Princeton Gives AI One Million Dollars – Most Go Bankrupt

According to tech optimists, so-called AI agents, or artificial intelligence that acts independently and makes decisions, are on the verge of a breakthrough. They are expected to book our vacations, shop for us, and even run entire companies. However, in practice, there are still several hurdles to this vision, as even the best agent-based AI currently operates unreliably and unpredictably – with sometimes disastrous consequences.

Researchers at Princeton University in New Jersey wanted to find out how AI agents based on various large language models would lead a fictional startup as CEOs. The models were given a starting capital of one million US dollars and were to establish a profitable business model over a simulated period of 500 days under realistic conditions, making independent decisions in areas such as pricing, marketing, and product development. The study was recently published as a preprint paper under the name CEO-Bench.

Can an AI Agent Take on the Role of a CEO?

As the researchers state, AI agents are currently capable of mastering individual tasks, such as independently programming an application or booking a hotel. They are suited for "isolated tasks with a short-term time horizon." However, they still struggle with long-term goals where various uncertainty factors arise. Here, humans still excel with their "leadership intelligence," as the researchers call it.

With CEO-Bench, the team aimed to find out whether the currently most powerful language models like GPT 5.5 and Claude Fable 5 could at least show some semblance of such leadership intelligence. Similar simulations where AI agents were promoted to CEO have occurred in the recent past, but the Princeton researchers added further factors to these.

Simulation for AI Agents: Zero Customers and One Million Dollars in Starting Capital

The AI agents had to run a fictional startup called Novamind, which generates profits through subscription payments from customers and monetization of advertising within the product. The simulation begins with zero customers and one million dollars in starting capital. Once per simulated week, the agents can access 34 tools via a Python interface, analogous to departments in a real company. These include:

• Marketing: The AI agents can utilize various channels, such as social media, and develop advertising campaigns to attract specific customer groups.
• Pricing: They can set prices for various subscriptions and discounts and advertising measures.
• Product and Research: They can invest in the development of the existing product or initiate long-term research.
• Infrastructure: They can purchase additional server capacity or expand customer service to minimize system failures or delays.
• Public Communication: They can monitor social media and respond to feedback to improve the company's public image.

From Market Research to Marketing – What AI Agents Must Consider

To make Novamind profitable, the agents must continuously navigate the interplay between subscription and advertising revenues, capacity and computing costs, support, development, acquisitions, market research, and research projects, adjusting their decisions accordingly.

The researchers defined 26 customer groups. The models do not know their preferences but must derive them from feedback on social media. Additionally, they must be able to handle random events over the course of the 500 days.

Not least, they receive some data with a delay: "Costs can arise immediately, while the effects on revenue, customer retention, research, and reputation may only become apparent weeks later," they state. This was intended to determine how the agents behave in a constantly changing environment and whether they can adjust their decisions not only for short-term profit but also for long-term strategy.

Most AI Agents Drive Their Startup to Bankruptcy

If the available capital fell below zero, the startup went bankrupt, and the simulation ended. In fact, most AI agents proved to be terrible CEOs; only Claude Fable 5, Claude Opus 4.8, and GPT-5.5 managed to increase the starting capital. However, the latter two only succeeded in their best simulations. In other runs, they did not make a profit but also did not go bankrupt. Speaking of bankruptcy: Grok 4.20 performed the worst. The agent based on Elon Musk's xAI language model crashed the business in less than 40 days.

An interesting finding from CEO-Bench is the different strategies the agents pursued. For instance, GPT 5.5 attempted to maintain a consistent customer base over the entire 500-day period by slightly increasing product quality in the short term to prevent customer churn. In contrast, Claude Opus 4.8 made a strategic shift after an initially aggressive growth phase in one simulation: it consciously accepted that the number of customers would drop to zero by the end but simultaneously stopped spending on advertising, development, and technology, thereby salvaging profit just over the "finish line." A strategy that would be rather impractical in the real world.

What Can Be Learned from the CEO-Bench Study

In the end, the study shows that AI agents still struggle to maintain a coherent strategy over 500 days. However, some models do show potential to manage resources in a constantly changing environment, plan proactively, and respond to market changes – such as shifts in customer preferences.

"CEO-Bench is a step towards developing agents and training models that not only answer queries but also help long-term organizations navigate uncertainties," the Princeton researchers conclude. However, it is better not to let them run a company at this time.