public, documenting the failures alongside the wins.
The people making the bet are not newcomers. Sam Pasupalak and Kaheer Suleman met at the University of Waterloo nearly two decades ago and went on to found Maluuba, an early deep-learning lab that worked alongside Turing Award winners Yoshua Bengio and Richard Sutton before Microsoft bought it for around $160 million in 2017. Reunited now under Skyfall, they are making an argument that is really a critique of the entire frontier-model industry. Running a business, Pasupalak says, is not like writing software, which is largely logical and deterministic. It demands abstract reasoning, long-term planning and high-stakes decisions made with incomplete information, the messy stuff that today's chatbots are worst at.
To back the claim, Skyfall is releasing a benchmark called Morpheus, which tests whether frontier models can keep adapting as a business environment shifts. Its finding is pointed: models such as GPT-5.5 and Gemini perform well in familiar conditions but deteriorate as competitors launch products and customer behaviour changes. They are brilliant at applying what they already know and poor at learning continuously from experience. Scaling the current architecture, Skyfall argues, will not close that gap. Its proposed answer is what it calls an Enterprise World Model, a system that predicts how a company evolves in an abstract "latent space" rather than word by word, focusing only on the variables that actually move the business.
The idea is less unprecedented than it sounds. Anthropic's Project Vend handed a Claude-powered agent a real office vending business; the first version was talked into stocking novelty tungsten cubes and selling them at a loss, but by mid-2026 the operation had expanded across San Francisco, New York and London and, in the words of its overseers, become "almost boring" because it ran so smoothly. Earlier stunts, like the Chinese gaming firm that named an AI its "CEO," were mostly theatre. What separates Skyfall is the insistence on making the hypothesis falsifiable: buy a company with real revenue, real customers and real consequences, and let the scoreboard settle the argument.
Even the founders draw a line. Pasupalak doubts AI should handle motivating employees or building relationships, and Suleman insists a human must remain accountable no matter how much the machine runs day to day. That caveat is the tell. The genuinely interesting outcome here is not whether an AI can double a small company's revenue, but what it fails at when nobody is steering, because that failure list, published honestly, would be worth more than any benchmark. Most AI demos are engineered to succeed. Skyfall is promising one designed so it can visibly lose. That alone makes it worth watching.