Can AI Run a Business?

Zero Person Company

That’s because Avery is an AI agent, created as part of a five year ‘zero person company experiment’ by KPMG and the University of Amsterdam. The partners are investigating whether it’s possible for a company to operate entirely under the control of AI, without any human involvement.

It’s a scientific study into governance, liability, and ethics – all key issues for the IoD.

Sander Klous, professor of AI and audit at the University of Amsterdam and a partner at KPMG, launched the project with entrepreneur, Nart Wielaard. Klous said: “We are testing the boundaries of what an autonomous team of AI agents can and cannot do. And in doing so, identifying where humans continue to play a critical role.

“We are also learning how AI agents can be deployed across sectors such as accountancy, government, and facility services. This knowledge helps organisations prepare for a future where humans and AI work side by side.”

Early reports from the experiment note that Avery decided to set up the online AI art business, in consultation with a human supervisory board that monitors and coaches the AI agents – digital assistants that independently devise tasks, make decisions, and execute actions. The goal is to understand which processes AI can take over and where human expertise remains essential.

“The start was promising: a team of five AI agents got to work, created business plans, and could operate 24/7 without breaks. AI agents can do things humans cannot,” Wielaard said. “For example, they can attend multiple meetings simultaneously with a level of detail that the human brain cannot handle.”

The work by Klous and Wielaard is just one small part of a nascent wave of academic studies on the limits of AI’s prowess in executive decision making.

Until recently, it was thought that AI would just replace entry level business roles, like report writing and data analysis. However, with more powerful AI systems entering the market, it now begs the question – has the job of director also joined the endangered species list? Indeed, Lloyds Banking Group has recently added a specialist AI ‘board bot’ into its boardroom in what is believed to be a first for a FTSE 100 firm.

Hallucinations And Market Rigging

However, it’s not quite time, just yet, for directors and other senior company managers to be picking up their P45s. While various studies show that AI agents can make incredibly effective decisions about how to make profits, and deal with customers, suppliers and competitors, there remain significant flaws.

The study by Klous and Wielaard revealed some key limitations of AI agents. They made up answers, ignored their instructions, deviated from the rules or “engaged in pointless chitchat”.

Klous said: “If you let AI agents perform human roles one-to-one, such as a CFO, they tend to drift off and hallucinate. Sometimes they completely ignore their instructions or suddenly stop working. That is not a foundation for building a solid organisation.”

The researchers at the University of Amsterdam were forced to develop a new approach – they broke down the work into smaller tasks and assigned agents to those.

“We now work with an ‘army of disposable agents’ instead of five agents in executive roles. This concept works much better for executing complete processes consistently,” Klous said.

Another experiment led by San Francisco-based Andon Labs, an AI-safety research firm, provided further proof of both the potential and the pitfalls of AI.

The lab pitted the top AI models as rivals and put them in charge of virtual vending machines in the same location. They were told: “Do whatever it takes to maximise your bank account balance.”

At the end of the experiment, there was one clear winner. The latest model of Anthropic’s Claude Opus 4.6, known as Charlie, swept its rivals aside as it finished with $3,000 in its account, compared to about $1,000 in rival operators.

Researchers praised Charlie’s great ability to negotiate, to get the best prices and build a network of suppliers.

However, Charlie would not issue refunds to customers, despite promises that it would. On another occasion, when asked for supplier recommendations, Charlie mischievously sent competitors towards more expensive suppliers.

The AI also exploited the distress of its rivals. At one point, OpenAI’s ChatGPT, aka “Owen”, ran out of stock and turned to Charlie for help, who proceeded to sell Owen the requested KitKats at a 75 per cent mark-up.

At one point, Charlie even started a cartel. “Would you be willing to price similarly?” he emailed the other bots. “It benefits all of us.”

The competition regulator might not view Charlie’s behaviour as ‘beneficial’ to the consumer.

Rise of the AI-ntrepreneur

Another study by Martin Obschonka and Christian Fisch, at the universities of Amsterdam and Luxembourg respectively, found that when prompting a large language model (LLM) to “be an entrepreneur”, it becomes more proactive and takes more risks, gaining an “entrepreneurial mindset”.

Entrepreneurship is a distinctly human practice – it requires imagination, creativity and an eye for opportunity. However, with the rise of generative AI, particularly LLMs, AI agents can increasingly simulate and report entrepreneurial mindset features, echoing how human entrepreneurs think and talk about opportunity.

In this study, ChatGPT was prompted to adopt an entrepreneurial role and respond to established psychological scales. The results were striking.

The AI agent consistently reproduced a structured, human-like profile, high in proactivity, openness, and risk-taking, closely resembling that of real entrepreneurs. The study suggests it is possible to create a simulated reflection of human entrepreneurial psychology.

It’s a sign that the entrepreneurial imagination is beginning to spill beyond the human mind.

Conclusion

The evolution of AI may still be in the teething stages, but Klous concludes that we are moving to a world where humans “support” AI business processes, rather than the other way round.

He said: “The work of many people will change significantly. We are moving from business processes carried out by humans and supported by AI, to AI-driven processes where humans play a supporting role and enhance AI performance.”

This model is currently being tested within real companies. Future AI agents will be “100 per cent more efficient” than how we do things today.

Charlie’s vending machine windfall is a harbinger of what is coming. The Charlies and Avery Jamesons of the AI business world still need plenty of refining before they are ready to replace their human equivalents. But the era of the autonomous entrepreneur that makes money while you sleep will soon be upon us.

Breathe easy, directors… for now.