AINEWS
Skyfall AI's model will run pricing, marketing, finance, and operations at a small business it acquires for up to $1 million, with a goal of doubling revenue in six months.
The training ground was a theme-park simulator. Before Skyfall AI committed money to a real acquisition, the team ran its model through OpenRCT2, the open-source remake of the 2002 management game RollerCoaster Tycoon 2, and watched it optimize park revenue over weeks of simulated time. The sim was the team's confidence check: if the system could not turn a profit on a virtual park, it was not ready to try on a real business. The dry run passed, the company says, so the next step is to buy a small SaaS or e-commerce company for up to $1 million and let the same system run it.
That is the announcement Forbes reported this week, with Gizmodo and Futurism following. The WSJ newsletter has also noted the company's emergence from stealth. Skyfall frames the experiment as a test of an "AI CEO," its own term rather than an industry one, that will oversee pricing, marketing, customer support, finance, and operations at the acquired business, with human involvement gradually reduced. The public success metric is a doubled top line within six months.
The founders behind it are Sam Pasupalak, Kaheer Suleman, and Sumit Pasupalak. The first two co-founded Maluuba, a deep-learning and reinforcement-learning startup that Microsoft acquired in 2017. That pedigree is the credibility core of the pitch: people who shipped reinforcement-learning research into a real product at hyperscaler scale. The thesis they are now testing, however, is a critique of the same hyperscaler playbook they came out of.
In a post on its own site, Skyfall argues that five years of LLM scaling, with more data, more compute, and bigger models, is running into a ceiling. Frontier models from Anthropic, OpenAI, and Google, the company writes, can ace ARC-AGI-style reasoning benchmarks but still fail at running a business, because business decisions have consequences that arrive days, weeks, or quarters later. A pricing change today affects next quarter's churn. A marketing spend today affects a brand that compounds over years. The argument is that the next paradigm is not a larger language model but an "enterprise world model": a system that simulates the business's own dynamics and updates its beliefs as conditions change, paired with continual learning rather than fixed training.
Skyfall built its own benchmark to test that claim, and the public results will be the evidence the experiment rests on. The company's own research, not yet independently replicated, says current frontier models handle simple tasks and break on long-horizon decisions with delayed consequences. That is the founders' own admission, and it is also the gap the experiment is designed to stress.
A $1 million small-business acquisition is a narrow evidence base, and the founders are not claiming it generalizes. Skyfall's framing is closer to a lab test than a market entry: the acquisition exists to generate real revenue data, real customer feedback, and a real failure surface for a world-model system that has, until now, only run inside simulators and benchmarks. Whether "AI CEO" becomes a job title or a marketing phrase will depend less on this one experiment and more on what the next round of public evidence shows.
The first signal will be the public release of Skyfall's benchmark comparing world models against frontier LLMs on long-horizon business tasks. The second is which business Skyfall actually buys, and at what multiple. The third is whether the revenue line moves at all, let alone doubles, in the stated six-month window. The fourth is on the customer side: does anyone notice, and is anyone worse off, when an AI starts changing prices and answering support tickets? Each of those answers is on a clock the company itself set.